What Makes a Good Image?

Airbnb Demand Analytics Leveraging

Interpretable Image Features

Shunyuan Zhang
Harvard Business School
[email protected]

Dokyun Lee
Questrom School of Business, Boston University
[email protected]

Param Vir Singh

Tepper School University, Carnegie Mellon University
[email protected]

Kannan Srinivasan
Tepper School University, Carnegie Mellon University
[email protected]

ABSTRACT

We study how Airbnb property demand changed after the acquisition of verified images (taken by Airbnb’s
photographers) and explore what makes a good image for an Airbnb property. Using deep learning and
difference-in-difference analyses on an Airbnb panel dataset spanning 7,423 properties over 16 months, we find
that properties with verified images had 8.98% higher occupancy than properties without verified images (images
taken by the host). To explore what constitutes a good image for an Airbnb property, we quantify 12 human-
interpretable image attributes that pertain to three artistic aspects—composition, color, and the figure-ground
relationship—and we find systematic differences between the verified and unverified images. We also predict the
relationship between each of the 12 attributes and property demand, and we find that most of the correlations
are significant and in the theorized direction. Our results provide actionable insights for both Airbnb
photographers and amateur host photographers who wish to optimize their images. Our findings contribute to
and bridge the literature on photography and marketing (e.g., staging), which often either ignores the demand
side (photography) or does not systematically characterize the images (marketing).

Keywords: sharing economy, Airbnb, property demand, computer vision, deep learning, image feature
extraction, content engineering

Electronic copy available at: https://ssrn.com/abstract=2976021

1. Introduction
The global sharing economy market has been rapidly expanding in recent years and is projected to generate
roughly $335 billion by 2025 (PwC report 2015). On Airbnb, travelers can find lodging in home environments
rather than hotels, and hosts can generate rental income from spare rooms or properties. As the world’s largest
home-sharing platform, Airbnb recently was valued 20% higher than Marriott and hosted 25% more guests per
night than Hilton Worldwide (Winkler and Macmillan 2015).
Despite its success, Airbnb faces a significant problem in resolving the uncertainty that consumers face when
evaluating property quality. The inefficiency of transferring information required by prospective guests creates
transactional friction and leads to the loss of users. Reports show that quality uncertainty leads many potential
consumers to choose trusted hotel brands over Airbnb (PwC report 2015; Ufford 2015).
Airbnb attempts to alleviate quality uncertainty with features such as customer reviews, host verification,
property descriptions, and property images. In particular, property images provide visual information and reduce
uncertainty about experiential aspects (e.g., cleanliness and mood) of units in ways that written reviews and
descriptions cannot. While hotel images are taken by professional photographers, however, most Airbnb property
images are taken by hosts—usually amateur photographers. Therein lies a source of inefficiency of information
transfer that causes uncertainty for potential guests. Hosts often lament the poor quality of their own amateur
photos and fear that their properties appear small.
To address this concern, in 2011, Airbnb launched a “photography program” that gives interested hosts free
access to local professional photographers who are commissioned by the company to take photos of the host’s
property. A “verified” mark appears on images shot and uploaded by Airbnb’s professional photographers.
Figure 1 juxtaposes the original amateur photographs with Airbnb’s professional photographs of the same room
to illustrate the improvement.
Figure 1. Comparison of Unverified and Verified Photos
Unverified
Verified

It remains unclear, however, whether the professional photography program has been beneficial for Airbnb
and its hosts. In fact, the photography program has raised much controversy among Airbnb hosts and guests. It
seems plausible that verified photos could attract more guests, but they also could oversell or misrepresent the
property, leading to a negative impact on property demand —a fear voiced by some Airbnb hosts on the host

Electronic copy available at: https://ssrn.com/abstract=2976021

forum.1 The Airbnb photography program raises a series of questions: (1) Do properties with verified photos
experience higher demand? (2) If so, what systematic differences between verified and unverified photos might
explain the higher demand? (3) Finally, what constitutes a good image for an Airbnb property?
To answer these questions, we collected panel data from 13,000 Airbnb listings with over 510,000 property
images in seven major US cities from January 2016 through April 2017. The dataset contains rich information
about each property’s monthly reservations, photos, price, and other details about the property and host. The
dataset captures variation in property images across both units and time, so we can observe some of the properties
transition from using unverified photos to verified photos.
We use Amazon Mechanical Turk (AMT), a crowdsourcing platform for on-demand human tasks, to classify
a random set of pictures into the binary categories of high- and low-quality. We use the manually classified
training set to develop a scalable model that takes advantage of computer vision and deep learning. Specifically,
taking pixel-level information from the images as input, we apply a convolutional neural network (CNN) to
classify the aesthetic quality of each image in the training sample. The CNN model is optimized to extract a
hierarchical set of features from images and learn the relationship between the set of features and the image’s
label (in this case, “high quality” or “low quality”). Then, we use our trained CNN image quality classifier to label
the property images in the Airbnb dataset.
We employ a difference-in-difference (DID) analysis to compare the demand for properties with verified
photos and with unverified photos. Hosts self-select into the Airbnb photography program, so we address
endogeneity concerns through propensity score weighting (PSW). The PSW approach assumes that selection into
the treatment group is captured by observed variables, and the algorithm matches treated and untreated units
using estimated treatment probabilities. The PSW approach mitigates the self-selection problem only so far as
the factors that affect both the treatment probability and property demand are captured in observed variables.
Thus, we conduct a Rosenbaum bounds analysis to test the robustness of our results to hypothetical
unobservables. The results of the Rosenbaum analysis are within the range reported in prior literature, and a
series of other tests suggests similarly robust results. Nevertheless, we cannot fully rule out the possibility that
selection is driven by unobserved variables in the absence of a random experiment, and we caution readers to
interpret our results with this caveat in mind.
We find that the occupancy rate is 8.985% higher for properties with verified photos than for properties with
unverified photos. The treatment coefficient is positive and significant even after controlling for other sources
of information, such as guest reviews. The estimated treatment coefficient decreases dramatically after we
incorporate photo characteristics into the demand model. Particularly, the coefficient decreases by 41.0% when
we control for image quality, suggesting that a significant portion of the coefficient is explained by the higher
quality of the verified images.

1 http://airhostsforum.com/t/professional-photography/3675/35.
3

Electronic copy available at: https://ssrn.com/abstract=2976021

 One of our key objectives is to determine what makes a good image for an Airbnb property. Our CNN model
is highly accurate at predicting image quality, but the CNN-extracted features are uninterpretable. To provide
better guidance for managers, we use the photography literature to identify 12 human-interpretable image
attributes that are relevant to image quality in the real estate context. We theorize the relationship between each
of the 12 image attributes and property demand. The 12 attributes fall under three key artistic aspects:
composition, color, and the figure-ground relationship. Composition is the arrangement of visual elements in the
photograph; ideally, the composition leads the viewer’s eyes to the center of focus (Freeman 2007). We capture
composition with four attributes: diagonal dominance, the rule of thirds, visual balance of color, and visual
balance of intensity. Color can affect the viewer’s emotional arousal. The marketing literature has studied the
impact of color on consumer behavior particularly in the context of web design, product packaging design, and
advertisement design (Gorn et al. 1997, Gorn et al. 2004; Miller and Kahn 2005). We include five aspects related
to color: warm hue, saturation, brightness, contrast of brightness, and image clarity. The principle of the figure-
ground relationship is one of the most basic laws of perception and is used extensively by expert photographers
to plan their photographs. In visual art, the figure refers to the key region (i.e., foreground), and the ground refers
to the background; photographs in which the figure is inseparable from the ground do not retain the viewer’s
attention. We include three attributes: the area difference, texture difference, and color difference between the
figure and ground.
We use computer vision methods to score the images on the 12 image attributes, and we find systematic
differences between the verified photos and both the unverified low-quality and unverified high-quality photos.
The verified photos dominate the unverified high-quality photos on all attributes except for the rule-of-thirds
and saturation. The verified photos score higher than the unverified high-quality photos across the attributes for
which the theorized effect of the attribute is positive; however, the verified photos score lower than the unverified
high-quality photos when the theorized effect of the attribute is negative.
We explore how the 12 image attributes might be related to property demand, though we cannot identify
causal effects of these 12 image attributes on property demand given the observational nature of our data. That
is, the image attributes change when a property acquires verified photos, but the choice to do so is endogenous,
making any changes in the image attributes also endogenous. As a result, our analysis of the relationship between
the image attributes and property demand should be viewed as exploratory. After controlling for the 12 attributes,
the estimated treatment coefficient becomes statistically insignificant. The results suggest that the positive
coefficient of verified photo acquisition may derive from the 12 image attributes, which together capture both
vertically-differentiated quality and horizontally-differentiated taste. Airbnb professional photographers seem to
be better than host photographers at capturing the attributes that matter for Airbnb property demand.
Of the 12 image attributes, the visual balance of color is most strongly related to property demand, followed
by image clarity and the contrast of brightness. The visual balance of color refers to color symmetry, which can
be affected by both the property itself and the position from which the image is captured. Image clarity refers to

Electronic copy available at: https://ssrn.com/abstract=2976021

the extent to which the image conveys visual information. The unverified low-quality images scored poorly on
image clarity; the verified photos scored almost twice as high. Even without employing a professional
photographer, hosts can improve image clarity through the effective use of lighting and access to a good camera.
Finally, the contrast of brightness captures the difference in illumination between the brightest and dimmest
points in the image; a low contrast of brightness indicates that illumination is relatively even across the image.
The verified photos have a significantly lower contrast of brightness than unverified high-quality images.
Interestingly, several hosts on the Airbnb community forums complained that the contrast of brightness is so
low in the verified photos that they appear washed out, but we find the predicted negative relationship between
the contrast of brightness and property demand. In other words, consumers seem to prefer the low contrast of
brightness that appears in verified photos.
This study makes several contributions. The ability to capture and analyze unstructured data, particularly
images, is a nascent field with few papers (e.g., Wang et al. 2018, Malik and Singh 2019, Netzer et al. 2019, Malik
et al. 2020, Liu et al. 2020). Ours is among the first to examine how Airbnb property demand changed with the
adoption of the Airbnb photography program. From a managerial perspective, our results suggest that the
program is successful for hosts who choose to take advantage of it—properties with verified images achieved an
8.98% higher occupancy rate than properties without verified images. However, more than half of the Airbnb
properties use low-quality images. Our findings may motivate hosts to obtain high-quality images of their
properties.
Although it seems intuitive that the demand should be higher for properties with good images, the
components of a “good image” are less obvious. A key practical finding of our paper is the identification of 12
interpretable image attributes that significantly correlate with property demand. Our investigation has parallels
with marketing studies on color as a key image attribute that affects consumers’ perceptions and product demand.
Beyond color, we consider composition and the figure-ground relationship, and we find that even the high-quality
unverified images are systematically inferior to the verified images on most of the 12 attributes. Our findings may
be valuable to photographers who are attempting to optimize the appeal of their images for Airbnb consumers.
Moreover, insights from our paper can guide image-content engineering efforts in short-term lodging
contexts beyond Airbnb (e.g., hotels) and real estate markets. Our approach efficiently computes an extensive
set of image attributes (at ~1.06 seconds per image) on a 2-Intel-Haswell (E5-2695 v3) CPU, so it can be
implemented as a scalable real-time model. Our application of an existing method for image-quality classification
and feature extraction can be adopted by photographers of firms and hosts to check their image quality and
identify the shortcomings of their photographs. Such a practice ultimately could improve the efficiency of image-
based information transfer and may increase the appeal of their properties.
Prior studies in the marketing literature have considered the effects of images and visual elements, but our
paper differs from existing papers in three ways. First, the extant studies focus on the viewer’s emotional arousal
and are restricted to either isolated image features (e.g., color; Gorn et al. 1997, Gorn et al. 2004) or high-level

Electronic copy available at: https://ssrn.com/abstract=2976021

image content or style (e.g., whether an image is “art”; see Hagtvedt and Patrick 2008). By contrast, we use the
photography literature to identify 12 image attributes on which any photograph can be evaluated, and we explore
their relationships with property demand. Second, the relevant literature focuses primarily on high-quality images
(Bertrand et al. 2010), but e-commerce (including the sharing economy) often uses low-quality, user-generated
product images. Finally, the evolving stream of marketing literature on the impact of images on consumers’
product perceptions focuses on the mere presence or absence of an image rather than specific image attributes
(e.g., Mitchell and Olsen 1981; Meyers-Levy and Peracchio 1992; Peracchio and Meyers-Levy 1994, 2005; Scott
1994; Valdez and Mehrabian 1994; Gorn et al. 1997; Larsen et al. 2004; Miller and Kahn 2005).

2. Empirical Framework
2.1 Data Description
We randomly selected more than 13,000 Airbnb property listings from seven US cities (Austin, Boston, Los
Angeles, New York, San Diego, San Francisco, and Seattle), and we collected data on the listings from January
2016 to April 2017. We obtained information about each property host from that host’s public profile on
Airbnb.com. Each profile specifies the date on which the host joined Airbnb and whether the host had a verified
Airbnb account at the time of the analysis. For each property, we obtained information about static
characteristics: location (city and zip code), property type (e.g., house, apartment), property size (number of beds),
amenities (e.g., pool, AC, proximity to a beach), and capacity (maximum guests). We also obtained information
about dynamic characteristics: property bookings, nightly prices, guests’ reviews, property photos, and whether
the photos were verified. In the appendix (Section VI), we provide detailed methods for data collection and
matched sample construction. In the next section, we describe the measures of our key variables and summarize
their measurement statistics.

2.2 Definitions and Measures of Key Variables

Treatment Group, Untreated Group, and Treatment Status
The panel data cover 16 one-month periods from January 2016 through April 2017. The “treatment” is the
adoption of verified photos; therefore, a property is “treated” if it used verified property photos during the
observation period. The sample for our main analysis consists of 7,423 unique properties that did not have
verified photos in January 2016;2 212 had verified photos by the end of April 2017 (the treatment group), and
the remaining 7,211 properties did not (the untreated group). We capture the treatment with three indicator
variables. TREATi equals 1 (0) if property i belongs to the treatment group (untreated group). AFTERit equals 1
(0) if period t is after (before) the period when property i was first observed to have verified photos. For example,
if a property acquired verified property photos in March 2016, then AFTERit equals 0 for the periods of January

2 Of the 13,000 listings in our original dataset, approximately 5,000 were already “treated” in January 2016. We exclude them
from our main analyses, but as a robustness check, we repeat our main analysis with the pre-treated units, and we obtain
consistent results (see Section V.7 in the appendix for details).
6

Electronic copy available at: https://ssrn.com/abstract=2976021

and February 2016 and equals 1 for the periods of March 2016 onward. Hence, the treatment status indicator
TREATINDit = TREATi * AFTERit equals 1 if i was treated in period t and equals 0 otherwise.

Property Demand
We purchased listing-level booking data from a company that specializes in collecting Airbnb property demand
data. The booking data includes the number of days in a month in which the property was open (i.e., available to
be booked), blocked (i.e., marked as unavailable by the host), and booked (i.e., by a guest). For property t in
period t, we operationalize property demand as the occupancy rate, that is, the fraction of the open days that
were booked, scaled by 100. For example, if a property in March was open for 24 days and booked for six days,
then its demand for that month was DEMANDit = (6/24) * 100 = 25.00.

Property Price
The property price for property i in period t, NIGHTLY_RATEit, refers to the average nightly price over the
days in period (Zhang et al. 2019). Property price is endogenous because it correlates with random demand
shocks in the current period, which also affects property demand. To address the endogeneity concern, we use a
set of instrument variables (IVs) for price. Following the extant literature, we include the characteristics of
competing properties (Berry et al. 1995; Nevo 2001). The logic is that the characteristics of competing products
are unlikely to be correlated with unobserved shocks in the demand for the focal property. However, the
proximity of the characteristics of a property and its competitors influences the competition and, as a result, the
property markup and price. 3 In addition, we collect cost-related variables, that is, the factors that enter the
cost/supply side but not the demand side. We use the local (zip-code level) residential utility fee obtained from
OpenEI and local rental information collected from Zillow.4 These factors affect the host’s expenses and thus
provide an indirect measure of cost but are unlikely to be correlated with demand in the short-term lodging
market.

Property Photos
The property photos are the set of photos posted on the property web page in the given period. Three variables
characterize the property photos: photo quantity, photo quality, and the distribution of photographed room
types. IMAGE_COUNTit is the number of photos of property available during period t. We calculate
IMAGE_QUALITYit using machine learning techniques that are appropriate for the size of the dataset (over
510,000 images).5 We build a supervised image-quality classifier that classifies each image as high-quality (value

3 We compute IVs based on property type, listing type, property capacity, and number of reviews, none of which are directly
controlled by the host. Competitors are defined as the properties in the same zip code.
4 The OpenEI dataset provides average residential, commercial, and industrial electricity rates by zip code, compiled from

ABB, Velocity Suite, and the US Energy Information Administration dataset 861: https://openei.org/doe-
opendata/dataset/u-s-electric-utility-companies-and-rates-look-up-by-zipcode-feb-2011. Zillow Research provides average
home values by zip code and home size (# of bedrooms): https://www.zillow.com/research/data/.
5The image data contains all images associated with all properties included in the dataset. That is, they include images for
properties that were verified before the observation started (and hence are not included in the sample for the DiD analyses)
and all images updated/added/deleted during observation periods.
7

Electronic copy available at: https://ssrn.com/abstract=2976021

of 1) or low-quality (value of 0). We calculate the average image quality, IMAGE_QUALITYit, of all the photos
associated with property i in period t. For example, if property i had 10 images in period t, and eight images are
labeled as high-quality, then IMAGE_COUNTit = 10 and IMAGE_QUALITYit = (8 * 1 + 2 * 0)/10 = 0.8.
Lastly, we include the distribution of photographed room types because professional photographers may have a
more advanced understanding of which types of rooms will be most appealing to guests and, thus, present more
of these aspects of properties. Specifically, we compute the proportion of the photographs that depict each of
five room types: bathroom, bedroom, kitchen, living room, and outdoor area. Then, for property i in period t,
the distribution is represented by a vector: {BATHROOM_PHOTO_RATIOit, BEDROOM_PHOTO_RATIOit,
KITCHEN_PHOTO_RATIOit, LIVINGROOM_PHOTO_RATIOit, OUTDOOR_PHOTO_RATIOit}.
Table 1 presents a summary of the statistics for the key variables at the group level. To show the overall trends
in the key variables, we report statistics for the pre-treatment period (January 2016, i.e., when none of the
properties in the sample were treated) and the post-treatment period (April 2017), when all the properties in the
treatment group had been treated. As expected, image quality was stable over time in the control group (0.29 in
January 2016 vs. 0.30 in April 2017) and improved dramatically in the treatment group (0.27 in January 2016 vs.
0.77 in April 2017), reflecting the high quality of photos shot by Airbnb professional photographers.

Table 1. Summary of the Statistics of Airbnb Properties

Control Group Treatment Group

Mean Std. Dev. Mean Std. Dev.
Sample Size: # Units (N) 7,211 212
Pre-treatment (January 2016)
DEMAND (occupancy rate * 100) 31.07 35.93 32.57 32.30
# RESERVATION DAYS (demand component-days 5.49 8.79 6.62 8.45
booked by guests)
# AVAILABLE DAYS (demand component-day not 12.91 12.34 14.87 11.86
booked, not blocked)
# BLOCKED DAYS (demand component-days blocked 12.60 13.42 9.51 12.43
by host)
IMAGE_QUALITY 0.29 0.27 0.27 0.25
IMAGE_COUNT 12.78 9.49 14.48 10.38
BATHROOM_PHOTO_RATIO 0.21 0.17 0.22 0.16
BEDROOM_PHOTO_RATIO 0.30 0.20 0.29 0.19
KITCHEN_PHOTO_RATIO 0.11 0.11 0.10 0.10
LIVINGROOM_PHOTO_RATIO 0.19 0.17 0.18 0.16
OUTDOOR_PHOTO_RATIO 0.19 0.20 0.20 0.20
SECURITY_DEPOSIT 181.97 347.75 202.77 333.07
CLEANING_FEE 48.02 53.04 54.69 58.95
MAX_GUESTS 3.19 2.11 3.50 2.23
SUPER_HOST 0.09 0.29 0.15 0.35
INSTANT_BOOK 0.11 0.31 0.11 0.31
MINIMUM_STAY 2.62 2.71 2.57 2.92
RESPONSE_RATE 91.00 16.29 92.25 14.31
RESPONSE_TIME (minutes) 261.07 365.66 225.12 338.39
NIGHTLY_RATE (average property price) 179.74 249.31 170.15 240.80
REVIEW_COUNT 16.41 26.25 20.56 26.98
HAS_RATING 0.68 0.47 0.78 0.41

Electronic copy available at: https://ssrn.com/abstract=2976021

RATING_OVERALL (overall score out of 100; 92.88 7.14 93.99 5.04
summarizes multi-dimensional sub-ratings)
RATING_COMMUNICATION 9.71 0.64 9.74 0.49
RATING_ACCURACY 9.50 0.76 9.54 0.56
RATING_CLEANLINESS 9.21 0.97 9.26 0.82
RATING_CHECKIN 9.67 0.67 9.72 0.48
RATING_LOCATION 9.39 0.82 9.40 0.70
RATING_VALUE 9.24 0.79 9.34 0.60
ZILLOW_RENTAL 2,459.95 800.40 2,474.75 845.05
UTILITY 0.18 0.03 0.17 0.04
INACTIVE PROPERTIES (proportion; a property in a
month is inactive if it’s blocked full month) 0.24 0.17
Post-treatment (April 2017)
DEMAND (occupancy rate * 100) 36.79 39.46 46.58 38.37
# RESERVATION DAYS (demand component-days 8.59 10.35 10.97 10.13
booked by guests)
# AVAILABLE DAYS (demand component-day 16.23 11.56 13.41 11.49
not booked, not blocked)
# BLOCKED DAYS (demand component-days 5.17 8.79 5.62 8.94
blocked by hosts)
IMAGE_QUALITY 0.30 0.27 0.77 0.22
IMAGE_COUNT 16.22 11.98 19.61 11.53
BATHROOM_PHOTO_RATIO 0.21 0.16 0.21 0.15
BEDROOM_PHOTO_RATIO 0.30 0.19 0.30 0.18
KITCHEN_PHOTO_RATIO 0.11 0.10 0.09 0.10
LIVINGROOM_PHOTO_RATIO 0.19 0.15 0.18 0.17
OUTDOOR_PHOTO_RATIO 0.20 0.19 0.21 0.21
SECURITY_DEPOSIT 149.53 325.71 179.82 280.53
CLEANING_FEE 41.51 55.93 46.02 51.01
MAX_GUESTS 3.37 2.26 3.50 2.31
SUPER_HOST 0.19 0.39 0.29 0.45
INSTANT_BOOK 0.11 0.32 0.16 0.37
MINIMUM_STAY 2.36 3.75 2.56 4.06
RESPONSE_RATE 96.63 12.17 95.82 12.50
RESPONSE_TIME (minutes) 150.42 266.86 167.72 287.81
NIGHTLY_RATE (average property price) 237.97 319.49 210.14 146.81
REVIEW_COUNT (number of guest reviews) 32.89 49.33 41.80 46.37
HAS_RATING 0.82 0.38 0.84 0.36
RATING_OVERALL (overall score out of 100; 93.80 5.34 94.16 4.45
summarizes multi-dimensional sub-ratings)
RATING_COMMUNICATION 9.80 0.45 9.82 0.40
RATING_ACCURACY 9.60 0.60 9.66 0.51
RATING_CLEANLINESS 9.40 0.75 9.40 0.73
RATING_CHECKIN 9.79 0.45 9.78 0.45
RATING_LOCATION 9.47 0.67 9.50 0.65
RATING_VALUE 9.36 0.64 9.39 0.57
ZILLOW_RENTAL 2,459.80 822.06 2,426.22 823.28
UTILITY 0.18 0.04 0.17 0.04
INACTIVE PROPERTIES (proportion) 0.41 0.32
Time-invariant
APARTMENT 0.67 0.47 0.60 0.49
ENTIREHOME 0.59 0.49 0.61 0.49
# BEDS 1.70 1.22 1.81 1.21
POOL 0.08 0.28 0.10 0.30
BEACH 0.01 0.10 0.02 0.14
AC 0.98 0.13 0.99 0.06
PARKING 0.44 0.50 0.50 0.50
9

Electronic copy available at: https://ssrn.com/abstract=2976021

 INTERNET 0.98 0.13 0.99 0.09
TV 0.74 0.44 0.79 0.41
WASHER 0.59 0.49 0.60 0.49
MICROWAVE 0.09 0.29 0.15 0.36
ELEVATOR 0.23 0.42 0.20 0.40
GYM 0.10 0.29 0.11 0.31
FAMILY_FRIENDLY 0.23 0.42 0.19 0.39
SMOKE_DETECTOR 0.46 0.50 0.55 0.50
SHAMPOO 0.35 0.48 0.45 0.50

2.3 Analysis of Property Images

2.3.1 Classifying Images as High- or Low-Quality
We build a supervised deep learning algorithm that classifies images as high- or low-quality.

Training set construction

We chose a random sample of images from our dataset to tag using AMT, a crowdsourcing platform for human
tasks. A stratified (by the crude metric of quality) random sample is necessary to create a sample that is balanced
as well as random. For each image, we asked the MTurkers (i.e., workers) to rate the image quality on a 1–7 Likert
scale, where 1 is “very bad” and 7 is “excellent.” For guidance, we provided examples of rated photos across the
quality spectrum, and we explained what we meant by “quality” (e.g., the image should be “visually pleasing” and
“clearly show room/house features”). Each image was evaluated by five qualified MTurkers. In the appendix, we
report details on image labeling using AMT. The image labels were further converted to a binary format (“high
quality” vs. “low quality”), following practices described in the computational aesthetics literature (Datta et al.
2006). The final training sample contained 1,155 high-quality images and 1,104 low-quality images.

Training step
CNN Approach: We apply a CNN, a deep learning framework widely applied in the field of computer vision with
breakthrough performances on tasks including object recognition and image classification (Krizhevsky et al.
2012). Our CNN image quality classifier (see Figure 2) represents the architecture of a classic CNN model. The
CNN consists of a sequence of neural layers (also called filters); the first layer extracts features from the input
image and summarizes them into an intermediate output, which becomes the input with which the next layer
generates a higher-level summary, and so on. The key component in the CNN is a convolution kernel (or
convolution filter), represented by an n by n weighting matrix. Given the intermediate output from the previous
layer, the convolution kernel extracts features through a matrix dot product operation between the weighting
matrix and the intermediate output. In other words, the sequence of layers in the CNN extracts a hierarchical set
of image features from the input. A training set (in which the label, “high-quality” or “low-quality,” is already
known for each image) enables the model to learn the relationships between the extracted features and the labels
such that the model can extract the features that have the most discriminative power for predicting the labels.
To reduce the overfitting problem in the training step, we increase the size of and random variation within the
training sample by randomly applying one of three transformations to each image (e.g., data augmentation, see

10

Electronic copy available at: https://ssrn.com/abstract=2976021

Krizhevsky et al. 2012): (1) flipping the input image horizontally, (2) rescaling the input image within a scale of
1.2, or (3) rotating the image within 20˚.

Transfer Learning and Fine-Tuning the Parameters of the CNN: Because a deep learning model has many filters and
parameters, it requires a large quantity of data to train the model. We overcome our limited training data by
leveraging transfer learning. We start with the pre-trained parameters of the widely-applied CNN model, VGG-
16 (trained on over one million images, Simonyan and Zisserman 2015) and fine-tune the parameters with 80%
of our training set of Airbnb property images. We use the remaining 20% as a hold-out sample to test the
performance of the trained CNN, and we achieve 90.4% accuracy). The high accuracy in predicting image quality
ensures a valid interpretation of our results of image quality in the demand model. In the appendix, we provide
a detailed description of the CNN architecture and technical notes on the training process.

Prediction step
Once the CNN classifier learns the relationship between image features and the image label, it can be applied to
the unlabeled images (Figure 2). The classifier takes the unlabeled image as the input, extracts the hierarchical set
of image features using the parameters of the trained classifier, and outputs the predicted label: either “1” for a
high-quality image or “0” for a low-quality image.
Figure 2. Description of Architecture and Layer Description of the CNN Classifier

Filters: The number of convolution windows (i.e., number of feature maps) on each convolution layer.
Zero-padding: Pads the input with zeros on the edges to control the spatial size of the output. Zero-padding has no
impact on the predicted output.
Max-pooling: Subsampling method. A 2 × 2 window slides through (without overlap) each feature map at that layer,
and the maximum value in the window is picked as the representation of the window. This reduces computation and
provides translation invariance.

2.3.2 Classifying the Image Room Type

We build another deep learning model to automatically classify the room in a property image as a bathroom,
bedroom, kitchen, living room, or outdoor area. Again, we use transfer learning—we start with Places205, which
was pre-trained on a large scene classification dataset (Zhou et al. 2014), and we fine-tune the classifier on our
own training dataset (i.e., with known labels) that we collected from real-estate-related websites on which a vast
number of indoor/outdoor images are classified into categories: 54,557 images of “bathrooms,” 59,082 images
of “bedrooms,” 88,030 images of “kitchens,” 81,819 images of “living rooms,” and 5,734 images of “outdoor
areas.” The trained classifier achieves 95.05% accuracy on the hold-out sample. We then apply the trained
11

Electronic copy available at: https://ssrn.com/abstract=2976021

classifier to the unlabeled property images. In the appendix, we provide a description of the training set and
technical notes for the training step of the room-type classifier.

3. Methods and Results

We implement a DiD analysis (Heckman et al. 1997) with PSW to mitigate the endogeneity concern.

3.1 DiD Analysis

The DiD analysis requires the identification of a treatment group and (comparable) control group. Our treatment
group consists of the 221 properties that did not have verified photos in January 2016 but acquired at least one
verified photo by April 2017. The control group consists of the 7,211 properties that had only unverified photos
during the observation window.

In an ideal setting, the two groups would be comparable such that the impact of treatment (i.e., the adoption
of verified photos) on demand would be reflected by the demand difference in the post-treatment period. In our
unrandomized (i.e., self-selecting) setting, however, the treatment is endogenous, and the two groups may not be
comparable. As shown in Table 1, the treatment and control groups differ on some pre-treatment covariates. If
certain differences affect both property demand and hosts’ decisions about whether to join the photography
program, then we cannot simply attribute any observed difference in the property demand to the treatment
(Athey and Imbens 2006). To mitigate the endogeneity concern, we use the PSW method to create groups that
are sufficiently comparable on the observed characteristics.

3.2 The PSW Method

The propensity score is the probability that an individual unit receives treatment, conditional on a set of observed
covariates (Rosenbaum and Rubin 1983). Propensity scores can be used to “balance” samples and are widely
used to make two groups comparable with respect to covariates (Rosenbaum 2002).
In practice, the true propensity scores are often unknown, so the PSW method estimates propensity scores
by modeling the treatment probability as a function of the observed covariates. The propensity score of unit i,
psi, is computed via a specified function, 𝑝𝑠 = 𝑓(𝑋 𝛽), where 𝑋 is a 1 * M dimensional vector of pre-treatment
observed covariates of unit i and 𝛽 is an M * 1 dimensional vector of parameters for X.
The model finds a set of parameters that maximizes the treatment likelihood in the sample data (Rosenbaum
and Rubin 1983). We estimate the parameters via logistic regression because the treatment assignment is a binary
variable (i.e., treatment or control). The selection of X is based on a covariate balance check (see the appendix).
The differences in the means of covariates between the treatment and control groups should be minimized. With
parameter vector 𝛽 estimated, we approximate for each unit i (with observed covariates 𝑋 ) the propensity score
𝑝𝑠 (𝑋 ), which is used to compute the sample weights.

In the X of the PSW model, we include covariates to control for the self-selection issue identified by Zhang
et al. (2019)—that a rational host should adopt high-quality images if they have an appropriately high-quality

12

Electronic copy available at: https://ssrn.com/abstract=2976021

property and level of service. We control for the service quality through the Response Rate and Response Time (similar
to Zhang et al. 2019) and capture the property quality with a rich set of property characteristics such as property
amenities (internet, gym, beach, pool, etc.), property size and type, and host quality (super host).

Computing Sample Weights Based on Propensity Scores

We use propensity scoring for a weighting strategy—the inverse probability of treatment weighting (IPTW)
method—in the DiD analysis (Austin and Stuart 2015). The IPTW method calculates a weight w for unit i by
taking the inverse of unit i's propensity score:

𝜔 (𝑇, 𝑋 ) =
𝑇
+
1−𝑇
,
(1)
𝑝𝑠 (𝑋 ) 1 − 𝑝𝑠 (𝑋 )

where 𝑝𝑠 (𝑋 ) is the estimated propensity score of unit i computed with its observed covariates 𝑋 , and T is a
dummy variable that equals 1 if is in the treatment group and 0 otherwise.
The PSW results are validated to ensure that the treatment and control groups in the weighted sample are
balanced on the included covariates (see the appendix for details on the PSW approach, results, and validation).
As shown in Table 2, the group means are statistically indifferent, suggesting that the PSW method successfully
removed systematic differences in host and property observed characteristics that might confound the treatment
assignment. See Section V.5 in the appendix for the analysis details.

Table 2. PSW Validation: Covariate Balance Check

Weighted Means
T-test
Variables Treated Untreated t p-value
REVIEW_COUNT 20.56 19.88 0.27 0.790
IMAGE_QUALITY 0.27 0.25 1.00 0.316
IMAGE_COUNT 14.48 15.1 -0.67 0.506
NIGHTLY_RATE 170.15 191.36 -1.14 0.257
MINIMUM_STAY 2.57 2.57 -0.00 1.000
MAX_GUESTS 3.5 3.67 -0.82 0.410
RESPONSE_RATE 92.25 91.19 0.79 0.431
RESPONSE_TIME (minutes) 225.12 260.98 -1.18 0.238
SUPER_HOST 0.15 0.11 1.05 0.292
INSTANT_BOOK 0.11 0.11 -0.14 0.888
# BLOCKED DAYS 9.51 8.32 1.10 0.271
# RESERVATION DAYS 6.62 6.74 -0.16 0.877
PARKING 0.5 0.49 0.09 0.929
POOL 0.1 0.08 0.76 0.445
BEACH 0.02 0.02 0.34 0.737
INTERNET 0.99 1 -0.58 0.563
TV 0.79 0.81 -0.55 0.579
WASHER 0.6 0.57 0.81 0.419
MICROWAVE 0.15 0.13 0.64 0.523
ELEVATOR 0.2 0.21 -0.22 0.826
GYM 0.11 0.13 -0.69 0.490
FAMILY_FRIENDLY 0.19 0.2 -0.56 0.576

13

Electronic copy available at: https://ssrn.com/abstract=2976021

 SMOKE_DETECTOR 0.55 0.52 0.62 0.534
SHAMPOO 0.45 0.44 0.09 0.929
BATHROOM_PHOTO_RATIO 0.22 0.21 1.05 0.295
BEDROOM_PHOTO_RATIO 0.29 0.29 -0.26 0.795
KITCHEN_PHOTO_RATIO 0.1 0.1 -0.15 0.880
LIVINGROOM_PHOTO_RATIO 0.18 0.19 -0.52 0.602
OUTDOOR_PHOTO_RATIO 0.2 0.21 -0.13 0.898
SECURITY_DEPOSIT 202.77 225.19 -0.65 0.517
# OPEN DAYS 21.49 22.68 -1.10 0.271
CANCELLATION_STRICT 0.26 0.29 -0.89 0.371
# BEDS 1.81 1.96 -1.17 0.242
APARTMENT 0.6 0.61 -0.27 0.786
ENTIRE_HOME 0.61 0.64 -0.64 0.522

3.3 Model Specification and DiD Estimator

We obtain our DiD estimator through a weighted least squares (WLS) regression, which computes sampling
weights using estimated propensity scores. Let 𝐷𝐸𝑀𝐴𝑁𝐷 denote the demand for property i (in city c) in
year y and month m (further, let t denote month m in year y ):

𝐷𝐸𝑀𝐴𝑁𝐷 = 𝐼𝑁𝑇𝐸𝑅𝐶𝐸𝑃𝑇 + 𝛼𝑇𝑅𝐸𝐴𝑇𝐼𝑁𝐷 + 𝜆𝐶𝑂𝑁𝑇𝑅𝑂𝐿𝑆 + 𝑃𝑅𝑂𝑃𝐸𝑅𝑇𝑌 + 𝑆𝐸𝐴𝑆𝑂𝑁𝐴𝐿𝐼𝑇𝑌 +𝜀 , (2)

where 𝑇𝑅𝐸𝐴𝑇𝐼𝑁𝐷 is the treatment status indicator, which equals 1 if property i has received treatment in
period t and equals 0 otherwise. The key coefficient 𝛼 estimates the percentage change in property demand that
is associated with the treatment (i.e., having verified photos), and 𝜀 is a random shock in period t on property
i’s demand, assumed to follow an i.i.d. normal distribution. The vector 𝐶𝑂𝑁𝑇𝑅𝑂𝐿𝑆 represents a set of control
variables that may be correlated with property demand, for example, the property rules and consumer reviews. 6
We include the property fixed effect term, 𝑃𝑅𝑂𝑃𝐸𝑅𝑇𝑌 , to capture time-invariant factors that may impact
property demand, such as geographic information and property-specific characteristics. We also include the time
fixed effect term 𝑆𝐸𝐴𝑆𝑂𝑁𝐴𝐿𝐼𝑇𝑌 , which captures any city-specific trends in property demand (i.e., we allow
each city to have its own seasonal pattern).
Note that the key outcome variable in this study is property demand, operationalized as the monthly
occupancy rate. As an empirical extension, in the appendix, we analyze whether the property price changed with
treatment (i.e., we replace the DV in Eq. 2 with the property’s nightly rate). We find that after controlling for

6 The vector CONTROLS includes two metrics that measure hosts’ responsiveness, RESPONSE_RATE (percentage of
messages/requests from guests that receive a response from the host) and RESPONSE_TIME (average number of minutes
to respond to a guest); MIN_STAYS, the minimum number of nights that a guest can book; MAX_GUESTS, the maximum
number of guests that may stay at once; SECURITY_DEPOSIT, the money that the guest will be charged if the host claims
that the guest damaged the property and Airbnb approves the claim; CANCELLATION_STRICT, whether the cancellation
policy is strict (1) or not strict (0); SUPER_HOST, whether the host has a “super host” badge (1) or not (0), which Airbnb
assigns based on consumers’ reviews, the host’s responsiveness, etc.; BUSINESS_READY, whether the property has
business-related amenities (1) or not (0); HAS_RATING, whether the average guest ratings are presented on the property
page (1) or not (0); and the interaction terms of HAS_RATING and the multi-dimensional ratings.
14

Electronic copy available at: https://ssrn.com/abstract=2976021

seasonality, property characteristics, and time-varying variables such as the number of reviews, the coefficient of
the treatment on property price is insignificant (see Section V8 in the appendix for detailed discussion). 7
As shown in Table 3, the estimated coefficient of the key variable TREATIND suggests a positive, significant
treatment effect. Specifically, properties that used verified photos had 8.985% higher occupancy than the control
group. On average, an untreated property was open 18.1 days per month, so the treatment coefficient
corresponds to an average of 18.1 days/month * 8.985% * 12 months/year = 19.5 additional booked days in a
year (or 1.6 additional days in a month). In terms of revenue, the acquisition of verified photos corresponds to
an average of $179.5/day * 19.5 days = $3,500.30 more per year. 8

Table 3. DiD Model: Regressing Property Demand on Verified Photos

VARIABLES Main DiD Model (Eq. 2)

ESTIMATES Robust S.E.

TREATIND 8.985*** 1.660
log REVIEW_COUNTt-1 9.375*** 0.930
NIGHTLY_RATE -0.146*** 0.0320
INSTANT_BOOK 4.156** 1.361
CLEANING_FEE 0.0808*** 0.0184
MAX_GUESTS 0.260 1.117
RESPONSE_RATE 0.0699 0.0430
RESPONSE_TIME (minutes) -0.000477 0.00161
MINIMUM_STAY 0.133 0.131
SECURITY_DEPOSIT 0.00177 0.00201
SUPER_HOST 3.801* 1.494
BUSINESS_READY 1.806 0.985
CANCELLATION_STRICT 1.016 1.271
HAS_RATING 14.32 12.25
HAS_RATING × COMMUNICATION -0.212 1.420
HAS_RATING × ACCURACY 0.878 1.211
HAS_RATING × CLEANLINESS -1.344 1.133
HAS_RATING × CHECKIN -2.060 1.526
HAS_RATING × LOCATION -0.757 1.183
HAS_RATING × VALUE 2.141 1.176
Fixed Effect Property
Seasonality City-Year-Month
Num. Observations 76,901
R-squared 0.6608
Note: The demand model does not include the number of open nights because the D.V. is computed as a
function of it (occupancy rate = #reservation days/#open days). This variable is used in the matching step
(see Table A1 in the appendix). *p < 0.05, **p < 0.01, ***p < 0.001.

7 A possible explanation is that a change in photo quality does not necessarily reflect changes in the property or host. In a
report on the smart pricing model, Airbnb did not describe property pictures as one of the many “factors at play” that the
algorithm used: https://blog.atairbnb.com/smart-pricing/.
8 The properties in our sample in the pre-treatment period (January 2016) were priced at $ 179.5 per night on average.

15

Electronic copy available at: https://ssrn.com/abstract=2976021

3.4 Validating the DiD Model
We implement a set of analyses to validate our combination of the DiD model and the PSW strategy. We begin
with a falsification check that examines the critical “common pre-treatment trends” assumption, followed by a
Rosenbaum bounds analysis for the selection on unobservables and several additional robustness checks.

3.4.1 Falsification Check: Pre-Treatment Trends

The validity of the DiD approach (Eq. 2) relies on a critical assumption of common pre-treatment trends: that
is, the two weighted groups should exhibit common trends in their demand before treatment (Angrist and Pischke
2008). We test the common trends assumption using a relative-time model of pre-treatment periods. Following
the extant literature (e.g., Wang and Goldfarb 2017), we decompose the pre-treatment periods into a series of
dummy variables:
𝐷𝐸𝑀𝐴𝑁𝐷 = 𝐼𝑁𝑇𝐸𝑅𝐶𝐸𝑃𝑇 + 𝛼𝑇𝑅𝐸𝐴𝑇𝐼𝑁𝐷 + ∑ 𝛽 (𝑃𝑅𝐸 (𝑗) ∙ 𝑇𝑅𝐸𝐴𝑇 ) + 𝜆𝐶𝑂𝑁𝑇𝑅𝑂𝐿𝑆 + (3)
𝑆𝐸𝐴𝑆𝑂𝑁𝐴𝐿𝐼𝑇𝑌 + 𝑃𝑅𝑂𝑃𝐸𝑅𝑇𝑌 + 𝜀 .

We define PRE(1) as the period prior to the treatment month and set it as the reference period (i.e., we
normalize its coefficient to zero). Then, PRE(2) is two months prior to treatment, PRE(3) is three months prior
to treatment, and PRE(4) represents all the periods from the beginning (i.e., January 2016) through four months
prior to treatment (Autor 2003). For properties with fewer than four pre-treatment periods (e.g., a property that
acquired verified photos in February 2016 would have only one pre-treatment period), the period dummies that
correspond to months before January 2016 are set to zero.
The set of coefficients 𝛽 allows us to validate the DiD model by comparing the trends in the property
demand of the weighted control and treatment groups prior to treatment. The common trends assumption is
validated if there is no significant pre-treatment difference between the groups, and Table 4 confirms that none
of 𝛽 the period dummy coefficients are statistically different from zero. Moreover, the set of 𝛽 does not exhibit
an increasing trend in the property demand for the treatment units compared with the control units. In other
words, the demand for the treated units was not already deviating from the demand for the control units prior
to treatment, which would suggest that the difference between groups was caused by an idiosyncratic shock that
affected both the treatment likelihood and property demand rather than by the treatment itself.
Table 4. Falsification Check: A Relative-Time Model of Pre-Treatment Trends

Relative-Time Model (Eq. 3)

VARIABLES ESTIMATES Robust S.E.
PRE (4) * TREAT 2.230 2.503
PRE (3) * TREAT -0.894 3.164
PRE (2) * TREAT 1.302 3.102
PRE (1) * TREAT (reference month) -- --
TREATIND 9.988*** 2.311
Property (Non-Photo) Characteristics
log REVIEW_COUNTt-1 9.338*** 0.933
NIGHTLY_RATE -0.147*** 0.0320
INSTANT_BOOK 4.166** 1.363
CLEANING_FEE 0.0811*** 0.0184
16

Electronic copy available at: https://ssrn.com/abstract=2976021

 MAX_GUESTS 0.199 1.108
RESPONSE_RATE 0.0730 0.0431
RESPONSE_TIME (minutes) -0.000403 -0.00161
MINIMUM_STAY 0.132 0.131
SECURITY_DEPOSIT 0.00175 0.002
SUPER_HOST 3.764* 1.495
BUSINESS_READY 1.821 0.985
CANCELLATION_STRICT 1.05 1.271
HAS_RATING 14.57 12.22
HAS_RATING × COMMUNICATION -0.229 1.421
HAS_RATING × ACCURACY 0.978 1.21
HAS_RATING × CLEANLINESS -1.403 1.134
HAS_RATING × CHECKIN -2.117 1.521
HAS_RATING × LOCATION -0.716 1.182
HAS_RATING × VALUE 2.118 1.177
Fixed Effect Property
Seasonality City-Year-Month
Num. Observations 76,901
R-squared 0.6609
Note: *p < 0.05, **p < 0.01, ***p < 0.001.

3.4.2 Selection on Unobservables

The PSW method addresses endogeneity concerns regarding observed variables, but it does not rule out a
possible hidden bias involving unobserved variables that might influence the treatment likelihood and outcome
variable simultaneously. We assess the sensitivity of our estimation to a potential hidden bias with the Rosenbaum
bounds test (Rosenbaum 2002), which determines how much of a change in the odds ratio of treatment would
need to be due to unobservables to nullify the treatment effect identified by the PSW method. We can be more
confident about the validity of the estimation results if the Rosenbaum bounds test suggests that unobservables
would have to cause a large change in the odds ratio to overturn the estimated treatment effect.
Our results for the Rosenbaum bounds test with an examination of the Hodges-Lehmann estimates
(Rosenbaum 1993) suggest that unobservables would need to increase the odds ratio of treatment by at least 55%
(i.e., gamma≥1.55) to overturn the positive estimated treatment effect on property demand. The results of our
sensitivity analysis are similar to the results obtained in the extant literature (DiPrete et al. 2004; Sun and Zhu
2014; Manchanda et al. 2015; Li et al. 2016), which reported gamma values ranging from 1.2 to 1.6. This suggests
that our study is robust to the hypothetical unobserved factors that may affect the treatment likelihood. We
provide detailed methods and results in the appendix.

3.4.3 Additional Robustness Tests

In addition to the falsification test (Section 3.4.1) and the Rosenbaum analysis (Section 3.4.2), we perform a
comprehensive set of seven tests that verify the robustness of our main findings. We briefly describe the tests
below; detailed descriptions and results are in Section V in the appendix.

First, we run two analyses to test for potential inflation in the treatment coefficient in the long term, which
could occur if Airbnb’s ranking algorithm favors properties with verified images. We estimate our main DiD

17

Electronic copy available at: https://ssrn.com/abstract=2976021

specification on a set of subsamples in which we include a shorter period (four months post-treatment) for each
treated unit. The estimated coefficient of TREATIND remains significant and positive. We also use a relative-
time model to estimate the treatment effect in each month following treatment. We find that the estimated
coefficient of TREATIND in the month following treatment is close to our main finding. The two analyses add
confidence that our main finding was not driven by long-term inflation.
Second, we test whether properties with particular amenities were more likely to acquire verified photos at a
time when these amenities were most attractive (e.g., a pool in the summer). We include interaction terms for the
dummy AFTER and meaningful amenities (pool, beach, AC) in the DiD model and find consistent estimated
results. The results suggest that the positive treatment coefficient is not due to higher demand for these amenities
following treatment or in particular seasons.
Third, we investigate whether the positive treatment effect was driven by unobserved enhancements of the
property or host. We regress the multidimensional review ratings (communication, cleanliness, etc.) on
TREATIND. The coefficient of TREATIND is statistically insignificant, suggesting no substantial change in the
quality of guests’ experiences after the acquisition of verified photos.
Fourth, we examine whether hosts in the treatment group changed their management of the property calendar
(e.g., made the property open for more nights per month) at the same time as they adopted verified photos. We
regress #_OPEN_DAYS on TREATIND and other controls and find an insignificant coefficient of
TREATIND, suggesting that the main results are not due to an increase in property availability after treatment.
Additionally, we include #_OPEN_DAYS in demand model and have similar finding of estimated TREATIND.
Fifth, we replicate the DiD model with a log transformation of the numeric variables and find no change in
the estimated treatment coefficient.
Sixth, Zhang et al. (2019) find that image quality affects the review probability, and the number of reviews
affects property demand. We incorporate the probability that a guest writes a review for the property. In this test,
we match properties based on review probability as well as other covariates. The DiD estimation results on the
matched sample are consistent with our main results, and the estimated coefficient of review probability. 9
Seventh, we use a different control group: the pre-treated units (i.e., those that adopted verified photos before
January 2016 and thus were excluded from the main analyses). The pre-treated units may represent a stronger
control because the pre-treated and to-be-treated properties might be more similar in the characteristics that drive
the treatment likelihood (Narang and Shankar 2019). Two analyses report estimated coefficients of TREATIND
that are consistent with our main finding.

9 We do not include the review writing probability in the main DiD model as: 1) review writing probability is indirectly
controlled for through number of reviews and property characteristics, and 2) including it will significantly reduce the sample
size. See Section V.9) for detailed discussions.
18

Electronic copy available at: https://ssrn.com/abstract=2976021

3.5 Examining the Aesthetic Qualities of Verified vs. Unverified Photos

Our main analyses in Section 3.3 established that the demand for properties with verified images was 8.98%
higher than the demand for properties with unverified images. In this section, we explore the potential sources
of the positive coefficient of verified photos correlationally, by extracting image aesthetic quality. We analyze the
images and then estimate a demand equation (see Eq. 4) that captures three aspects of a property’s image set:
IMAGE_COUNT (the number of property images), IMAGE_QUALITY (the average of the binary labels of
“high-quality” [1] and “low-quality” [0]), and the distribution of the five main types of photographed rooms. In
Section 2.3 and the appendix, we define the variables and provide technical details regarding their measurements.
𝐷𝐸𝑀𝐴𝑁𝐷 = 𝐼𝑁𝑇𝐸𝑅𝐶𝐸𝑃𝑇 + 𝛼𝑇𝑅𝐸𝐴𝑇𝐼𝑁𝐷 + 𝜇𝐼𝑀𝐴𝐺𝐸_ 𝐶𝑂𝑈𝑁𝑇 + 𝛾𝐼𝑀𝐴𝐺𝐸 𝑄𝑈𝐴𝐿𝐼𝑇𝑌
+ 𝜌 𝐵𝐴𝑇𝐻𝑅𝑂𝑂𝑀 𝑃𝐻𝑂𝑇𝑂 𝑅𝐴𝑇𝐼𝑂 + 𝜌 𝐵𝐸𝐷𝑅𝑂𝑂𝑀 𝑃𝐻𝑂𝑇𝑂 𝑅𝐴𝑇𝐼𝑂 (4)
+ 𝜌 𝐾𝐼𝑇𝐶𝐻𝐸𝑁 𝑃𝐻𝑂𝑇𝑂 𝑅𝐴𝑇𝐼𝑂 + 𝜌 𝐿𝐼𝑉𝐼𝑁𝐺 𝑃𝐻𝑂𝑇𝑂 𝑅𝐴𝑇𝐼𝑂 + 𝜆𝐶𝑂𝑁𝑇𝑅𝑂𝐿𝑆
+ 𝑆𝐸𝐴𝑆𝑂𝑁𝐴𝐿𝐼𝑇𝑌 + 𝑃𝑅𝑂𝑃𝐸𝑅𝑇𝑌 + 𝜀
Table 5 reports the estimation results; the model specification does not include IMAGE_QUALITY in
column (1) and does include it in column (2). Recall that the estimated coefficient of the key variable TREATIND
was 8.985 in the original model; the coefficient decreases to 7.453 in column (1) with the inclusion of
IMAGE_COUNT, which has a positive and significant coefficient (none of the room type ratios have a
significant coefficient). The treatment coefficient decreases by another 41% (from 7.453 to 4.397) in column (2)
with the inclusion of IMAGE_QUALITY, which has a positive and significant coefficient. This significant
reduction suggests that the high quality of the verified images explains some but not all of the treatment effect,
as there is a significant residual treatment coefficient even when we control for both covariates.

Table 5. DiD Model: Controlling for the Number and Quality of Property Images

VARIABLES (1) (2)

Without Image Quality Including Image Quality
ESTIMATES Robust S.E. ESTIMATES Robust S.E.
TREATIND 7.453*** 1.777 4.397* 2.190
Property (Non-Photo) Characteristics
log REVIEW_COUNTt-1 9.754*** 0.944 9.570*** 0.942
NIGHTLY_RATE -0.183*** 0.0325 -0.187*** 0.0325
INSTANT_BOOK 3.768** 1.357 3.664** 1.349
CLEANING_FEE 0.0931*** 0.0188 0.0955*** 0.0187
MAX_GUESTS 0.247 1.098 0.285 1.094
RESPONSE_RATE 0.0868* 0.0427 0.0886* 0.0427
RESPONSE_TIME (minutes) -0.000232 0.00159 -0.000151 0.00160
MINIMUM_STAY 0.172 0.132 0.171 0.133
SECURITY_DEPOSIT 0.00211 0.00201 0.00210 0.00200
SUPER_HOST 3.999** 1.495 3.890** 1.494
BUSINESS_READY 1.805 0.977 1.813 0.974
CANCELLATION_STRICT 1.282 1.277 1.308 1.278
HAS_RATING 15.82 12.27 14.96 12.28
HAS_RATING × COMMUNICATION 0.0499 1.418 0.0702 1.423

19

Electronic copy available at: https://ssrn.com/abstract=2976021

 HAS_RATING × ACCURACY 0.588 1.209 0.681 1.206
HAS_RATING × CLEANLINESS -1.214 1.133 -1.144 1.141
HAS_RATING × CHECKIN -2.260 1.525 -2.267 1.516
HAS_RATING × LOCATION -0.785 1.186 -0.745 1.193
HAS_RATING × VALUE 2.124 1.176 2.023 1.182
AFTER × POOL 5.841 4.610 6.053 4.608
AFTER × BEACH -11.29 10.91 -10.39 10.66
AFTER × AC 0.331 3.978 -0.198 3.963
Property Image Characteristics (+)
log IMAGE_COUNT 6.874*** 1.724 5.518** 1.777
BATHROOM_PHOTO_RATIO 0.777 8.457 0.779 8.439
BEDROOM_PHOTO_RATIO -2.575 7.539 -2.128 7.499
KITCHEN_PHOTO_RATIO 18.04 11.27 17.20 11.22
LIVINGROOM_PHOTO_RATIO -12.07 8.430 -12.41 8.405
IMAGE_QUALITY 8.984** 3.371
Fixed Effect Property Property
Seasonality City-Year-Month City-Year-Month
Num. Observations 76,901 76,901
R-squared 0.6623 0.6628
Note: + The coefficient for OUTDOOR_PHOTO_RATIO is not estimated; *p < 0.05, **p < 0.01, ***p < 0.001.

3.6. What Makes a Property Image Appealing?

Our CNN classifier predicted image quality with high accuracy, but the CNN-generated features are
uninterpretable. That is, the features do not provide any insights for managers or photographers regarding how
to improve their images based on CNN generated features. Furthermore, Airbnb property images may be
appealing for more reasons than image quality. Image quality captures only the vertical differentiation among
images, but images can also differ horizontally in terms of taste. In an effort to provide actionable insights, we
evaluate the relationship between property demand and 12 key image attributes identified in the photography
literature. We first explain these features and then theorize their impact on property demand.

3.6.1 Interpretable Image Features

Multiple image attributes may affect consumers’ perceptions and choices. To investigate which image attributes
correlate most strongly with Airbnb property demand, this section defines 12 key dimensions along which
photographs can be compared and categorized. We use the art and photography literature to define the image
attribute dimensions that are most relevant to property photos. Then, drawing from studies in this literature that
pertain to the role of images in viewer perception, we theorize how each attribute could affect property demand.

The photography literature highlights 12 image attributes across three components: composition, color, and
the figure-ground relationship (Datta et al. 2006; Freeman 2007; Wang et al. 2013). The features affect both the
vertical quality of the photographs and horizontal qualities that might affect a property’s appeal to potential guests
who browse the images.

20

Electronic copy available at: https://ssrn.com/abstract=2976021

Component: Composition
Composition is the way in which visual elements are arranged, which guides the viewer’s eyes to certain elements
of the photograph (Freeman 2007). Expert photographers compose images that enable viewers to quickly identify
the element that is the center of focus (Grill and Scanlon 1990). The most appropriate compositional attribute
depends on the context; the following three compositional techniques are relevant for real estate photography.

Attribute 1: Diagonal Dominance. A photographer can guide the viewer’s eyes with leading lines, and in a
rectangular frame, the two diagonals of the photograph are the longest possible straight lines. In a diagonally
dominant photograph, the most salient visual elements are placed close to the two diagonals (Grill and Scanlon
1990). Diagonal dominance creates a perception of spaciousness, so we predict a positive relationship between
diagonal dominance and property demand. In Figure 3, it is likely that viewers will perceive that the image on the
right represents a larger room than the one on the left.
Figure 3. Comparing Images on Diagonal Dominance

Image without Diagonal Dominance Image with Diagonal Dominance

Attribute 2: Rule of Thirds (ROT). An image can be divided into nine equal parts with (imaginary) horizontal
and vertical third lines. The ROT states that the main visual elements should be placed along the imaginary third
lines or close to the four intersections (Krages 2005). These off-center focal points introduce movement into the
photograph, engaging the viewer and making the image aesthetically pleasing and dynamic (Meech 2004), so we
predict a positive relationship between the use of the ROT and property demand. In Figure 4, the image on the
right follows the ROT better than the image on the left. Hence, when looking at the image, the viewer’s attention
first goes to the bed and then its counterpoint—the other vertical third line. By contrast, in the image on the left,
the focus and key objects are not obvious.
Figure 4. Comparing Images on the ROT

Image That Does Not Follow the ROT Image That Follows the ROT

21

Electronic copy available at: https://ssrn.com/abstract=2976021

Attributes 3 and 4: Visual Balance of Intensity and Visual Balance of Color. Visual balance is the symmetry
of visual elements (in this case, intensity and color) within an image (Krages 2005). The more symmetrical the
elements within the image, the greater the visual balance; the extreme case being perfect symmetry. Humans
subconsciously consider visual balance to be aesthetically pleasing, and symmetry increases visual interest
(Arnheim 1974; Bornstein et al. 1981). Visually balanced real estate images give viewers a feeling of order and
tidiness, minimizing the cognitive demand required to process the images (Kreitler and Kreitler 1972; Machajdik
and Hanbury 2010). We predict a positive relationship between visual balance and property demand. In Figure
5, the image on the right is more visually balanced than the one on the left, so the image on the right can be
processed very quickly and provides a sense of order and cleanliness.

Figure 5. Comparing Images on Visual Balance

Image without Visual Balance Image with Visual Balance

Component: Color
Color, one of the most significant elements in photography, affects the viewer’s emotional arousal. Building on
past research, Gorn et al. (1997) identified two dimensions of arousal: from boredom to excitement, and from
tension to relaxation. Excitement is preferred to boredom, and relaxation is preferred to tension. Three
dimensions of color—hue, saturation (chroma), and brightness (value)––can affect the level of arousal. Each
dimension has been widely studied in the marketing literature, particularly in the contexts of web design, product
packaging design, and advertisement design (Gorn et al. 2004; Miller and Kahn 2005). We add another attribute,
image clarity, that is affected by the combination of the aforementioned three.

Attribute 5: Warm Hue. Hues (e.g., red, green) are believed to be a major driver of emotion. The warmth in an
image is affected by the actual colors of the subject, but the photographer can also manipulate hue by warming
up or cooling down a picture during post-processing. Warm hues (such as red and yellow) elicit higher levels of
excitement (Gorn et al. 2004; Valdez and Mehrabian 1994), while cool hues (such as blue and green) elicit higher
levels of relaxation. Hence, we predict a positive relationship between warm hues and property demand. In Figure
6, we present a cool photo of a living room on the left and a warm photo of a living room on the right.

22

Electronic copy available at: https://ssrn.com/abstract=2976021

 Figure 6. Comparing Images on Warm Hue

Image with Cool Colors Image with Warm Colors

Attribute 6: Saturation. Saturation refers to the richness of color. Highly saturated images are colorful, while
weakly saturated images contain low levels of pigmentation. Saturation is positively associated with happiness
and purity; less-saturated colors are associated with sadness and distress (Valdez and Mehrabian 1994; Gorn et
al. 2004). Thus, we predict that real estate images with saturated colors can induce positive emotions in viewers
and that there is a positive relationship between saturation and property demand. To illustrate the difference in
emotional arousal, Figure 7 displays two images of the same room with different levels of saturation.

Figure 7. Comparing Images on Saturation

Image with Low Saturation Image with High Saturation

Attributes 7 and 8: Brightness and the Contrast of Brightness. The photography literature identifies two
image attributes regarding image illumination: brightness and its contrast. Brightness is the level of overall
illumination; viewers prefer bright images as they induce a sense of relaxation but do not affect the level of
excitement (Valdez and Mehrabian 1994; Gorn et al. 1997). Furthermore, sufficient illumination makes the
content of an image clear to viewers because images convey information through pixel brightness, so we predict
a positive relationship between brightness and property demand. Meanwhile, the contrast of brightness is the
variance in the illumination and describes whether the illumination is evenly distributed over the image with a
smooth flow. In other words, a low contrast of brightness indicates an even distribution of illumination across
the photograph. An uneven distribution (i.e., high contrast) of brightness may induce a feeling of harshness, so
we predict a negative relationship between the contrast of brightness and property demand. For example, in

23

Electronic copy available at: https://ssrn.com/abstract=2976021

Figure 8, the image on the right has a higher level of brightness and more uniform illumination (i.e., lower contrast
of brightness) than the image on the left.
Figure 8. Comparing Images on Illumination (Brightness and its Contrast)

Image with Low Brightness and High Contrast Image with High Brightness and Low Contrast

Attribute 9: Image Clarity. Clarity reflects the intensity of hues in the HSV (i.e., hue, saturation, value) space
(Levkowitz and Herman 1993). An image is “dull” if it is dominated by desaturated colors or has near-zero hue
intensities in some color channels (He et al. 2011). Amateur photographers often shoot dull photos, inducing a
so-called haze effect in which some parts of the image look unclear and ill-focused. By contrast, images with
“clear” color reduce the friction in information transfer, so we predict a positive relationship between image
clarity and property demand. In Figure 9, the photo on the right has higher clarity than the one on the left.
Figure 9. Comparing Images on Clarity

Image with Dull Color Image with Clear Color

Component: The Figure-Ground (FG) Relationship

Attributes 10, 11, and 12: Area Difference, Color Difference, and Texture Difference. The FG relationship
within an image is evaluated in relation to three aspects—area, color, and texture. The principle of the FG
relationship is one of the most basic laws of perception and is used extensively by expert photographers to plan
their photographs. In visual art, the figure refers to the key region (i.e., foreground), and the ground refers to the
background. The FG relationship describes the separation between the figure and ground. Gestalt theory states
that objects that share visual characteristics, such as size, color, and texture, are seen as belonging together
(Arnheim 1974). Hence, the figure is more salient if it differs from the ground in size, color, and texture. In
advertising research, consumers pay more attention to images with clear FG relationships (Larsen et al. 2004;
24

Electronic copy available at: https://ssrn.com/abstract=2976021

Schloss and Palmer 2011), and we predict a positive relationship between FG separation and property demand.
Figure 10 presents one set of images in which the figure is clearly separable from the ground and another set in
which the separation is not obvious.

Figure 10. Comparing Images on the Figure-Ground Relationship

10a. Clear Separation of Figure from Ground

10b. Unclear Separation of Figure from Ground

3.6.2 Measurement of Image Attributes and Related Statistics

Table 6 summarizes and briefly describes the 12 attributes.
Table 6. The 12 Image Attributes and Their Descriptions

COMPONENT ATTRIBUTE DESCRIPTION

1 Diagonal Dominance Alignment of the key objects with the diagonals
2 Visual Balance-Intensity Symmetry of key objects around the imaginary vertical central
Composition line
3 Visual Balance-Color Symmetry of colors around the imaginary vertical central line
4 Rule of Thirds Alignment of the key objects with the intersections of two
imaginary horizontal and two vertical lines
5 Warm Hue Dominance of warm colors (e.g., red and yellow) over cold
colors (e.g., blue and purple)
6 Saturation The richness/vividness of image colors
Color 7 Brightness Overall level of illumination
8 Contrast of Brightness The variance in the illumination flow across the image.
9 Image Clarity Whether image colors have sufficient intensity
Figure-Ground 10 Size Difference Difference in area between the figure and ground
Relationship 11 Color Difference Difference in color between the figure and ground
12 Texture Difference Difference in texture between the figure and ground

25

Electronic copy available at: https://ssrn.com/abstract=2976021

Statistics of Image Attributes
We measure the 12 attributes using computer vision algorithms; detailed methods appear in the appendix. In
general, an image-processing task involves the segmentation of an image into patches, detection of salient regions,
and computation involving those regions. We compare the attributes across three groups of property images:
Group LQ: All low-quality images (all unverified; N = 368,626)
Group HQ_UN: All high-quality unverified images (N = 69,380)
Group HQ_V: All verified images (all high-quality; N = 72,608)
The first three columns of Table 7 present the means (standard deviations) for the image attributes by group.
The fourth (rightmost) column presents the difference between the means of the HQ_UN and HQ_V groups
along with the two-sample t-statistics in parentheses; statistically significant differences (at a 5% significance level)
are bolded.

The low-quality images rate lower than the high-quality images on all image attributes except for the contrast
of brightness, which is theorized to have a negative relationship with the property demand. More interestingly,
the unverified high-quality images also perform significantly worse than the verified high-quality images on most
attributes. We conclude that there is a systematic difference in the high-quality images taken by Airbnb
photographers versus by other photographers.

Table 7. Summary Statistics: Mean (Standard Deviation) of Image Attributes of Verified vs. Unverified
High-Quality Images

COMPONENT IMAGE LQ HQ_UN HQ_V HQ_V

ATTRIBUTE Low Quality High-Quality High-Quality V.S.
Unverified Verified HQ_UN
368,626 Obs. 69,380 Obs. 72,608 Obs.
Mean Mean Mean Difference
(Std. Dev.) (Std. Dev.) (Std. Dev.) (t-statistic)
Diagonal Dominance -0.342 -0.281 -0.236 0.045***
(0.160) (0.109) (0.081) (88.56)
Visual Balance of -0.865 -0.774 -0.757 0.017***
Composition10 Intensity (0. 110) (0.103) (0.105) (30.78)
Visual Balance of Color -59.281 -53.093 -50.096 2.997***
(19.460) (15.509) (15.070) (36.93)
Rule of Thirds -0.147 -0.089 -0.089 0.0003
(0.082) (0.045) (0.047) (1.23)
Warm Hue 0.738 0.751 0.789 0.038***
(0.230) (0.208) (0.181) (36.77)
Saturation 59.023 73.942 73.683 -0.259
(37.528) (31.300) (26.929) (0.87)
Color Brightness 136.029 154.212 175.802 21.590***
(32.488) (27.558) (22.593) (161.75)

10Note that the composition measurements are negative because they reflect distances, and we subtract all distances from
zero to preserve the absolute magnitude when the direction is reversed. A higher value (i.e., less negative) suggests better
performance on that composition attribute. For example, a higher value of diagonal dominance suggests that the image is
more diagonally dominant.
26

Electronic copy available at: https://ssrn.com/abstract=2976021

 Contrast of Brightness 60.601 58.029 53.996 -4.033***
(13.628) (13.056) (12.990) (58.33)
Image Clarity 0.324 0.413 0.595 0.182***
(0.232) (0.217) (0.195) (166.38)
Size Difference -0.405 -0.181 -0.140 0.041***
Figure-Ground (0.181) (0.188) (0.153) (45.16)
Relationship Color Difference 23.090 33.054 39.063 6.009***
(20.056) (17.552) (15.580) (68.29)
Texture Difference 0.043 0.057 0.059 0.002***
(0.033) (0.026) (0.018) (16.92)
Note: SD in parentheses for the first three columns; t-statistics in parentheses for the rightmost column; *p < 0.05,
**p < 0.01, ***p < 0.001.

3.6.3 Relating the Interpretable Image Attributes to Property Demand

The 12 human-interpretable attributes have well-studied relationships with image quality, but the tastes of Airbnb
consumers do not necessarily align perfectly with image quality. The lack of exogenous variation in the 12
attributes prevents us from drawing causal conclusions. Nevertheless, the DID analysis may offer valuable
exploratory insights. Table 8 presents the results from the estimation of Eq. 5:11
𝐷𝐸𝑀𝐴𝑁𝐷 = 𝐼𝑁𝑇𝐸𝑅𝐶𝐸𝑃𝑇 + 𝛼𝑇𝑅𝐸𝐴𝑇𝐼𝑁𝐷 + 𝜇𝐼𝑀𝐴𝐺𝐸_ 𝐶𝑂𝑈𝑁𝑇
+ 𝜌 𝐵𝐴𝑇𝐻𝑅𝑂𝑂𝑀 𝑃𝐻𝑂𝑇𝑂 𝑅𝐴𝑇𝐼𝑂 + 𝜌 𝐵𝐸𝐷𝑅𝑂𝑂𝑀 𝑃𝐻𝑂𝑇𝑂 𝑅𝐴𝑇𝐼𝑂 (5)
+ 𝜌 𝐾𝐼𝑇𝐶𝐻𝐸𝑁 𝑃𝐻𝑂𝑇𝑂 𝑅𝐴𝑇𝐼𝑂 + 𝜌 𝐿𝐼𝑉𝐼𝑁𝐺 𝑃𝐻𝑂𝑇𝑂 𝑅𝐴𝑇𝐼𝑂
+ 𝜂𝐼𝑀𝐴𝐺𝐸 𝐴𝑇𝑇𝑅𝐼𝐵𝑈𝑇𝐸𝑆 + 𝜆𝐶𝑂𝑁𝑇𝑅𝑂𝐿𝑆 + 𝑆𝐸𝐴𝑆𝑂𝑁𝐴𝐿𝐼𝑇𝑌 + 𝑃𝑅𝑂𝑃𝐸𝑅𝑇𝑌 + 𝜀
Recall that we theorized that the contrast of brightness has a negative effect on property demand, and the
remaining 11 image attributes have positive effects. The coefficients in Table 8 are of the expected sign, that is,
significantly negative for the contrast of brightness and significantly positive for the other attributes. We can
conclude that, controlling for all other variables, the 12 image attributes are significantly correlated with property
demand.

Among the four composition attributes, the largest coefficients belong to the visual balance of color and
visual balance of intensity. Both features describe the symmetry of the image, which is determined by both the
physical arrangement of elements in the photograph and the photographer’s position from which the image is
taken.

Among the five color attributes, the largest magnitudes of coefficients belong to image clarity and the contrast
of brightness. Image clarity is obviously desirable; images that do not convey information clearly cannot make
elements of the property appear attractive to the consumer. One would expect most skilled photographers to
prioritize image clarity, but Table 7 shows a significant difference in image clarity between the unverified and
verified high-quality images; unsurprisingly, the verified photos score almost twice as high as the unverified low-
quality images. Meanwhile, a low value is usually preferred for the contrast of brightness, though several hosts

11For ease of understanding, we use standardized image attributes (i.e., variables are normalized to zero-mean and unit-
variance).
27

Electronic copy available at: https://ssrn.com/abstract=2976021

have complained (on the Airbnb host forums) that the contrast of brightness is so low in the verified photos that
they appear washed out, which might drive guests away. 12 Our results do not find evidence to support this
concern. The lower contrast of brightness in the verified photos (Table 7) may actually make property images
more appealing to viewers.

Among the three FG relationship features, the most prominent one is the size difference, as suggested by its
largest coefficient. When the figure has a much larger size than the background, it is better able to capture the
viewer’s attention. A photographer’s ability to manipulate the size difference may be physically constrained; for
instance, if there is little size difference between the two, it will be hard to separate the figure from the ground.

With the inclusion of the 12 image attributes, the coefficient of the key variable TREATIND falls to 1.721
and is statistically insignificant (Table 7), suggesting that the treatment effect is associated with the ability of
Airbnb professional photographers to achieve more optimal values of the 12 interpretable attributes.

Table 8. DiD Model: Controlling for Interpretable Image Attributes

COMPONENT VARIABLES ESTIMATES Robust S.E.

TREATIND 1.721 6.575
Property (Non-Photo) Characteristics
log REVIEW_COUNT 9.279*** 0.927
NIGHTLY_RATE -0.194*** 0.0325
INSTANT_BOOK 3.245* 1.351
CLEANING_FEE 0.0955*** 0.0185
MAX_GUESTS 0.159 1.093
RESPONSE_RATE 0.0946* 0.0424
RESPONSE_TIME (minutes) -0.000203 0.00158
MINIMUM_STAY 0.159 0.130
SECURITY_DEPOSIT 0.00189 0.00208
SUPER_HOST 3.614* 1.518
BUSINESS_READY 2.182* 0.971
CANCELLATION_STRICT 1.670 1.256
HAS_RATING 14.02 11.76
HAS_RATING × COMMUNICATION 0.0507 1.393
HAS_RATING × ACCURACY 0.156 1.200
HAS_RATING × CLEANLINESS -0.879 1.112
HAS_RATING × CHECKIN -2.198 1.472
HAS_RATING × LOCATION -0.473 1.156
HAS_RATING × VALUE 2.059 1.165
AFTER × POOL 6.789 4.455
AFTER × BEACH -9.047 10.31
AFTER × AC 1.474 2.410
Property Image Characteristics
log IMAGE_COUNT 4.530* 1.824
BATHROOM_PHOTO_RATIO 3.431 8.526

12 See discussion at https://airhostsforum.com/t/airbnb-plus-program-anyone-else-who-thinks-the-new-photos-

suck/30120.
28

Electronic copy available at: https://ssrn.com/abstract=2976021

 BEDROOM_PHOTO_RATIO 0.282 7.642
KITCHEN_PHOTO_RATIO 15.57 11.28
LIVINGROOM_PHOTO_RATIO -10.39 8.484
12 Human-Interpretable Image Attributes
1 DIAGONAL_DOMINANCE 2.516** 0.945
Composition 2 VISUAL_BALANCE_INTENSITY 4.618*** 1.350
3 VISUAL_BALANCE_COLOR 8.869*** 2.143
4 RULE_OF_THIRDS 3.537** 1.106
5 WARM_HUE 4.715* 2.363
6 SATURATION 3.920* 1.927
Color 7 BRIGHTNESS 3.434* 1.679
8 CONTRAST_OF_BRIGHTNESS -4.897* 2.411
9 IMAGE_CLARITY 6.212** 2.175
Figure-Ground 10 SIZE_DIFFERENCE 3.807* 1.541
Relationship 11 COLOR_DIFFERENCE 2.728* 1.372
12 TEXTURE_DIFFERENCE 2.313* 1.090
Fixed Effect Property
Seasonality City-Year-Month
Num. Observations 76,901
R-squared 0.6670
Note: *p < 0.05, **p < 0.01, ***p <0.001.

4. Discussion and Conclusions

This study employs a DiD analysis combined with the PSW method to investigate Airbnb’s photography
program. We find that the properties that acquired verified photos had an 8.98% higher occupancy rate,
translating into an average increase of $3500.30 in annual revenue, than the properties without verified photos.
The treatment coefficient decreased by 41% when we controlled for image quality, suggesting that a sizeable
portion (but not all) of the coefficient of verified photos on property demand is attributable to the high quality
of the professional images.
Furthermore, we shed light on what makes a good image for an Airbnb property. We show that the verified
and unverified photos differ significantly on 12 interpretable image attributes identified by the photography
literature. We theorize the impact of each of the 12 attributes on property demand, and we find supportive
evidence for all 12 hypothesized relationships. We cannot draw causal conclusions from our observational study,
but our findings nevertheless may inform content-engineering strategies among Airbnb photographers, hosts
who take their own photographs, and real estate marketers more broadly.
We use the PSW method to mitigate the effect of self-selection on our DID analysis. There are, however,
possible reasons that may cause risk to the PSW and DiD analyses: in a related study, motivated by our findings
and leveraging the foundations of our model, Zhang et al. (2019) investigate the reason as to why a large
proportion of Airbnb hosts do not adopt high quality images even when Airbnb offers them for free. They find
that high-quality images lead to higher expectations for the property and level of service. If guests hold
inappropriately high expectations, they are more likely to be dissatisfied and less likely to write a review, if such
expectations are not met. In the long run, a property may face lower demand from its relatively low number of
29

Electronic copy available at: https://ssrn.com/abstract=2976021

accumulated reviews. That is, the service quality and property quality, affect hosts’ verified photo adoptions, and
hence beside other factors, create a self-selection issue for our study. As discussed in Section 3.2, our PSW model
helps to mitigate the issue, through controlling for Response Rate and Response Time, and a rich set of property
characteristics. Additionally, note Zhang et al. (2019) finds that the direct effect of image quality on demand is
always positive, in the short- and long- term for all properties (consistent with our long-term robustness test, see
V.7 of the appendix). The negative effect of images’ quality on property demand comes indirectly through a lower
rate of review accumulation for properties using high-quality images. Similar to Zhang et al. (2019), we control
for the number of reviews in the demand model.

The PSW approach accounts for and matches properties on observed characteristics, but unobserved aspects
of property quality may also affect both the property demand and treatment likelihood. We assess the robustness
of our results to unobserved factors with two follow-up analyses: a sensitivity analysis (Rosenbaum Bounds Test,
Section 3.4.2) and a DiD estimation with the pre-treated units (i.e., those that acquired verified photos before
January 2016) as the control group, such that any unobservables that drive self-selection should affect the
treatment and control groups equally (Section 3.4.3; Section V of the appendix). The results increase our
confidence that the PSW method adequately mitigated the endogeneity concern.

There are a few limitations to this research. First, our trained CNN classifier has a relatively high accuracy of
90.4%, but it is not perfectly accurate;13 future studies may further improve the classifier’s performance with
more training data. Second, we ignore an important element of the user search process on Airbnb. Typically, a
potential guest browses several properties that appear on an Airbnb search page based on user-specified criteria
(e.g., location, dates,). A single image appears for each property on the search result page and may influence
which properties the guest chooses to evaluate further (at which point the guest could see all images available for
the property). Without access to consumer search processes, we cannot explicitly incorporate relevant
information into our analysis, but future research may pursue such an analysis as more data (e.g., on the search
process and transaction) become available to researchers. Lastly, as noted previously, our observational data do
not enable us to make causal claims about the relationship between image features and property demand. Future
research can investigate the causal impact of image attributes on property revenue if the data and context allow
(e.g., in a randomized controlled trial).
This paper relates property demand to both high-level and low-level dimensions of image features. Certain
industries could benefit from the documented insights regarding the 12 image attributes. For example, home
rental markets, such as Airbnb and VRBO, could reduce the issue of quality uncertainty by incentivizing their
hosts to display high-quality property images. In the related industry of real estate (Zillow.com, Redfin.com,

13In our analyses, the IMAGE_QUALITY variable in the econometric model is the mean of the quality labels for the
property’s set of images. Though some images likely were misclassified during the machine learning step, some
misclassifications should cancel out in the computed mean quality, leading to a relatively accurate mean value.
30

Electronic copy available at: https://ssrn.com/abstract=2976021

RE/MAX, etc.), a platform such as Zillow.com could use our results to improve listing images. Finally, our study
is among the first to investigate the difference in demand generated by properties with verified and unverified
photos and to examine the correlations between identifiable image attributes and property demand.

References
Angrist, J., and Pischke, J. 2008. Mostly Harmless Econometrics: An Empiricist’s Companion. Princeton University Press.
Arnheim, R. 1974. Art and Visual Perception: A Psychology of the Creative Eye. Berkeley: University of California Press.
Athey, S., and Imbens, G.W. 2006. Identification and Inference in Nonlinear Difference-in-Differences Models. Econometrica
74(2): 431–497.
Austin, P.C., and Stuart, E.A. 2015. Moving towards Best Practice When Using Inverse Probability of Treatment Weighting
(IPTW) Using the Propensity Score to Estimate Causal Treatment Effects in Observational Studies. Statistics in Medicine
34(28): 3661–3679.

Autor, D. 2003. Outsourcing at Will: The Contribution of Unjust Dismissal Doctrine to the Growth of Employment
Outsourcing. Journal of Labor Economics 21(1): 1–42.
Berry, S., Levinsohn, J., and Pakes, A. 1995. Automobile Prices in Market Equilibrium. Econometrica 63(4): 841–890.
Bertrand, M., Karlan, D., Mullainathan, S., Shafir, E., and Zinman, J. 2010. What’s Advertising Content Worth? Evidence
from a Consumer Credit Marketing Field Experiment. The Quarterly Journal of Economics, 125(1): 263–306.

Bornstein, M.H., Ferdinandsen, K., and Charles, G.G. 1981. Perception of Symmetry in Infancy. Developmental Psychology
17(1): 82–86.
Datta, R., Joshi, D., Li, J., and Wang, J.Z. 2006. Studying Aesthetics in Photographic Images Using a Computational
Approach. ECCV 3953: 288–301.
DiPrete, T.A., and Gangl, M. 2004. Assessing Bias in the Estimation of Causal Effects: Rosenbaum Bounds on Matching
Estimators and Instrumental Variables Estimation with Imperfect Instruments. Sociological Methodology, 34: 271–310.
Freeman, M. 2007. The Photographer’s Eye: Composition and Design for Better Digital Photos. (1st ed.). Focal Press.
Gorn, G.J., Chattopadhyay, A., Sengupta, and Tripathi, J.S. 2004. Waiting for the Web: How Screen Color Affects Time
Perception. Journal of Marketing Research 41(2): 215–225.
Gorn, G.J., Chattopadhyay, A., Yi, T., and Dahl, D.W. 1997. Effects of Color as an Executional Cue in Advertising: They’re
in the Shade. Management Science 43(10): 1387–1400.
Grill, T., and Scanlon, M. 1990. Photographic Composition. Watson-Guptill.

Hagtvedt, H., and Patrick, V.M. 2008. Art Infusion: The Influence of Visual Art on the Perception and Evaluation of
Consumer Products. Journal of Marketing Research 45(3): 379–389.
He, K., Sun, J., and Tang, X. 2011. Single Image Haze Removal Using Dark Channel Prior. Pattern Analysis and Machine
Intelligence. IEEE Transactions 33(12): 2341–2353.

Heckman, J., Ichimura, H., and Todd, P. 1997. Matching as an Econometric Evaluation Estimator: Evidence from
Evaluating a Job Training Programme. The Review of Economic Studies 64(4): 605–654.
Krages, B. 2005. Photography: The Art of Composition. Allworth Press, USA.

Kreitler, H., and Kreitler, S. 1972. Psychology of the Arts. Duke University Press.
Krizhevsky, A., Sutskever I., and Hinton, G.E. 2012. Imagenet Classification with Deep Convolutional Neural Networks.
In NIPS’12: Proceedings of the 25th International Conference on Neural Information Processing Systems, Vol. 1, (pp. 1097–1105).

31

Electronic copy available at: https://ssrn.com/abstract=2976021

Larsen, V., Luna, D., and Peracchio, L.A. 2004. Points of View and Pieces of Time: A Taxonomy of Image Attributes. Journal
of Consumer Research 31(1): 102−111.

Levkowitz, H., and Herman, G.T. 1993. GLHS: A Generalized Lightness, Hue and Saturation Color Model. CVGIP:
Graphical Models and Image Processing 55(4): 271–285. doi:10.1006/cgip.1993.1019.

Li, J., Moreno, A., and Zhang, D. 2016. Pros vs Joes: Agent Pricing Behavior in the Sharing Economy. Available at
SSRN: https://ssrn.com/abstract=2708279.

Liu, L., Dzyabura, D., and Mizik, N. 2020. Visual Listening In: Extracting Brand Image Portrayed on Social Media. Marketing
Science 39(4): 669-686. https://doi.org/10.1287/mksc.2020.1226

Machajdik, J., and Hanbury, A. 2010. Affective Image Classification Using Features Inspired by Psychology and Art Theory.
In Proceedings of the International Conference on Multimedia (pp. 83–92) ACM.
Malik, N., Singh, P. V., and Srinivasan, Kannan, 2019. A Dynamic Analysis of Beauty Premium (February 26, 2019).
Available at SSRN: https://ssrn.com/abstract=3208162 or http://dx.doi.org/10.2139/ssrn.3208162

Malik, N., and Singh, P.V. 2019. Deep Learning in Computer Vision: Methods, Interpretation, Causation and Fairness,
Tutorials in Operations Research (pp. 73-100).

Manchanda, P., Packard, G., and Pattabhiramaiah, A. 2015. Social Dollars: The Economic Impact of Customer Participation
in a Firm-Sponsored Online Customer Community. Marketing Science 34(3): 367–387.

Meech, S. 2004. Contemporary Quilts: Design, Surface and Stitch. Batsford: London, UK.
Meyers-Levy, J., and Peracchio, L. A. 1992. Getting an Angie in Advertising: The Effect of Camera Angle on Product
Evaluations. Journal of Marketing Research 29: 454–461.

Miller, E.G., and Kahn, B.E. 2005. Shades of Meaning: The Effect of Color and Flavor Names on Consumer Choice. Journal
of Consumer Research 32(1): 86–92.
Mitchell, A., and Olsen, J.O. 1981. Are Product Attribute Beliefs the Only Mediator of Advertising Effects on Brand
Attitude? Journal of Marketing Research 18: 318–332.

Narang, Unnati & Shankar, Venkatesh. (2019). Mobile App Introduction and Online and Offline Purchases and Product
Returns. Marketing Science. 38. 10.1287/mksc.2019.1169.

Nevo, A. 2000. Mergers with Differentiated Products: The Case of the Ready-to-Eat Cereal Industry. The RAND Journal of
Economics 31(3): 395–421.

Netzer, O., Lemaire, A., & Herzenstein, M. (2019). When Words Sweat: Identifying Signals for Loan Default in the Text of
Loan Applications. Journal of Marketing Research, 56(6), 960–980.

Peracchio, L.A., and Meyers-Levy, J. 1994. How Ambiguous Cropped Objects in Ad Photos Can Affect Product
Evaluations. Journal of Consumer Research 21: 190–204.
PwC Report. 2015. Consumer Intelligence Series: The Sharing Economy.

Rosenbaum, P.R. 1993. Hodges-Lehmann Point Estimates of Treatment Effect in Observational Studies. Journal of the
American Statistical Association 83: 1250–1253.
Rosenbaum, P. R. 2002. Observational Studies. (2nd ed.). New York: Springer.
Rosenbaum, P.R., and Rubin, D.B. 1983. The Central Role of the Propensity Score in Observational Studies for Causal
Effects. Biometrika 70(1): 41–55.
Schloss, K. B., and Palmer, S. E. 2011. Aesthetic Response to Color Combinations: Preference, Harmony, and Similarity.
Attention, Perception, & Psychophysics 73(2) 551–571. Available at http://doi.org/10.3758/s13414-010-0027-0.

32

Electronic copy available at: https://ssrn.com/abstract=2976021

Scott, L.M. 1994. Images in Advertising: The Need for a Theory of Visual Rhetoric. Journal of Consumer Research 21: 252–273.

Simonyan, K., and Zisserman, A. 2015. Very Deep Convolutional Networks for Large-Scale Image Recognition. Presented
at International Conference on Learning Representations 2015 (ICLR 2015).
Sun, M., and Zhu, F. 2013. Ad Revenue and Content Commercialization: Evidence from Blogs. Management Science 59(10):
2314–2331.

Ufford, S. 2015. The Future of the Sharing Economy Depends on Trust. Available at
http://www.forbes.com/sites/theyec/2015/02/10/the-future-of-the-sharing-economy-depends-on-trust/#21a3370b58ff.
Valdez, P., and Mehrabian, A. 1994. Effects of Color on Emotions. Journal of Experimental Psychology: General 123(4): 394–409.

Wang, Q., Li, B., and Singh, P.V. 2018. Copycats versus Original Mobile Apps: A Machine Learning Detection Method and
Empirical Analysis. Information Systems Research, 29(2): 273–291.

Wang, X., Jia, J., Yin, J., and Cai, L. 2013. Interpretable Aesthetic Features for Affective Image Classification. In IEEE
International Conference on Image Processing (pp. 3230–3234). Melbourne, Australia.

Wang, K., and Goldfarb, A. 2017. Can Offline Stores Drive Online Sales?. Journal of Marketing Research 54(5): 706–719.

Winkler, R., and Macmillan, D. 2015. The Secret Math of Airbnb’s $24 Billion Valuation. Available at
https://www.wsj.com/articles/the-secret-math-of-airbnbs-24-billion-valuation-1434568517.

Zhou, B., Lapedriza, A., Xiao, J., Torralba, A., and Oliva, A. 2014. Learning Deep Features for Scene Recognition Using
Places Database. In Proceedings of the 27th International Conference on Neural Information Processing Systems. Vol. 1. December 2014.

Zhang, Shunyuan, Nitin Mehta, Param Vir Singh, and Kannan Srinivasan. Can Lower-quality Images Lead to Greater
Demand on AirBnB?. Technical report, Working Paper, Carnegie Mellon University, 2019.

33

Electronic copy available at: https://ssrn.com/abstract=2976021

An Appendix to
What Makes a Good Image? Airbnb Demand Analytics Leveraging Interpretable Image Features

Table of Contents
An Appendix to .................................................................................................................................................................34

I. Classifying Image Quality Using a Deep Learning-based Classification Model .............................................36

1. Training Set Construction ..................................................................................................................................36

2. Image Quality Classifier Training ......................................................................................................................38

II. Algorithm and Concepts for Image Attribute Computation ...........................................................................42

1. Visual (Image) Saliency .......................................................................................................................................42

Definition ..........................................................................................................................................................................42

Calculation ........................................................................................................................................................................42

2. Salient Region .......................................................................................................................................................42

Definition ..........................................................................................................................................................................42

Detection ...........................................................................................................................................................................42

3. FG ..........................................................................................................................................................................43

Definition .....................................................................................................................................................................43

Detection ...........................................................................................................................................................................43

III. Measurement of Interpretable Image Attributes ..............................................................................................45

a. Composition..........................................................................................................................................................45

b. Color ......................................................................................................................................................................46

c. FG Relationship ...................................................................................................................................................46

IV. Room Type Classification.....................................................................................................................................48

Training Set of Indoor/Outdoor Photos .............................................................................................................48

Training a Room Type Classifier on a Collected Training Set ..........................................................................49

V. Robustness Checks, Empirical Extension, and Exclusion of Alternative Explanations .............................51

1) Validating the Propensity Score Method .........................................................................................................51

2) Sensitivity Analysis of the Propensity Score Method (Rosenbaum Bounds Test) ....................................52

3) Addressing the Possibility of an Inflated Long-Term Effect .......................................................................54

34

Electronic copy available at: https://ssrn.com/abstract=2976021

 4) Adding Interaction Terms with Meaningful Amenities.................................................................................59

5) Testing Changes in the Property or the Host’s Unobserved Quality (via Multidimensional Ratings) ..61

6) Alternative Specifications: Logging All Numeric Variables ..........................................................................63

7) Including Pre-treated Units in the Control Group ........................................................................................64

8) Additional Analyses .............................................................................................................................................67

9) Testing Host Behavior – Changes in the Number of Open Days ..............................................................70

VI. Data Description and Sample Construction ......................................................................................................73

VII. Example of Property Images with 1 SD of Improvement in Key Image Features ...................................75

35

Electronic copy available at: https://ssrn.com/abstract=2976021

I. Classifying Image Quality Using a Deep Learning-based Classification Model

1. Training Set Construction

Image Quality Assessment Survey on Amazon Mechanical Turk

We describe the steps for creating the dataset to train our classifier using Amazon Mechanical Turk (AMT). AMT
is a platform of Amazon Web Services that enables users to outsource small tasks to a large group of workers at
a relatively low cost. It has been widely used for human intelligence tasks, such as data collection and data
cleaning. As a crowdsourcing method, AMT has been found to be quite efficient and accurate (Casalboni 2015;
Laws et al. 2011).
To construct a labeled training set for supervised learning, we selected 3,000 Airbnb property images from
our dataset and used AMT to tag each image based on its quality. In the selection of images for AMT tagging,
full random sampling was not optimal, as we did not have information on the distribution of image quality
beforehand. We used stratified random sampling to ensure that a sufficient number of images were evaluated
and labeled for different categories of image quality. A random sample stratified by a crude metric of quality was
necessary, as it ensures that the sample is balanced and random. We randomly selected 500 images from the pool
of verified images, as these were guaranteed to be taken by professional photographers and are most likely to be
of high quality. Then, from all the unverified images, a human judge chose 8,000 images that looked bad, and 500
images from this group were randomly sampled. From the unverified images, we chose 5,000 images that we
judged to be in between excellent and very bad. Of these, 500 images were randomly sampled. Lastly, we randomly
sampled 1,500 images from the entire sample. Constructing the AMT data in this way ensured that we would
have a sufficient random sample of images from each stratum (i.e., subgroup of images with a certain quality).
We also manually reviewed the selected images to ensure that no image was repeatedly sampled.
For the AMT tagging task, we created a survey instrument that asked the Turker to assign a score to a displayed
image based on its aesthetic quality. To provide accurate guidelines for image evaluation, we borrowed
instructions from professional photography forums, as well as Airbnb’s guidelines for shooting good property
photos. We also provided example photos. The quality measurement was based on a Likert scale from 1 to 7, in
which 1 is very bad and 7 is excellent. Figure A1 shows an example question on the survey given to the Turkers. To
ensure high-quality and consistent responses from the Turkers, we required them to have an approval rate of
higher than 95% and to have completed at least 50 approved tasks.

36

Electronic copy available at: https://ssrn.com/abstract=2976021

 Figure A1. Example of the AMT Aesthetic Quality Assessment Task

Constructing the Training Set from Tagged Images

After the AMT survey, we obtained 3,000 tagged images, each of which was evaluated by five Turkers, to be pre-
processed for the construction of our training set. Following previous studies on aesthetic quality (Datta et al.
2006; Datta et al. 2008; Marchesotti et al. 2011), we computed for each image, i, the mean aesthetic score, score ,
given by five Turkers. We then set two thresholds: 𝜃 = 𝑠𝑐𝑜𝑟𝑒 + 𝑔𝑎𝑝/2 and 𝜃 = 𝑠𝑐𝑜𝑟𝑒 − 𝑔𝑎𝑝/2. Next, we
annotated image i with the term high quality if the average score ≥ θ and with the term low quality if the average
score ≤ θ . The mean score for all 3,000 images was 𝑠𝑐𝑜𝑟𝑒 = 4.5. Images with an average score of θ and θ
were excluded from the training set, leaving an artificial gap between high- and low-quality images. As argued by
Datta et al. (2006), the reasons for creating a gap between high- and low-quality images is that close aesthetic
scores (e.g., 4.4 and 4.5) are unlikely to reflect that the images differ in aesthetic quality; rather, they indicate noise
in the quality measurement process. Increasing the value of the gap makes it easier to distinguish between high-
and low-quality images, but it leads to a smaller training set, as images lying within the gap are dropped. To
choose an optimal value, we varied the value of the gap and selected the one that resulted in the best performance
of the trained classifier on a hold-out set (Datta et al. 2006; Marchesotti et al. 2011). Using this method, we chose
a gap of 0.8, which left us with a training set of 2,259 images. As 50% of the training set was selected at random,
we repeated the analysis with the selected 50% and found similar results for the image labeling tasks.

37

Electronic copy available at: https://ssrn.com/abstract=2976021

2. Image Quality Classifier Training

Architecture of the Convolutional Neural Network (CNN) Framework

After a training set was constructed, the next task was to build an image quality classifier using labeled data. We
applied CNN, an emerging deep learning framework that is widely applied in the field of computer vision and
has been shown to perform very well for tasks, such as object recognition and image classification (Krizhevsky
et al. 2012; Simonyan and Zisserman 2015).
As shown in Figure A2, a CNN model consists of a sequence of layers, each having multiple neurons. The
number of neurons can vary from one to thousands. These neuron layers perform matrix multiplication based
on an input, generating an output to serve as the input for the next layer. Both the input and output take the
form of multi-dimensional matrices. The sequence of layers makes the neural network deep.
Images serve as the first input for the deep learning framework. In our training task, we resized all the images
to 224 × 224 pixels, determined the pixel intensity of each image, and represented the image with a 3D array
(matrix) that contains pixel information for the three channels (RGB). This was done to alleviate the
computational burden and ensure that the image size aligned with the pre-trained VGG16 model (described
below).
The last output layer predicts the binary label for its input, which, after passing through the whole network,
is an N-dimensional vector extracted from the image. For an image in our training set (represented by 𝐼𝑀𝐺 ),
the output layer applies a sigmoid function and predicts the label regarding the image’s quality:

1 (ℎ𝑖𝑔ℎ 𝑞𝑢𝑎𝑙𝑖𝑡𝑦) 𝑖𝑓 ≥ 0.5

( ( ))
Label Image Quality 𝐼𝑀𝐺 = ,
0 (𝑙𝑜𝑤 𝑞𝑢𝑎𝑙𝑖𝑡𝑦) 𝑖𝑓 < 0.5
( ( ))

where 𝑋 represents the output from the layer preceding the output layer (in our model, this was the FC2 layer,
which produces a 4096*1 vector), 𝑊 represents the weight parameters, 𝑊 represents the bias (a constant)

connecting the preceding layer to the output layer, and is the probability that the image is of
( ( ))

high quality, given the 𝑋 and 𝑊 and 𝑊 values for the output layer.
Throughout the CNN model, a sequence of weights on each layer define the intermediate extracted vectors
from each layer, including 𝑋 . These weights are adjusted during the training process to optimize the model’s
predictive performance.

Operation of Key Layers in the CNN

A few key layers determine the performance of a CNN: the convolution layer, the zero-padding layer, and the
max-pooling layer. We describe these below.

38

Electronic copy available at: https://ssrn.com/abstract=2976021

Convolution Layer

The convolution layer is the most important and unique layer in the CNN. It consists of a stack of convolution
filters or convolution kernels. For example, the two convolution layers in Layer Block A (shown in Figure A2)
consist of 64 and 128 convolution filters, respectively. A convolution filter is a matrix in which each element
represents a numeric value. For example, in Layer Block A, the convolution layers have a size of 3 × 3 and hence
consist of nine numeric values.14 This matrix, treating an image or an intermediate input as a matrix, operates dot
production by sliding through the input. For an input of a relatively large size (e.g., 224 × 224), a 3 × 3 convolution
filter operates dot production for every 3 × 3 section. The convolution operation is beneficial because it reduces
the dimensionality of parameters and well explores and reserves the (local) spatial relationships of the input.
Regarding the second benefit, if a convolution kernel extracts an oriented edge of an object, operating this kernel
on every small square (e.g., 3 × 3) of an image would enable the extraction of all edges with that direction from
the image. Many kernels that extract edges do so in all directions, potentially constructing the contour of an
object. As can be seen in Figure A2, each block consists of a varying number of convolutional filters (e.g., 64,
128, 256, and 512 filters). These kernels extract features from the input data, which represent the extracted
features from the preceding layers. Toward the output layer (i.e., layers closer to the output layer) in the CNN,
the filters combined together extract higher- and lower-level features. That is, the CNN can extract a hierarchical
structure of related features to predict the output labels.

Zero-Padding Layer

The zero-padding layer adds numerical arrays consisting of all 0 values to the edge of an intermediate output
from a layer. The size of the zero-padding layer is a hyperparameter in the CNN model. Typically, as was done
in our model, the intermediate output is padded with 1*M 0 vectors on each side, causing the width and height
to both increase by 1 after the zero-padding value. As zeros do not contribute to the matrix multiplication
procedure, the zero-padding layer does not affect the features extracted by the layers. In addition, the zero-
padding layer allows us to control the spatial size of the intermediate outputs, and it can prevent the outputs
from decreasing too quickly after layers of convolution operations.

Max-Pooling Layer

Inserting a max-pooling layer between successive convolution layers is common in CNNs. A max-pooling layer
is a small square filter (in our model, a 2 × 2 matrix). Similar to the convolution filter, a max-pooling layer applies
to every 2 × 2 square patch in input data. Its function is to select and preserve only the maximum value in that 2

14 The size of a convolution layer is a choice of the model architecture. A 3 × 3 or 5 × 5 configuration is common.
39

Electronic copy available at: https://ssrn.com/abstract=2976021

× 2 square. Adding max-pooling layers can reduce the spatial size of the intermediate features and the dimension
of the trained parameters in the model, and it can help to efficiently prevent overfitting.

Figure A2. Description of the Architecture and Description of the Layers of the CNN Classifier

Filters: Indicate the number of convolution windows (i.e., number of feature maps) on each convolution
layer.
Zero-padding layer: Pads the input with zeros on the edges to control the spatial size of the output. It has
no impact on the predicted output.
Max-pooling: Subsampling method. A 2 × 2 window slides through each feature map (without overlap) at
that layer, and then the maximum value in the window is selected as a representation of the window. This
reduces computation and ensures translation invariance.

Training the CNN

We randomly split the dataset into training and validation sets, with 80% of the examples forming the training
set and the remainder being used as the validation set. To reduce the overfitting problem in the training step, we
use data augmentation and implement a real-time (i.e., during training) image transformation for each image in
the training sample by randomly (1) flipping the input image horizontally, (2) rescaling the input image within a
scale of 1.2, and (3) rotating the image within 20˚ (images being rotated by a random value between 0˚ and 20˚).
This method introduces random variation in the training sample, increases the training set size, and reduces
overfitting (Krizhevsky et al. 2012).
To effectively learn features from a relatively small sample size, we apply the idea of transfer learning, which
involved building our model on top of an existing well-trained CNN model and then fine-tuning it. As the
features extracted from images are generic to some extent (e.g., almost all CNNs extract edge information at the
first layer), transfer learning is quite common in deep learning and is suggested as an effective approach for
dealing with problems caused by limited data (Girshick et al. 2014; Lin et al. 2015; Zhang et al. 2015). In this
study, we used VGG16 (Simonyan and Zisserman 2015), as it is a conceptually simple and popular pre-trained
40

Electronic copy available at: https://ssrn.com/abstract=2976021

model. We removed the last fully connected layers from the original VGG16 model, as they contain more data-
specific features, and then we added the output layer as the last layer. Figure 2A presents the architecture of our
image classification model. The parameters were initialized using the pre-trained weights, except for the output
layers, for which the parameters were initialized following LeCun’s uniform scaled initiation method (LeCun et
al. 1998). Then, we fixed the parameters on the first 25 layers and fine-tuned the model. The model was trained
on the training set using a NVIDIA K80 GPU, and then performance was tested on the hold-out set at the end
of the training. Optimization was performed with an adaptive method of gradient descent (Adadelta optimization;
Zeiler 2012) for each mini-batch of 16 examples.

41

Electronic copy available at: https://ssrn.com/abstract=2976021

II. Algorithm and Concepts for Image Attribute Computation

In this section, we define the key concepts used in the process of image attribute computation. These key
concepts include image saliency, the key/salient region, and the figure-ground (FG). We first discuss each
concept’s definition and then present the image algorithm used to detect, extract, or compute the concept.

1. Visual (Image) Saliency

The basic unit determining image saliency is visual saliency at the pixel level. The overall saliency score for a local
patch of an image can be computed based on the pixel saliency within the local region.

Definition

Saliency describes a concept that originates from visual unpredictability. In images, it is often captured by
variations in, for example, boundaries and colors. Studies on cognitive psychology and computer vision have
investigated how humans process and pay attention to visual information and have found that we allocate our
attention to parts of given information (e.g., the regions of an image) while cognitively ignoring other parts. Visual
uniqueness is salient in the sense that it easily attracts the attention of viewers.

Calculation

In general, the models proposed for calculating visual saliency are based on local contrast to surroundings. The
contrasts are determined using features, such as color, intensity, edge density, and orientation. A simple example
is the gradient of pixel intensity. A pixel with great contrast is assigned a high saliency value.

2. Salient Region

Definition

Following the definition of visual saliency, salient regions are defined as the regions of an image that have a high
overall saliency score.

Detection

The detection of a salient region involves four steps: (1) segmenting an image into local patches, (2) assigning a
saliency score to each path, (3) merging similar patches into a region, and (4) finding the most salient region. These
steps are discussed in greater depth below.

1) Segmenting an image: This process generally involves grouping the pixels of an image into multiple parts
containing pixels that are similar to one another. Segmentation can be based on edges (detected edges
are assumed to define the boundaries of objects), colors, or other factors. A segmentation algorithm is

42

Electronic copy available at: https://ssrn.com/abstract=2976021

 intended to compare two intensity differences: the difference across the boundary of patches and the
difference between two neighboring pixels within the same patch.
2) Assigning a saliency score to a patch: Each pixel is assigned a saliency value. The saliency score of a patch is
calculated by averaging the scores of all pixels within the patch.
3) Merging patches into a region: If neighboring patches have similar colors, then they are merged into a larger
region.
4) Finding the most salient region: The salient region is found by selecting the region with the highest average
salient score.

3. FG

Definition

The figure is the foreground of an image, and the ground is the background. Only one figure and one ground can
be detected for each image. This is different from the detection of salient regions, for which multiple regions can
be detected.

Detection

The figure is detected and extracted from an image, and then the ground is defined as the rest of the image.
Detection of the foreground is an extension of image segmentation, as a pixel is assigned a value of either 1
(foreground) or 0 (background).

Detailed Algorithm

We used GrabCut, a state-of-the-art model for foreground extraction (Rother et al. 2004). In this model, an image
is treated as a graph, with each pixel serving as a node and pixel similarity being defined as an edge. GrabCut
implements the expectation–maximization (EM) algorithm and min-cut algorithm to iteratively assign a
foreground/background label to pixels and to cut the graph into two subgraphs: one representing the foreground
and the other representing the background.

1) Initially, an arbitrary rectangle separates the image into two parts. The pixels in the rectangle are labeled
“1” (foreground), and those outside it are labeled “0” (background). The initial position of the rectangle
can be arbitrary. Alternatively, one can specify the rectangle’s location or hard label some pixels with
good prior knowledge of where the foreground might be.
2) A Gaussian mixture model (GMM) is trained with the EM algorithm based on the distribution of pixel
color statistics. From the GMM, we can determine the probability that each pixel belongs to a particular
mode (or cluster). That is, the GMM labels a pixel as a probable foreground or a probable background.

43

Electronic copy available at: https://ssrn.com/abstract=2976021

3) A graph is built in which each node represents a pixel, and the edge weight between two pixels represents
pixel similarity. Similar pixels will be assigned a low edge weight and vice versa. Pixel similarity can be
computed based on intensity, color, or texture.
4) Two additional nodes are created on the graph: the source node and the sink node. All pixel nodes
labeled as a probable foreground (probable background) are connected to the source node (sink node),
with the edge weight between each pixel node and the source node (sink node) representing the
probability that the pixel belongs to the foreground (background).
5) Next, the graph is cut into two parts, and the min-cut algorithm is implemented by minimizing the cost
function, defined as the sum of the edge weights across all edges that are cut. The min-cut algorithm
penalizes a cut if it will cause two similar pixels to be separated into two subgraphs. Intuitively, if two
pixels both have a high probability of being in the foreground (background), then it is desirable for them
to be labeled “1” (“0”) at the end of this iteration.
6) Steps 2–5 are repeated until convergence in the pixel labeling is achieved.

44

Electronic copy available at: https://ssrn.com/abstract=2976021

III. Measurement of Interpretable Image Attributes

Using the computer vision algorithm described above, we perform image processing tasks to segment images
into patches and detect key/salient regions. After salient regions are detected, subsequent computation is
performed to measure image attributes. This section discusses the steps for computing image attribute
measurements after image processing.

a. Composition

Four image attributes are categorized during the composition step. How well an image performs concerning a
particular attribute is evaluated by distance, such as the distance between two pixels. A smaller distance indicates
better performance. For all four composition attributes, we compute the distance metrics and then subtract the
metrics from zero.

Diagonal Dominance (Attribute 1): Diagonal dominance captures how close an image’s key region is positioned
to the two diagonals of the image. For an image, we first identify the key region and then measure the weighted
Manhattan distance from the key region to each diagonal (Liu et al. 2010; Wang et al. 2013).15 The measurement
of diagonal dominance is computed by subtracting the minimum weighted distance from zero.
A greater diagonal dominance value suggests that the image is more diagonally dominant.

Rule of Thirds (ROT) (Attribute 2): We first divide an image into nine equal parts with two (imaginary) equally
spaced vertical lines and two (imaginary) equally spaced horizontal lines. Then, we calculate the Euclidean
distance from the centroid of the key region to each of the intersections (Wang et al. 2013). If the minimum
distance is small, then the image follows the ROT, with its key region close to at least one intersection. The ROT
is measured by subtracting the minimum distance from zero. If an image follows the ROT more closely, the ROT
value is greater.

Visual Balance Intensity (Attribute 3): In this measurement, we split the image along its vertical central line.
On each half of the image, we identify a key region and compute the distance from its centroid to the vertical
central line (Liu et al. 2010). A relative distance measure is calculated by subtracting the shorter distance from the
longer one and dividing the difference by the longer distance. Then, visual balance intensity is computed by
subtracting the relative distance from zero. A greater value for this measure suggests that the image is more
(vertically) visually balanced in terms of pixel intensity.

Visual Balance Color (Attribute 4): The color measurement of visual balance compares the left half of an image
with the right half based on color. We first calculate the Euclidean distance in terms of color intensity (i.e., RGB

15The Manhattan distance between two points on an image is measured as the number of pixels between them, with only
horizontal and vertical paths allowed.
45

Electronic copy available at: https://ssrn.com/abstract=2976021

channel) between each pixel and its symmetrical pixel (across the vertical line). Then, visual balance color is
computed by subtracting the average difference from zero. A greater value suggests that the image is more visually
balanced in terms of color around its vertical central line.

b. Color

Five image attributes related to color are computed. The measurements are based on pixel intensity or related
values (e.g., hue and saturation).

Warm Hue (Attribute 5): The warm hue measurement captures the warmth of an image, which is defined by the
relative proportion of warm hues (e.g., yellow) to cool hues (e.g., green). The measurement is computed in the
HSV (hue, saturation, and volume) space. Specifically, we calculate the pixel hues that fall outside the cool range
(30–110) on the hue spectrum (Wang et al. 2013). If an image contains more warm hues, such as yellow and
orange, it will have a greater warm hue value.

Saturation (Attribute 6): We compute pixel saturation in the HSV space by averaging the saturation values of all
pixels in the image. A greater value indicates higher saturation (for example, the image contains more saturated
colors).

Brightness (Attribute 7): The brightness of an image is defined as the overall level of illumination. We calculate
the intensity of each pixel and then average the intensity values across all pixels in the image. A brighter image
has a greater brightness value.

Contrast of Brightness (Attribute 8): Contrast is calculated as the SD of pixel intensity in the whole image. A
lower contrast of brightness value suggests that the brightness is more evenly distributed across the image.

Image Clarity (Attribute 9): Image clarity captures the portion of hues with sufficient intensity. We measure
pixel brightness on a 0–1 scale and then compute the portion of pixels whose brightness is 0.7–1 (Wang et al.
2013). A clear image has a higher image clarity value.

c. FG Relationship

The FG relationship is described as the difference between the figure and its ground in terms of three metrics:
size, color, and texture. An image with a good FG relationship has a clearly separable figure and ground (i.e.,
greater differences).

Size Difference (Attribute 10): The size difference attribute compares the size of the figure with that of the
ground. We detect the figure and ground in an image and calculate the size of each in relation to the whole image
(Cheng et al. 2011). Then, the size difference is computed by subtracting the ground’s size ratio from the figure’s

46

Electronic copy available at: https://ssrn.com/abstract=2976021

size ratio. A greater size difference value indicates that the image has a figure occupying a relatively larger area,
standing out from the ground.

Color Difference (Attribute 11): The color difference attribute captures the difference in colors between the
figure and the ground. We compute the Euclidean distance between the mean color of the figure and that of the
ground. A high color difference value suggests that the figure and ground contain distinct colors. In such cases,
the figure can be easily distinguished from the ground.

Texture Difference (Attribute 12): Texture difference measures the difference between the figure and the
ground in terms of texture, which is captured by the edge density within a local region. For the figure and ground,
we use the Canny edge detector to detect the edge and then compute edge density. Then, we measure the absolute
difference between the two densities. A high texture difference value suggests that the figure and ground have a
clear separation based on texture.

47

Electronic copy available at: https://ssrn.com/abstract=2976021

IV. Room Type Classification

We build a deep learning model to automatically categorize the type of scene photographed. The goal is to
compute the distribution of different types of rooms depicted in property images. Controlling the distribution in
our demand model helps us address the concern that professional photographers know which types of places
appeal more to consumers and thus present these aspects of properties.
We build a deep learning model to automatically classify the room type (bathroom, bedroom, kitchen, living
room, or outdoor area) depicted in a given property image. Using transfer learning with a deep learning model
that was pre-trained on a large scene classification dataset, Places205 (Zhou et al. 2014), we optimize the classifier
for a dataset we collected, which consists of 54,557 images of bathrooms, 59,082 images of bedrooms, 88,030
images of kitchens, 81,819 images of living rooms, and 5,734 images of outdoor areas. The average classification
accuracy is 95.05% on a hold-out set across the five categories.

Training Set of Indoor/Outdoor Photos

To train a room type classifier, we need a large number of room images labeled with a room type. We would
have preferred to use original images from Airbnb.com, but the images on property web pages are not labeled,
and it would incur a high labor cost to have someone (e.g., AMT) manually label images for us. Therefore, we
crawled data from real estate–related websites on which a vast number of indoor/outdoor images are classified
into categories. Figure A3 shows a portion of a web page displaying 23 images of kitchens at multiple properties.

48

Electronic copy available at: https://ssrn.com/abstract=2976021

 Figure A3 A Real Estate Web Page Showing Kitchens

From the website, we collected images aligning with the five room types: bathroom, bedroom, living room,
kitchen, and outdoor area. We then split the dataset, 80% of which was used as a training set and the remaining
20% used as a hold-out test set.

Training a Room Type Classifier on a Collected Training Set

We used the VGG16 ConvNet model, which was pre-trained on the Places205 dataset (Zhou et al. 2014) and
then fine-tuned on our training set.16 The model was used for a task to classify 205 categories of places. To
transfer the pre-trained model to our study, we removed the output layer in the pre-trained model and then added
an output layer designed for our specific task. The added output layer was a 5*1 vector in which each element
indicated the predicted probability of assigning the corresponding label. To ensure that all the probabilities add
up to 1 (as we assume that a room belongs to only one category), the Softmax function was used to calculate the
predicted probability (the function ensures that all the predicted probabilities add up to 1). For example, the
probability that an image is assigned room type k is computed as follows:

Prob Room Type = k X = ,

∑ (1)

16 The pre-trained VGG16 model (including the architecture and parameters) can be accessed at
http://places.csail.mit.edu/downloadCNN.html.
49

Electronic copy available at: https://ssrn.com/abstract=2976021

where j = 0,…,4 represents the room type (bathroom, bedroom, kitchen, living room, or outdoor area,
respectively), 𝑋 represents the output from FC2 (the second fully connected layer), 𝑊 represents the weight

connecting FC2 to the 𝑗 node on the output layer, and 𝑊 represents the bias connecting FC2 to the 𝑗 node
on the output layer.

50

Electronic copy available at: https://ssrn.com/abstract=2976021

V. Robustness Checks, Empirical Extension, and Exclusion of Alternative Explanations

This section reports a series of analyses performed to test the robustness of our main results and/or exclude
alternative explanations. We begin by presenting the validation of the propensity score method used in our
empirical model, which included a propensity score weighting (PSW) strategy and a sensitivity assessment of
unobservables.

1) Validating the Propensity Score Method

To ensure that the propensity score (PS) approach effectively eliminates potential systematic imbalances between
the treatment and control groups, one needs to show that the PSs have balanced the covariates for matched or
weighted samples.
We implemented a balance check, which compares, in terms of the covariates, the weighted means of the
∑ 𝑿𝒊 ∑ 𝑿𝒊
treatment group, 𝑿𝒕𝒓𝒆𝒂𝒕𝒎𝒆𝒏𝒕 = ∑
, and the control group, 𝑿𝒄𝒐𝒏𝒕𝒓𝒐𝒍 = ∑
. Here, 𝑋 is a

1*M dimensional vector of the pre-treatment covariates (i.e., the covariates observed before the treatment) of
unit 𝑖, and 𝜔 is the sample weight for unit 𝑖, computed based on the estimated PSs.
Table 2 presents the weighted group means and tests for the differences in the means for each variable 𝑋
(m = 1, 2, …, M). As shown by the t-statistics in the table, the weighted samples are not statistically different at
the 95% significance level. That is, the systematic differences in the weighed samples are negligible after
performing the PSW method. Therefore, we validated that our PSW method effectively eliminated imbalances
in the sample and that our weighted treatment and control groups are comparable in terms of the observed
covariates that may affect the treatment selection process.

Table A9. PSW Validation: Covariate Balance Check

Weighted Means in
Groups T-test
Variables Treated Untreated t p-value
REVIEW_COUNT 20.56 19.88 0.27 0.790
IMAGE_QUALITY 0.27 0.25 1.00 0.316
IMAGE_COUNT 14.48 15.1 −0.67 0.506
NIGHTLY_RATE 170.15 191.36 −1.14 0.257
MINIMUM_STAY 2.57 2.57 −0.00 1.000
MAX_GUESTS 3.5 3.67 −0.82 0.410
RESPONSE_RATE 92.25 91.19 0.79 0.431
RESPONSE_TIME (minutes) 225.12 260.98 −1.18 0.238

51

Electronic copy available at: https://ssrn.com/abstract=2976021

 SUPER_HOST 0.15 0.11 1.05 0.292
INSTANT_BOOK 0.11 0.11 −0.14 0.888
# BLOCKED DAYS 9.51 8.32 1.10 0.271
# RESERVATION DAYS 6.62 6.74 −0.16 0.877
PARKING 0.5 0.49 0.09 0.929
POOL 0.1 0.08 0.76 0.445
BEACH 0.02 0.02 0.34 0.737
INTERNET 0.99 1 −0.58 0.563
TV 0.79 0.81 −0.55 0.579
WASHER 0.6 0.57 0.81 0.419
MICROWAVE 0.15 0.13 0.64 0.523
ELEVATOR 0.2 0.21 −0.22 0.826
GYM 0.11 0.13 −0.69 0.490
FAMILY_FRIENDLY 0.19 0.2 −0.56 0.576
SMOKE_DETECTOR 0.55 0.52 0.62 0.534
SHAMPOO 0.45 0.44 0.09 0.929
BATHROOM_PHOTO_RATIO 0.22 0.21 1.05 0.295
BEDROOM_PHOTO_RATIO 0.29 0.29 −0.26 0.795
KITCHEN_PHOTO_RATIO 0.1 0.1 −0.15 0.880
LIVINGROOM_PHOTO_RATIO 0.18 0.19 −0.52 0.602
OUTDOOR_PHOTO_RATIO 0.2 0.21 −0.13 0.898
SECURITY_DEPOSIT 202.77 225.19 −0.65 0.517
# OPEN DAYS 21.49 22.68 −1.10 0.271
CANCELLATION_STRICT 0.26 0.29 −0.89 0.371
# BEDS 1.81 1.96 −1.17 0.242
APARTMENT 0.6 0.61 −0.27 0.786
ENTIRE_HOME 0.61 0.64 −0.64 0.522

2) Sensitivity Analysis of the Propensity Score Method (Rosenbaum Bounds Test)

In the model for estimating the propensity scores, we included a rich set of variables and their interactions. The
inclusion of a complete set of covariates reduced the odds that our main results would be affected by variables
that were not accounted for when computing the propensity scores.
Despite the long list of included covariates for propensity scores estimation, some variables that may affect
one’s decision regarding participation in a professional program were omitted. As is commonly acknowledged,

52

Electronic copy available at: https://ssrn.com/abstract=2976021

PSs are computed based on observed variables; therefore, there may be a hidden bias if unobserved variables
simultaneously affect the selection process (i.e., the treatment assignment) and the outcome variables. To assess
the sensitivity of our estimation to potential hidden bias, we implement a widely adopted approach of sensitive
analysis—the Rosenbaum bounds test (Rosenbaum 2002).
The logic of the Rosenbaum bounds analysis is as follows. Suppose the participation probability of unit 𝑖 is
𝑃 (𝑡𝑟𝑒𝑎𝑡𝑒𝑑 = 1|𝑥 , 𝑢 ) = 𝑓(𝛽𝑥 + 𝛾𝑢 ) , where 𝑥 and 𝑢 are the vectors of observed and unobserved
variables, respectively, and 𝛾 is 0 if there are no unobserved variables affecting the treatment selection process.
Units 𝑖 and 𝑗 with 𝑥 = 𝑥 have the same probability of receiving the treatment if and only if 𝛾(𝑢 − 𝑢 ) = 0.
Rosenbaum bounds evaluate the degree of minimum change in the odds ratio (OR) of participation due to
unobservables would be required to nullify the treatment effect identified with the PS method. The estimation
results inspire more confidence if a greater change in the OR caused by unobservables is needed to overturn the
estimated treatment effect.
The results of the Rosenbaum bounds test are provided in Table A10. As our main DiD (Difference in
Differences) analysis identified that verified photos had a positive effect on property demand, we were more
concerned with potential upward (positive) than downward (negative) bias in the DiD estimator. Therefore, in
Table A11, we are mostly interested in the sig+ column. The gamma results suggest that even when a hypothetical
unobservable increases the OR by 1.4 times, our causal inference of the positive treatment coefficient of
professional images on property demand, identified with the PS method, will be robust at the 95% confidence
level (and will remain positively significant at the 90% confidence level until the gamma is increased to 1.55). The
results suggest that for a positive estimated treatment effect on property demand to be overturned, the potential
unobserved factors affecting the treatment assignment process would have to be large enough to increase the
OR of participation by at least 60%. Moreover, if we look at Hodges–Lehmann’s estimates (Rosenbaum 1993),
they suggest a more robust result, as the upper (t-hat+) and lower (t-hat) bounds do not contain 0, at least until
1.6.
The results of our sensitivity analysis reveal a similar increase in gamma to that reported in the extant literature
(1.2- to 1.6-fold; DiPrete et al. 2004; Li et al. 2016; Manchanda et al. 2015; Sun and Zhu 2014). Although the
treatment effect will be insignificant if the unobservable is large enough to change propensities, this does not
mean that the PS method is invalid because Rosenbaum bounds analysis gives us a lower bound of confidence
when making a causal inference in the worst-case scenario with hypothetical hidden selection bias (note that
hidden bias caused by unobservables does not necessarily exist). For this reason, we are confident that our study
is, to some extent, robust to hidden bias caused by hypothetical unobserved factors affecting the selection
process.

53

Electronic copy available at: https://ssrn.com/abstract=2976021

 Table A10. Sensitivity Analysis: Rosenbaum Bounds Test

Gamma Significance Level Hodges–Lehmann Point 95% Confidence

Estimate Interval

sig+ sig- t-hat+ t-hat CI+ CI-

1 0.000205 0.000205 13.9456 13.9456 5.50612 24.1935

1.05 0.00052 0.000075 12.9032 15.8307 4.83871 25

1.1 0.00119 0.000027 12.5 17.5824 3.22581 26.4631

1.15 0.002482 9.50E-06 11.2903 18.3333 2.2043 27.7778

1.2 0.004777 3.30E-06 10.2716 19.3548 1.25353 29.6461

1.25 0.008569 1.10E-06 9.05707 20.2419 −4.40E-07 30

1.3 0.014442 3.80E-07 8.06452 21.4286 −4.40E-07 30.8405

1.35 0.02304 1.30E-07 6.45162 22.4194 −4.40E-07 32.0079

1.4 0.035005 4.20E-08 6.45161 23.5526 −4.40E-07 33.1349

1.45 0.050917 1.40E-08 5.50612 24.1935 −4.00E-06 33.8095

1.5 0.071236 4.50E-09 5 24.7878 −1.0326 35.0824

1.55 0.096249 1.50E-09 4.07337 25.8064 −2.01613 35.7692

1.6 0.126037 4.70E-10 3.22581 26.7374 −3.10676 36.7374

Note: Gamma log odds of differential assignment because of unobserved factors; sig + (−): upper (lower)
bound of the significance level; t-hat + (−): upper (lower) bound of the Hodges–Lehmann point estimate
CI + (−): upper (lower) bound of the confidence interval (a = 0.95).

3) Addressing the Possibility of an Inflated Long-Term Effect

i. Analyses of a Shorter-Period Sample

This is an alternative approach to address the concern about the possible long-term inflation of the treatment
effect if Airbnb’s search algorithm favors properties with professional photos. To address this concern, we

54

Electronic copy available at: https://ssrn.com/abstract=2976021

estimate our main DiD specification on a set of subsamples in which we included a shorter period (8 months)
for each treated unit.
For each treated unit, the observation spans exactly four periods per property before and after treatment
adoption. For the untreated (control) units, we used the full periods (16 months) for estimation. This is because
for untreated units, there is no reference period for which we can define the pre- and post-spans.
Table A11 presents the estimation results for the robustness analysis of the selected subsets. As shown, the
estimated coefficients of the key variable, TREATIND, remain consistently positive and significant. The
consistent estimated treatment effect when shorter periods are applied adds confidence that our main finding
was not driven by long-term inflation caused by Airbnb’s search ranking algorithm.

55

Electronic copy available at: https://ssrn.com/abstract=2976021

 Table A11. DiD Robustness Tests: Selecting Shorter Periods of Samples (Limited Eight
Periods/Month Spans)

Subset of four pre- and four post-treatment periods

Variables for treated units and a whole period for untreated
units

Estimates Robust S.E.

TREATIND 10.33*** 1.624

log REVIEW_COUNTt-1 10.03*** 0.768
NIGHTLY_RATE −0.175*** 0.0262
INSTANT_BOOK 4.436*** 1.169
CLEANING_FEE 0.0898*** 0.0125
MAX_GUESTS −2.142* 1.034
RESPONSE_RATE 0.0767* 0.0348
RESPONSE_TIME (minutes) −0.00152 0.00136
MINIMUM_STAY 0.0279 0.0873
SECURITY_DEPOSIT 0.00298* 0.00143
SUPER_HOST 1.989* 1.003
BUSINESS_READY 0.902 0.919
CANCELLATION_STRICT 0.521 1.045
HAS_RATING 6.380 8.310
HAS_RATING × COMMUNICATION 0.204 1.336
HAS_RATING × ACCURACY 0.00741 1.056
HAS_RATING × CLEANLINESS −2.015 1.035
HAS_RATING × CHECKIN −1.142 1.398
HAS_RATING × LOCATION 0.805 1.074
HAS_RATING × VALUE 1.552 0.998

Fixed Effect Property

Seasonality City-Year-Month

Observations 75,406
R-squared 0.6943

Note: The number of observations is smaller than that in the main analyses (Tables 2, 3, 4, and 7 in the
main paper) because a shorter period was used for the treated units. Specifically, instead of using the whole
panel of 16 periods of data, we used up to eight periods for the estimation in this table.
*p < 0.05, **p < 0.01, ***p < 0.001.

56

Electronic copy available at: https://ssrn.com/abstract=2976021

 ii. Potential Dynamics in the Treatment Effect (Robustness Test of Possible Long-term Inflation)
To address the concern that the estimated effect could be inflated in the long term if Airbnb’s ranking algorithm
favors properties with professional images, we implemented a relative-time model to break the after period into
a series of short periods, and we estimated the treatment effect for each month following treatment adoption.
We again relied on a relative model to estimate the following:

𝐷𝐸𝑀𝐴𝑁𝐷 = 𝐼𝑁𝑇𝐸𝑅𝐶𝐸𝑃𝑇 + ∑ 𝛽 (𝑃𝑅𝐸 (𝑗) ∙ 𝑇𝑅𝐸𝐴𝑇 ) + ∑ 𝛽 (𝑃𝑂𝑆𝑇 (𝑘) ∙

𝑇𝑅𝐸𝐴𝑇 ) + 𝜆𝐶𝑂𝑁𝑇𝑅𝑂𝐿𝑆 + 𝑆𝐸𝐴𝑆𝑂𝑁𝐴𝐿𝐼𝑇𝑌 + 𝑃𝑅𝑂𝑃𝐸𝑅𝑇𝑌 + 𝜀 .

The inclusion of the period leads, POST(k), allowed us to examine the possible dynamics in the treatment
effect across periods after adopting the treatment. That is, if the treatment had a greater impact a certain number
of months after the treatment adoption, we expected that the estimated coefficient associated with that period
dummy would be greater.
As shown in Table A12, the treatment effect exhibits dynamics in the post-treatment period. Specifically, the
effect size of the treatment increases over time before stabilizing (decreasing). We interpret this as follows: in the
month in which the treatment was adopted, (POST(0) *TREAT), we may have underestimated the treatment
effect, as some bookings may have been made in previous periods. In later periods, the estimated coefficient of
the verified photos initially increases before stabilizing. There are two forces that may lead to this change:
(1) All the demand in these periods is due to the verified photos and not the unverified ones.
(2) With an increased demand, Airbnb ranking algorithms may have started showing this property higher in
search rankings, leading to more demand.
Comparing the effect size in the month following the treatment month (i.e., the coefficient of POST(1)
*TREAT) to the treatment effect obtained from the main model, we see that the results are consistent (8.958
versus 9.303). The coefficient of interaction term with the last period dummy, POST (5) *TREAT, decreased
compared with the estimated coefficient for {𝑃𝑂𝑆𝑇 (𝑘) ∗ 𝑇𝑅𝐸𝐴𝑇} . This is likely because we have fewer
properties for which more than four post-treatment periods are observed (for example, a property that adopted
treatment in February 2017 contributes to the estimation of coefficients for {𝑃𝑂𝑆𝑇 (𝑘) ∗ 𝑇𝑅𝐸𝐴𝑇} , with k
= 0–2 only).

Note that to conclusively rule out or separate any potential effect of Airbnb’s ranking system that could be
confounded with verified photo adoption, we would like to know whether Airbnb’s ranking system considers
property images and, if so, how the inclusion of images affects the ranking algorithm. Unfortunately, we do not
have information about whether and how (if at all) Airbnb increases the ranking of treated properties, as the
algorithm is proprietary. In addition, if the data allow, we would like to use more granular information for
investigating whether Airbnb ranks properties with verified photos higher immediately after the adoption of
verified photos. For example, using the available data, one could examine whether a property received more

57

Electronic copy available at: https://ssrn.com/abstract=2976021

bookings within the next couple of hours or days. Yet, such an analysis is not feasible because our demand and
treatment data are at a monthly level. We knew the month in which a property adopted the treatment condition
but did not have more granular information about the treatment time (e.g., time stamp, exact date, or exact week).
In addition, our demand data were at a monthly level (e.g., number of days booked in a month). As a result, we
could not test whether demand increased within a very short period after the properties adopted verified photos.
We acknowledge that more granular data are needed to resolve this issue. The results should therefore be
interpreted with this caveat in mind.

Table A12. Falsification Checks of Pre-Treatment Trends: Relative-Time Model

DiD Model (Relative Time)

Variables Estimates Robust S.E.
PRE (−4) *TREAT 2.674 2.494
PRE (−3) *TREAT −0.687 3.163
PRE (−2) *TREAT 1.706 3.102
PRE (−1) *TREAT (reference month) -- --
POST (0) *TREAT 6.858* 3.130
POST (1) *TREAT 9.303** 2.917
POST (2) *TREAT 12.78*** 2.932
POST (3) *TREAT 12.92*** 2.918
POST (4) *TREAT 13.00*** 2.927
POST (5) *TREAT 8.818*** 2.609
Property (Non-Photo) Characteristics
log REVIEW_COUNTt-1 9.367*** 0.931
NIGHTLY_RATE −0.145*** 0.0319
INSTANT_BOOK 4.059** 1.364
CLEANING_FEE 0.0808*** 0.0183
MAX_GUESTS 0.0581 1.098
RESPONSE_RATE 0.0705 0.0433
RESPONSE_TIME (minutes) −0.000424 0.00162
MINIMUM_STAY 0.132 0.130
SECURITY_DEPOSIT 0.00162 0.00198
SUPER_HOST 3.698* 1.489
BUSINESS_READY 1.809 0.982
CANCELLATION_STRICT 1.068 1.270
HAS_RATING 13.86 12.19

58

Electronic copy available at: https://ssrn.com/abstract=2976021

 HAS_RATING × COMMUNICATION −0.183 1.425
HAS_RATING × ACCURACY 1.028 1.208
HAS_RATING × CLEANLINESS −1.509 1.136
HAS_RATING × CHECKIN −2.039 1.523
HAS_RATING × LOCATION −0.686 1.176
HAS_RATING × VALUE 2.057 1.180
Fixed Effect Property
Seasonality City-Year-Month
Num. Observations 76,901
R-squared 0.6616
* p < 0.05, **p < 0.01, ***p < 0.001.

4) Adding Interaction Terms with Meaningful Amenities

One concern regarding the self-selection issue is that properties with particular amenities may be more likely to
adopt the treatment. For example, if some amenities make the properties more attractive in particular seasons
(e.g., a pool or beach in summer), and the hosts adopt the treatment at that time, then some of the increase in
demand could be brought about by those attractive amenities. The effects of these amenities on demand (which
are fixed characteristics of the property) cannot be fully accounted for by the property’s fixed effect terms, as
amenities’ effects may be time varying (e.g., a pool has a greater effect in summer than in winter).
To address this concern, when estimating the PSs for PSW, we obtained additional data on the properties’
amenities and incorporated amenity information (e.g., AC, pool, whether the property is close to a beach) to
account for the possible factors that may be correlated with both property demand and the hosts’ treatment
adoption decision in particular seasons but that cannot be captured by property fixed effects. In addition, in the
model specification, we added interaction terms for the dummy AFTER and meaningful amenities (e.g., pool,
beach, AC) to account for the higher effect of these time-invariant variables after the treatment or in particular
seasons.
In Table A13, we present the estimation results. Column (1) presents the results of Equation (2a), which is
our main specification, and column (2) presents the results of Equation (2b), in which we incorporated the
interaction of AFTER with meaningful amenities. The consistency of the estimated results confirms a positive
and significant treatment coefficient of more than an 8% increase in the property occupancy rate after controlling
for area-specific seasonality, as well as the time-varying effect of particular time-invariant property amenities. In
the panel “Interacting with Meaningful Property Amenities,” we present the coefficients of the interaction terms.
As can be seen, all coefficients are insignificant.

59

Electronic copy available at: https://ssrn.com/abstract=2976021

 𝐷𝐸𝑀𝐴𝑁𝐷 = 𝐼𝑁𝑇𝐸𝑅𝐶𝐸𝑃𝑇 + 𝛼 𝑇𝑅𝐸𝐴𝑇𝐼𝑁𝐷 + 𝜆𝐶𝑂𝑁𝑇𝑅𝑂𝐿𝑆 + 𝑆𝐸𝐴𝑆𝑂𝑁𝐴𝐿𝐼𝑇𝑌
+ 𝑃𝑅𝑂𝑃𝐸𝑅𝑇𝑌 + 𝜀 . (2a)
𝐷𝐸𝑀𝐴𝑁𝐷 = 𝐼𝑁𝑇𝐸𝑅𝐶𝐸𝑃𝑇 + 𝛼 𝑇𝑅𝐸𝐴𝑇𝐼𝑁𝐷 + 𝜆𝐶𝑂𝑁𝑇𝑅𝑂𝐿𝑆 + ϑ 𝐴𝐹𝑇𝐸𝑅 × 𝑃𝑂𝑂𝐿
+ ϑ 𝐴𝐹𝑇𝐸𝑅 × 𝐵𝐸𝐴𝐶𝐻 + ϑ 𝐴𝐹𝑇𝐸𝑅 × 𝐴𝐶 + 𝑆𝐸𝐴𝑆𝑂𝑁𝐴𝐿𝐼𝑇𝑌 (2b
+ 𝑃𝑅𝑂𝑃𝐸𝑅𝑇𝑌 + 𝜀 . )

Table A13. DiD Model: Regressing Property Demand on Verified Photos

Variables Main DiD Model Interacting with Amenities

(Equation 2a) (Equation 2b)
Estimates Robust S.E. Estimates Robust S.E.
TREATIND 8.985*** 1.660 8.668*** 1.747
(TREAT × AFTER)
Property (Non-Photo) Characteristics
log REVIEW_COUNTt-1 9.375*** 0.930 9.403*** 0.933
NIGHTLY_RATE −0.146*** 0.0320 −0.149*** 0.0320
INSTANT_BOOK 4.156** 1.361 4.168** 1.362
CLEANING_FEE 0.0808*** 0.0184 0.0814*** 0.0185
MAX_GUESTS 0.260 1.117 0.305 1.111
RESPONSE_RATE 0.0699 0.0430 0.0693 0.0430
RESPONSE_TIME (minutes) −0.000477 0.00161 −0.000591 0.00161
MINIMUM_STAY 0.133 0.131 0.136 0.131
SECURITY_DEPOSIT 0.00177 0.00201 0.00171 0.00199
SUPER_HOST 3.801* 1.494 3.781* 1.494
BUSINESS_READY 1.806 0.985 1.791 0.984
CANCELLATION_STRICT 1.016 1.271 1.040 1.271
HAS_RATING 14.32 12.25 14.56 12.24
HAS_RATING × COMMUNICATION −0.212 1.420 −0.145 1.421
HAS_RATING × ACCURACY 0.878 1.211 0.837 1.211
HAS_RATING × CLEANLINESS −1.344 1.133 −1.323 1.134
HAS_RATING × CHECKIN −2.060 1.526 −2.118 1.526
HAS_RATING × LOCATION −0.757 1.183 −0.767 1.183
HAS_RATING × VALUE 2.141 1.176 2.142 1.173
Interacting with Meaningful Property Amenities
AFTER × POOL 6.157 4.692
AFTER × BEACH −10.77 10.83

60

Electronic copy available at: https://ssrn.com/abstract=2976021

AFTER × AC 1.034 3.990
Fixed Effect Property Property
Seasonality City-Year-Month City-Year-Month
No. of Observations 76,901 76,901
R-squared 0.6608 0.6609
Note: *p < 0.05, **p < 0.01, ***p < 0.001.

5) Testing Changes in the Property or the Host’s Unobserved Quality (via Multidimensional Ratings)

It is possible that properties or hosts self-select the adoption of professional Airbnb images when there are
substantial changes in the quality of the experience delivered to guests. Such changes could include an upgrade
of the house/room or the delivery of more friendly/better services to guests. This would introduce an upward
bias in the estimated coefficient, as it could happen at the same time as the treatment adoption. Although we
were not able to observe and control for everything, we adopted a few measures to help control for or alleviate
this issue.
First, in the demand model, we added the measurement of host responsiveness (host response rate and host
response time) to address the concern that hosts are more responsible in the post-treatment period.
Second, in the demand model, we added a complete set of multidimensional guest ratings to capture and
account for any potential changes in the stay experience or hosting quality.
Third, as we show below, for the properties for which ratings are available, we compared the average ratings
in terms of multidimensional aspects for a few periods before and after the treatment. The goal is to examine
whether there are substantial changes in the property characteristics that were unobserved by us but captured in
guests’ ratings. To implement this robustness test, we estimated the following:

𝑅𝑎𝑡𝑖𝑛𝑔 = 𝐼𝑁𝑇𝐸𝑅𝐶𝐸𝑃𝑇 + 𝛼 𝑇𝑅𝐸𝐴𝑇𝐼𝑁𝐷 + 𝜆𝐶𝑂𝑁𝑇𝑅𝑂𝐿𝑆 + 𝑆𝐸𝐴𝑆𝑂𝑁𝐴𝐿𝐼𝑇𝑌 + 𝑃𝑅𝑂𝑃𝐸𝑅𝑇𝑌 +

𝜀 ,

where the specification is the same as that in our main demand model, and the dependent variable is replaced
with one of the following guest ratings capturing, to some extent, how good a stay or host was: communication,
accuracy, cleanliness, and check-in. The metrics for cleanliness capture potential changes in the property, whereas
those for communication capture the quality of communication between a host and a guest. The coefficient 𝛼
captures changes in ratings along a particular dimension after treatment adoption.
As shown in Table A14, the coefficients of TREATIND are insignificant across all specifications, suggesting
that for the treated units, there was no substantial change in the stay or hosting quality delivered to guests before
and after adopting verified photos.

61

Electronic copy available at: https://ssrn.com/abstract=2976021

 Table A14. Robustness Test: Changes in Multidimensional Guest Ratings after Verified Photo
Adoption

DV: Multidimensional Guest Review Rating

Variables Communication Accuracy Cleanliness Check-in
TREATIND −0.0155 0.0175 −0.0261 −0.0245
(0.0118) (0.0167) (0.0193) (0.0177)
log REVIEW_COUNTt-1 −0.0371 0.0277 −0.0582* 0.0261
(0.0189) (0.0265) (0.0286) (0.0285)
NIGHTLY_RATE 0.00189*** −0.000621 0.00123* 0.000400
(0.000297) (0.000344) (0.000546) (0.000394)
INSTANT_BOOK −0.0134 0.0154 −0.0174 0.0162
(0.0153) (0.0187) (0.0197) (0.0158)
CLEANING_FEE −0.000325 0.000604 −0.000343 −0.000104
(0.000208) (0.000378) (0.000665) (0.000318)
MAX_GUESTS −0.0101 0.00335 0.0449*** 0.0106
(0.00728) (0.00814) (0.0125) (0.00988)
RESPONSE_RATE −0.000951** 0.000425 −0.000782 0.000130
(0.000332) (0.000367) (0.000504) (0.000357)
RESPONSE_TIME (minutes) −0.00000573 0.000000617 −0.00000225 −0.0000180
(0.0000131) (0.0000126) (0.0000163) (0.0000133)
MINIMUM_STAY −0.00129** −0.000554 −0.000550 0.000187
(0.000395) (0.000612) (0.000739) (0.000565)
SECURITY_DEPOSIT 0.00000253 0.0000174 0.0000104 −0.00000397
(0.0000107) (0.0000112) (0.0000168) (0.0000101)
SUPER_HOST −0.0246* 0.0169* 0.0188 0.00510
(0.0117) (0.00826) (0.0154) (0.00905)
BUSINESS_READY −0.00618 0.0166 0.0270 0.00700
(0.00943) (0.0109) (0.0140) (0.00982)
CANCELLATION_STRICT −0.0281 0.0284 −0.0152 0.0236
(0.0147) (0.0164) (0.0200) (0.0194)
Fixed Effect Property Property Property Property
Seasonality City-Year-Month City-Year-Month City-Year-Month City-Year-Month
Observations 45,386 45,386 45,386 45,386
62

Electronic copy available at: https://ssrn.com/abstract=2976021

R-squared 0.8771 0.8931 0.9155 0.8562
Note: The number of observations (45,386) is lower than that in our main analyses (76,901). This is because the
set of analyses presented here used a subset of samples that included only properties with guest ratings presented
on the property page. Note that Airbnb computes and presents ratings only when a property has more than three
guest reviews. As a result, when a property had less than three reviews in a given period, that observation was
automatically dropped from the series of analyses presented in this table.
*p < 0.05, **p < 0.01, ***p < 0.001.

6) Alternative Specifications: Logging All Numeric Variables

We performed a log transformation of all numeric variables (security deposit, price, etc.) and reported the results
of the estimation of the main DiD regression (Table 2 in the main paper). Table 3 presents the estimation results.
As can be seen, the estimated coefficient of the verified photos is consistent with the treatment effect reported
in the main paper.

Table A15. Alternative DiD Model Specification: Regressing Property Demand on Verified Photos

Variables DiD Model

Estimates Robust S.E.
TREATIND 9.135*** (1.647)
log REVIEW_COUNTt-1 9.044*** (0.867)
logNIGHTLY_RATE −6.182*** (1.145)
INSTANT_BOOK 3.725** (1.388)
logCLEANING_FEE 1.898*** (0.491)
logMAX_GUESTS −1.125 (5.412)
logRESPONSE_RATE 5.569*** (1.368)
logRESPONSE_TIME (minutes) −0.0876 (0.258)
logMINIMUM_STAY 1.480 (1.429)
logSECURITY_DEPOSIT 0.280 (0.318)
SUPER_HOST 3.937** (1.493)
BUSINESS_READY 1.666 (0.992)
CANCELLATION_STRICT 1.426 (1.298)
HAS_RATING 13.42 (12.26)
HAS_RATING × COMMUNICATION −0.0991 (1.423)
HAS_RATING × ACCURACY 0.975 (1.224)
HAS_RATING × CLEANLINESS −1.328 (1.118)
HAS_RATING × CHECKIN −2.237 (1.520)
HAS_RATING × LOCATION −0.666 (1.174)

63

Electronic copy available at: https://ssrn.com/abstract=2976021

 HAS_RATING × VALUE 2.267 (1.174)
Fixed Effect Property
Seasonality City-Year-Month
Num. Observations 76,901
R-squared 0.6613
Note: *p < 0.05, **p < 0.01, ***p < 0.001.

7) Including Pre-treated Units in the Control Group

As a robustness test, we included pre-treated units in the control group (which we excluded from our main
analyses for the sake of conceptual clarity, as these pre-treated units had verified photos prior to the start of our
observation window). As such, we estimated the change in property demand when the units to be treated (i.e.,
properties that were untreated prior to our observation window and adopted verified photos during our
observation window) used verified photos compared with the pre-treated and untreated units. The logic is that
the properties (or hosts) in the pre-treated and to-be-treated groups might be more similar in terms of
characteristics in obtaining verified photos. Therefore, including pre-treated units in the comparison group can
serve as a falsification check, as the treatment effect should occur only after a property changes from untreated
to treated. A similar logic was used in previous work (Manchanda et al. 2015; Narang and Shankar 2019) as a
robustness test when self-selection for the treatment was a concern.
Table A16 presents the estimation results. As can be seen, we obtained a consistent estimated treatment effect
(the coefficient for TREATIND is 7.977 and statistically significant) when we included the pre-treated units in
the control group17.
Table A16. DiD Model: Regressing Property Demand on Verified Photos on Property Demand,
Including Pre-treated Properties in the Control Group

Variables DiD Model

Estimates Robust S.E.
TREATIND 7.977*** (1.756)
log REVIEW_COUNTt-1 8.082*** (0.882)
NIGHTLY_RATE −0.126*** (0.0354)
INSTANT_BOOK 4.595*** (1.230)
CLEANING_FEE 0.107*** (0.0237)
MAX_GUESTS 1.385 (1.215)

17Note that although that the effect size was 11% lower (compared with the value of 8.985 that we obtained from the main
DiD regression), the estimated treatment effect was very close, and it remained at the same level. For example, Narang and
Shankar (2019) reported that the estimated effects from their robustness test were 16%–35% lower than their main results.
64

Electronic copy available at: https://ssrn.com/abstract=2976021

 RESPONSE_RATE 0.0735 (0.0448)
RESPONSE_TIME (minute) −0.000490 (0.00172)
MINIMUM_STY 0.198 (0.171)
SECURITY_DEPOSIT 0.00313 (0.00223)
SUPER_HOST 3.346* (1.495)
BUSINESS_READY 0.969 (0.963)
CANCELLATION_STRICT 1.362 (1.305)
HAS_RATING 12.37 (16.05)
HAS_RATING × COMMUNICATION −0.637 (1.482)
HAS_RATING × ACCURACY 0.712 (1.236)
HAS_RATING × CLEANLINESS −0.437 (1.299)
HAS_RATING × CHECKIN −1.162 (1.610)
HAS_RATING × LOCATION −1.165 (1.376)
HAS_RATING × VALUE 1.632 (1.305)
Fixed Effect Property
Seasonality City-Year-Month
Observations 123186
R-squared 0.6611
Note: The number of observations here is larger than that in the main analyses (76,901). This is
because the regression presented in this table was performed on a PSW-matched sample that included
pre-treated units in the control group.
*p < 0.05, **p < 0.01, ***p < 0.001

Using Only Pre-treated Units as the Control Group

In Table A8, we present the robustness checks by including the pre-treated units in the control group. In this
subsection, we use a stronger control—using the pre-treated units only. Similar to the robustness test in the
studies of Manchanda et al. (2015) and Narang and Shankar (2019), this analysis provides a conservative view by
comparing the dependent variable among adopters only.

Table A17 presents the estimation results. As can be seen, we obtained consistent results (the estimated
coefficient of TREATIND is 6.77 and significant at the 0.05 significance level) when using only the pre-treated
units as the control18. This robustness analysis, along with Table A8, suggests that our PSW combined with DiD
can mitigate the self-selection issue in the verified photo adoptions.

18Note that despite the effect size being 24% lower, the estimated effect was very close and remained at the same level (e.g.,
Narang and Shankar [2019] reported that the estimated effects from this robustness analysis were 16%–35% lower than
their main results across multiple DVs of interest).
65

Electronic copy available at: https://ssrn.com/abstract=2976021

Table A17 Regressing Property Demand on Verified Photos on Property Demand Using Only Pre-
treatment Properties as a Control Group

VARIABLES DiD Model

ESTIMATES Robust S.E.
TREATIND 6.769*** (1.011)
log REVIEW_COUNTt-1 5.943*** (0.345)
NIGHTLY_RATE −0.0433*** (0.0102)
INSTANT_BOOK 4.259*** (0.488)
CLEANING_FEE 0.0228* (0.0107)
MAX_GUESTS −0.185 (0.297)
RESPONSE_RATE 0.0245 (0.0169)
RESPONSE_TIME (minute) −0.000291 (0.000650)
MINIMUM_STY 0.00502 (0.0503)
SECURITY_DEPOSIT −0.0000970 (0.000710)
SUPER_HOST 1.393** (0.487)
BUSINESS_READY 0.402 (0.354)
CANCELLATION_STRICT −1.046 (0.844)
HAS_RATING 5.406 (6.542)
HAS_RATING × COMMUNICATION 0.150 (0.617)
HAS_RATING × ACCURACY −0.118 (0.490)
HAS_RATING × CLEANLINESS −0.302 (0.424)
HAS_RATING × CHECKIN −0.299 (0.586)
HAS_RATING × LOCATION −0.275 (0.430)
HAS_RATING × VALUE 0.709 (0.456)
Fixed Effect Property
Seasonality City-Year-Month
Observations 51918
R-squared 0.6814
Note: The number of observations is different from that in the main analysis (76,901), as this
regression was performed on a different sample, in which the 4,932 pre-treated units (i.e., properties
that adopted verified photos prior to our observation) formed the control group. The DV is the
property monthly occupancy rate.
*p < 0.05, **p < 0.01, ***p < 0.001

66

Electronic copy available at: https://ssrn.com/abstract=2976021

8) Additional Analyses
i. Regressing Property Price as the DV
We examined how price changed in the post-treatment period. We replicated our DiD regression but replaced
the DV with the property’s nightly price to estimate the coefficient of verified photos on price. As shown in
Table A18, the insignificant coefficient of TREATIND is positive, suggesting an upward change in price.
However, the coefficient is insignificant at a significance level of 95%. The results imply that after controlling for
the property fixed effects and city-specific seasonality and yearly effects, a property’s price does not change the
price based on having obtained verified photos.
Note that this result does not mean that a property’s price cannot increase or change during the post-treatment
period. For example, a property’s price may increase when it has accumulated more reviews or the host has gained
more experience, or as a response to seasonality effects. Rather, the insignificant coefficient of TREATIND in
Table A18 should be interpreted as indicating that controlling for all other factors (time-varying property
characteristics, city-specific seasonality), a property’s price would not increase solely because of the use of verified
photos. A possible explanation is that changing photos is not automatically associated with improvements in
other aspects of the property and host. Thus, everything else being equal, a host would not increase the price
after adopting verified photos. Another explanation is that many Airbnb hosts lack rich information about
demand and price, so they rely on Airbnb’s smart pricing algorithm to automatically set prices. As a result, it is
possible that the price would be higher because of the use of verified photos if the smart pricing algorithm did
not include the property pictures as an input. Of course, Airbnb does not disclose the details of its proprietary
algorithm. Our understanding is that the algorithm considers many factors, such as seasonality, market popularity,
and review history. In Airbnb’s report on its algorithm, property pictures were not included or described as one
of the many factors at play in the algorithm.19

Table A18. Regressing Property Price on the Treatment Indicator: DiD Model

Regressing Property Price as the DV

Variables Estimates Robust S.E.
TREATIND 4.388 (3.047)
log REVIEW_COUNTt-1 −14.29*** (2.245)
INSTANT_BOOK −6.833** (2.145)
CLEANING_FEE 0.721*** (0.0514)
MAX_GUESTS 4.898 (2.821)

19 See https://blog.atairbnb.com/smart-pricing/.

67

Electronic copy available at: https://ssrn.com/abstract=2976021

 RESPONSE_RATE 0.241** (0.0900)
RESPONSE_TIME (minute) 0.000684 (0.00445)
MINIMUM_STY −0.947** (0.310)
SECURITY_DEPOSIT 0.0443 (0.0237)
SUPER_HOST −4.432** (1.569)
BUSINESS_READY 1.909 (2.344)
CANCELLATION_STRICT −8.532*** (2.318)
HAS_RATING −142.2*** (21.38)
HAS_RATING × COMMUNICATION −3.999 (2.071)
HAS_RATING × ACCURACY 3.298 (2.604)
HAS_RATING × CLEANLINESS 7.547*** (1.775)
HAS_RATING × CHECKIN 3.812 (2.279)
HAS_RATING × LOCATION 11.37*** (1.899)
HAS_RATING × VALUE −8.340** (2.677)
Fixed Effect Property
Seasonality City-Year-Month
Observations 76901
R-squared 0.8810
Note: The DiD model was estimated by regressing property price on the same set of variables as the
main DiD model, except for instrumented price on the right-hand side, which was excluded. Regression
was performed on a dataset including properties that had at least one open day in a month.
*p < 0.05, **p < 0.01, ***p < 0.001.
ii. Including Review Probability in the PSW and DiD Model
Zhang et al. (2019) found that image quality affects the probability that a customer (after the stay) will write a
review for that property. We follow Zhang et al. (2019) and investigate the aspect of review probability as a key
variable affecting property demand. To do so, we match properties based on review probability to control for this
pre-treatment characteristic. In the DiD model, we also control for review probability (the review probability
from the previous month). The review probability for a property in each month is computed as the number of
received reviews divided by the reservations in that month.

As can be seen in Table A19, the estimated coefficient of TREATIND (b = 7.271, p < 0.001) is consistent
with our main results20. In addition, the estimated coefficient of REVIW_PROBABILITY is insignificant at the
0.05 significance level.

20 Note that this regression is estimated on those observations in which review probability can be defined and computed, that
is, the observations when a property in a month had at least one booking. By including review probability in the model, we
automatically excluded those observations with zero booking (demand is 0) in a month, which is more likely for untreated
properties than for treated properties. As a result, the coefficient of TREATIND in this analysis is a conservative estimate.
68

Electronic copy available at: https://ssrn.com/abstract=2976021

 Our interpretation is that although in our main analyses, we did not explicitly include review probability
in the PSW and DiD analyses, this variable was indirectly controlled for. This is because review probability can
be computed from (as a function of) a set of covariates, including image quality, host response time, price, and
other property characteristics (Zhang et al. 2019). These covariates were exactly used in the PSW model and the
DiD model. Furthermore, when we thought about how review probability may influence property demand, we
found that review probability should not have an impact on the present demand. This is because review
probability affects the evolution of Number of Reviews, which is a key driver in attracting bookings. Following the
arguments and findings of Zhang et al. (2019), review probability affects a property’s future demand by affecting
the speed of accumulating reviews. We believe that review probability alone should not directly affect demand (in
fact, Zhang et al. (2019) did not include this variable in their demand model). Although review probability is
related to property quality and guest satisfaction, it is unobserved by guests. We can compute review probability
because we obtain the # reviews and # bookings in each month. For an average Airbnb guest, however, such
information is unknown (a guest observes the total number of reviews but does not know how many bookings
a property has received). As a result, guests cannot use review probability as an important variable to aid their
booking decisions. It then follows that review probability alone should not have an impact on the present demand
for a property.

Table A19 Including Review Writing Probability in Demand Regression

VARIABLES DiD Model

ESTIMATES Robust S.E.
TREATIND 7.271*** (2.022)
log REVIEW_COUNTt-1 7.362*** (1.249)
NIGHTLY_RATE −0.149*** (0.0374)
INSTANT_BOOK 5.775*** (1.501)
CLEANING_FEE 0.0779** (0.0279)
MAX_GUESTS −0.312 (1.263)
RESPONSE_RATE 0.120 (0.0670)
RESPONSE_TIME (minute) −0.00152 (0.00245)
MINIMUM_STY 0.227 (0.158)
SECURITY_DEPOSIT 0.00486 (0.00261)
SUPER_HOST 2.094 (1.576)
BUSINESS_READY 0.503 (1.171)
CANCELLATION_STRICT 0.666 (1.557)

69

Electronic copy available at: https://ssrn.com/abstract=2976021

 HAS_RATING 21.64 (20.22)
HAS_RATING × COMMUNICATION −1.632 (1.842)
HAS_RATING × ACCURACY −0.430 (1.588)
HAS_RATING × CLEANLINESS −0.865 (1.336)
HAS_RATING × CHECKIN 0.885 (1.807)
HAS_RATING × LOCATION −1.783 (1.620)
HAS_RATING × VALUE 1.932 (1.292)
REVIW_PROBABILITYt-1 2.06 (1.507)
Fixed Effect Property
Seasonality City-Year-Month
Observations 40590
R-squared 0.5797
Note: The DV is the property monthly demand. The number of observations in this regression is
different from that in the main analysis (76,901). This is because this regression includes review
probability in the previous month as a control variable. Therefore, the model was regressed on
observations in which a property had at least one booking in the previous month (other: in addition
to this, the property had at least one open day in the current month).
*p < 0.05, **p < 0.01, ***p < 0.001

9) Testing Host Behavior – Changes in the Number of Open Days

One self-selection concern is that hosts may change their behavior (in a way that will increase the property
booking) along with the treatment adoption. To test this, we used the number of open days to serve as a
falsification test for endogeneity. Note that the treatment and control properties differ in the observed variables
(the treatment is endogenous in our setting). To address this, we performed a propensity score-based adjustment
to make the treatment and control groups comparable on the observed measures. If the propensity score-
weighted treatment and the control properties were to exhibit a different number of nights open subsequent to
the treatment, then it would indicate that part of the treatment may be driven by hosts’ desire to open for more
nights.
To check this, we produced the following results. Here, we regressed open nights as a function of the
treatment indicator and other controls (the dependent variable was log (#open nights +1)). We found that the
treatment had no impact on the number of open nights. Hosts’ adoption of treatment did not appear to be driven
by the intent to open the property for more days in the future. As shown in Table A20, the insignificant
coefficients across the two model specifications suggest that the increase in demand was not introduced by hosts
being more likely to open nights after adopting Airbnb professional photos. Note that in the model specification
in column (2), we added the number of reservation days in the previous period as an additional control.
70

Electronic copy available at: https://ssrn.com/abstract=2976021

 As another robustness test, we include # Open Days in the DiD demand model, control for possible changes
in the number of open days. As presented in Table A21, the estimated coefficients are similar to our main DiD
estimation results. Specifically, the estimated coefficient of TREATIND on demand was 8.984 (p<0.001).

Table A20 Robustness Test: Changes in # of Open Days along with Verified Photo Adoption

DV: log # Open Days in Period t

VARIABLES Model (1) Model (2)
TREATIND 0.0201 −0.0267
(0.0339) (0.0317)
log # RESERVATION_DAYSt-1 0.194***
(0.00831)
log REVIEW_COUNTt-1 −0.00463 −0.0650***
(0.0181) (0.0176)
NIGHTLY_RATE 0.00140* 0.00184***
(0.000554) (0.000517)
INSTANT_BOOK 0.0431 0.01000
(0.0226) (0.0213)
CLEANING_FEE −0.000765* −0.00103**
(0.000359) (0.000346)
MAX_GUESTS 0.0112 0.00981
(0.0232) (0.0204)
RESPONSE_RATE 0.000359 0.000376
(0.00106) (0.000983)
RESPONSE_TIME (minutes) −0.0000379 −0.0000210
(0.0000331) (0.0000307)
MINIMUM_STAY −0.00824** −0.00884**
(0.00296) (0.00269)
SECURITY_DEPOSIT −0.0000389 −0.0000309
(0.0000398) (0.0000369)
SUPER_HOST −0.0588 −0.0759*
(0.0342) (0.0297)

71

Electronic copy available at: https://ssrn.com/abstract=2976021

BUSINESS_READY 0.00305 −0.00470
(0.0199) (0.0177)
CANCELLATION_STRICT −0.0156 −0.0277
(0.0270) (0.0254)
HAS_RATING −0.157 −0.255
(0.239) (0.219)
HAS_RATING × COMMUNICATION −0.114*** −0.100***
(0.0304) (0.0276)
HAS_RATING × ACCURACY 0.0581* 0.0439
(0.0267) (0.0257)
HAS_RATING × CLEANLINESS −0.000146 0.0128
(0.0224) (0.0207)
HAS_RATING × CHECKIN 0.0426 0.0531
(0.0317) (0.0280)
HAS_RATING × LOCATION 0.0242 0.0230
(0.0243) (0.0221)
HAS_RATING × VALUE −0.00595 −0.0193
(0.0233) (0.0214)
INTERCEPT 2.913*** 2.729***
(0.144) (0.134)
Fixed Effect Property Property
Seasonality City-Year-Month City-Year-Month
Observations 76,901 76,901
R-squared 0.4444 0.5092
Note: The difference between the two specifications is that in Model (2), we control for the number
of reservation days in the previous period; robust standard errors are in parentheses; *p < 0.05, **p <
0.01, ***p < 0.001.

Table A21 Robustness Test: Including # Open Days in Demand Model

VARIABLES Main DiD Model (Eq. 2)

ESTIMATES Robust S.E.

TREATIND 8.984*** (1.659)
log REVIEW_COUNTt-1 9.377*** (0.931)

72

Electronic copy available at: https://ssrn.com/abstract=2976021

 NIGHTLY_RATE -0.147*** (0.0320)
INSTANT_BOOK 4.150** (1.360)
CLEANING_FEE 0.0809*** (0.0185)
MAX_GUESTS 0.260 (1.117)
RESPONSE_RATE 0.0699 (0.0430)
RESPONSE_TIME (minutes) -0.000472 (0.00161)
MINIMUM_STAY 0.133 (0.131)
SECURITY_DEPOSIT 0.00178 (0.00201)
SUPER_HOST 3.808* (1.492)
BUSINESS_READY 1.806 (0.985)
CANCELLATION_STRICT 1.016 (1.271)
HAS_RATING 14.32 (12.24)
HAS_RATING × COMMUNICATION -0.203 (1.419)
HAS_RATING × ACCURACY 0.873 (1.208)
HAS_RATING × CLEANLINESS -1.344 (1.133)
HAS_RATING × CHECKIN -2.063 (1.526)
HAS_RATING × LOCATION -0.761 (1.182)
HAS_RATING × VALUE 2.143 (1.176)
# OPEN DAYS 0.00610 (0.0583)
Fixed Effect Property
Seasonality City-Year-Month
Num. Observations 76,901
R-squared 0.6608
Note: the D.V. is computed as a function of Open Days (occupancy rate = #reservation days/#open days).
*p < 0.05, **p < 0.01, ***p < 0.001.

VI. Data Description and Sample Construction

We started with over 13,000 unique Airbnb properties in seven US cities, collecting images from the property
pages in each month from January 2016 to April 2017. Of the full sample, 4,932 properties were pre-treated (i.e.,
had verified photos prior to our observation in January 2016). For the analyses presented in this study, we
removed these properties from our sample for the sake of conceptual clarity (in an additional robustness test, we
included these units in the control group and obtained consistent results; see Section V.7) in this Appendix for
detailed results and discussions). The remaining properties totaled 8,858 (13,790 – 4,932). As described in the
main paper, we removed missing data points and properties for which there were web page errors during data
collection. Errors may have occurred because Airbnb blocked our data collection requests, and/or hosts unlisted
their properties from Airbnb temporarily or permanently.

Next, using the valid units, we performed a propensity score analysis to match the treated and untreated units
by comparison (matching) based on a set of observed covariates. Of the 7,825 valid properties identified during
the observational window, there were 231 treated units and 7,594 untreated units before PSW. Next, we
performed a propensity score estimation to find properties in the two groups that were close or matched with
each other in terms of the property and host characteristics in the pre-treatment period. We first estimated the
propensity score as a logit function of a set of property, host, and neighborhood covariates measured at the start
73

Electronic copy available at: https://ssrn.com/abstract=2976021

of the first period (i.e., the pre-treatment period). The set of variables used in the propensity score estimation
included time-invariant characteristics (e.g., number of bedrooms, property type, property amenities) and time-
varying dynamic characteristics (e.g., property price, number of reviews).
We first denoted the set of variables for the units to be matched on as X. Following prior literature that
applies the propensity score approach, we computed the propensity score as a logit function of X. Unit i has an
estimated propensity score of 𝑝𝑠 = 𝑓(𝑋 ). Then, as we described in the paper, we computed a weight for each

unit based on its estimated propensity score. Specifically, we calculated weight as 𝑤 = if i was a treated unit

and as 𝑤 = if i was an untreated unit. As we described in the paper, this is a widely used PSW approach

called inverse probability treatment weighting. Later, we used the sample weights for DiD regressions.
Furthermore, to address the concern that a very low estimated probability can result in extremely large (and
possibly unstable) weights, we followed common practice in PSW and used trimmed weights, excluding those
that were outside 1% and 99% of the distribution (Austin and Stuart 2015; Cole and Hernan 2008). The PSW
generated a sample that included 7,211 of the 7,594 untreated units and 212 of the 231 treated units. This matched
sample was used in the main regressions (Tables 2, 3, 4, and 7 in the main paper). The total number of
observations used in the analyses was 76,90121.

21Note that the total observations in the regressions was fewer than # units * # periods because for properties that did not
have open days in a month, the dependent variable—property demand (occupancy rate)—would be indefinable.
74

Electronic copy available at: https://ssrn.com/abstract=2976021

VII. Example of Property Images with 1 SD of Improvement in Key Image Features

As an example, we show what an increase of 1 SD may look like visually in Figure A1. The figure shows the
original photos and a 1 SD increase for two example attributes—brightness and saturation.

Figure A1 Examples of Images with a 1 SD Increase in Image Attributes

Attribute Original Photo +1 SD in Attribute

Brightness

Saturation

75

Electronic copy available at: https://ssrn.com/abstract=2976021