Overview

HUMBEHV 3ST3
• Part 1 - Introduction: What is Statistics?
Basic Concepts & Graphical Displays • Part 2 - Basic Concepts: Variables & measurement scales
Week 1 • Part 3 - Graphical displays & shapes of distributions
Prof. Patrick Bennett

Part 1: Introduction What is “statistics”?

• Statistics refers to statistical methods & procedures…

• and to the results/outputs of those methods

- Averaging is a statistical method and the result of averaging

data in a sample — i.e., the mean — is a statistic.
Role of statistics in research Role of statistics in research
Statistics helps us to collect, organize,
summarize, analyze, interpret, & present data

“The role of statistics is not to discover

truth. The role of statistics is to resolve
“Statistics is the grammar of
disagreements between people.” -
science.” - Karl Pearson
Milton Friedman

Role of statistics in research Role of statistics in research (PPDAC)

“...the purpose of statistics is to

organize a useful argument from David Spiegelhalter

quantitative evidence, using a form of

principled rhetoric*.” - Robert P. Abelson
*rhetoric: the art of effective/persuasive
speaking or writing
Understanding & defining the problem of interest
Statistics is used in many contexts

• Statistics & data-based decision making go hand-in-hand

• Statistics plays vital roles in:
- many areas of research in engineering, natural sciences,
social sciences, & humanities
- machine learning & data mining; artificial intelligence
- development of public policy in a wide range of areas

Alan Smith TEDxExeter

Why you should love statistics

https://www.ted.com/talks/alan_smith_why_we_re_so_bad_at_statistics

Statistics is the science of us, and that is why we should be fascinated by numbers.
 Many Ways of Collecting Data Data Management & “Cleaning”
Statistics plays important roles at beginning of studies

• Two ways we will consider in this course: • Brain-imaging studies often use only right-handed adults participants
• Correlational Studies - - what is an “adult”? how should we measure “handedness”?
- measure associations between predictor & criterion variables • UK study compared survival rates of children who had heart surgery
- “subjects” come with their own set of variables - what is a “child”? what counts as “heart surgery”? when can a
• Designed Experiments - death be attributed to the surgery?
- measure effects of independent variables on dependent variables • Political surveys often measure preferences of likely voters… how
- random assignment of “subjects” to experimental conditions should we define or measure a “likely” voter?

Descriptive Statistics Z. Hussain et al. / Vision Research 49 (2009) 2624–2634 2627

Populations vs. Samples
Data Summary & Communication ðD ¼ 0:26; CI95% ¼ ½0:21; 0:31&Þ, and declined to an average of (inferential vs. descriptive statistics)
Textures 0.13, CI95% ¼ ½0:11; 0:15& in the last three bins. Nevertheless, as
0.6

was the case with textures, the simple main effect of Day was sig-
nificant at each Bin (tð26Þ P 5:42; p < :0001 in all cases). Again,
0.5

• describes/summarizes important characteristics of data the analyses suggest that there was more within-session learning

• Population: all events (subjects, scores, etc) of interest

Proportion Correct

on Day 1 than on Day 2 for this group.

0.4

Inspection of Fig. 3 shows that average response accuracy on

• uses graphs & statistics e.g., mean or standard deviation Day 1 was significantly greater in the face condition than in the
0.3

texture condition ðCI95% ¼ ½0:02; 0:16&; tð53Þ ¼ 2:49; p ¼ 0:015Þ.

• describes interesting features of sample

On Day 2, average response accuracy also was numerically higher
• Sample: subset of population
0.2

for faces than for textures, but the difference between the groups
40-trial was not statistically significant ðCI95% ¼ ½' :03; :14&; tð53Þ ¼
10-trial
1:24; p ¼ 0:22Þ.
0.1

5-trial
1-trial The current results are consistent with previous reports that 40
trials per condition on Day 1 are sufficient to produce learning in - random sample: each member of population has equal chance
0.0

these texture- and face

et al.identification tasks (Hussain et al., 2005,
10 Perceptual
20 50 Learning
100 200 500 1000
C.M. Gaspar / Vision Research 48 (2008) 1084–1095
Face Perception 1089
2009a).
For each level of stimulus contrast variance used during trials with inversion of being selected
0.35

external noise, a percentage of agreement, Pa, was calculated for replicated contrast
3.3. Effects ofalso
reduced practice:
for eachtexture identification
Faces
0.6

trials. A percentage of correct responses, Pc, was estimated

combined
0.30

stimulus contrast by using the fitted Weibull (psychometric) function

described earlier. By pairing Pa and In this section,
Pc according and contrast,
to stimulus the next, we compare response accuracy
- convenience sample (e.g., psychology undergraduates)
0.5

we were thus able to obtain a unique mappinginbetween

measured Pa and on
all groups Pc. An
Day 2. These analyses addressed the
0.25
log10(threshold ratio)
Proportion Correct

observer modeled with different levels of internal noise, relative to a con-

issue of whether any exposure to textures or faces on Day 1 im-
stant amount of externally added noise, responds with systematic changes
0.4

proved
to the slope s of this equation (Gold performance relative to the 0-trials groups, and whether
et al., 1999b):

• Can we infer characteristics of population from sample?

0.20

groups that received 1–10 trials per condition performed worse

P c ¼ log10 ðP a =100Þs þ 100 ð2Þ
0.3

than the 40-trials groups. Fig. 4 clearly shows that the 10- and
0.15

The relationship between internal 40-trials

noise and s groups performed
was measured better than the 0-trials group. The
by running
0.2

Monte Carlo simulations of a ideal observer performing in this experiment

for40-trial
50 different levels of internal noise. By comparing a participant’s slope
0.10

20-trial
to the modeled observer’s slope, we were thus able to obtain an estimate of
0.1

10-trial
their total internal noise (ri), relative to external noise (rn). This inter-
Day 2 (Textures)
0.7

0.05

5-trial
nal:external noise ratio, ri/rn, was calculated for each participant in all
1-trial
conditions.
0.0

0.6

0.00
Proportion Correct

10 20 50 100 200 500 1000 none high

3. Results
Trial Number (Day 1) External Noise
0.5

Statistical analyses were performed with R v2.5.1 (R

MeanFig.face identification
2. Proportion correct on Dayaccuracy
1 plotted as ain 5 groups
function plotted
of trial number. as a function of
Proportion Mean face
Fig. 6.identification
Effects of inversion thresholds in six
and contrast-reversal on experimental
face identification conditions.
Development Core Team, 2007). The strength of associa- thresholds. Threshold ratios were calculated by dividing thresholds in the
the number
correct was of trials.
measured for Hussain et al,
blocks consisting of 102009, Viseach
trials, and Research.
point represents Error bars are ±1 SEM. Gaspar et al, 2008, Vis Research.
0.4

tion between the dependent and independent variables

the average taken across subjects. The solid line in each figure represents the least- inverted, contrast-reversed, and combined conditions by thresholds in the
squares regression line fit to the data. was expressed as partial omega-squared (x2p ) using formu- normal-face Trials/Condition
condition. Threshold ratios were averaged across stimulus
lae described by Kirk (1995). When appropriate, the 40-trials±1 SEM.
gender. Error bars represent
0.3

Huynh-Feldt estimate of sphericity, ~!, was used to adjust 10-trials

p values of F tests conducted on within-subject variables 5-trials
1-trials
0.2

(Maxwell & Delaney, 2004).

0.7

were averaged0-trials
across face gender. A 2(External
Noise) % 3(Condition) within-subjects ANOVA on the
3.1. Face identification thresholds 9 10 11 12log-transformed
13 14 ratios
15 revealed
16 a significant main effect
0.6

of Condition (F(2, 12) = 6.04, ~! ¼ 1, p = 0.0153,

Bin
ct

2
 Inferential Statistics Descriptive & Inferential Statistics
Bennett, PJ PSYCH 710

• uses sample to make claims about a population population

population parameters (Inferential Statistics)
µ, 2

- e.g., estimate population parameters from sample statistics

- e.g., investigate differences among population by examining
statistical theory
differences among groups/samples e.g., sampling
distributions
• Psychology experiments typically test undergraduate students
& generalize results to all young adults Bennett, PJ PSYCH 710
sample statistics
sample mean, variance (Descriptive Statistics)
Ȳ , s2

Exploratory vs Confirmatory Analyses Part 1 Summary

• Exploratory Data Analysis • what is statistics?

- first major proponent was John Tukey • why study statistics?

- goal: discover & summarize interesting aspects of data • ways of collecting data

- discover interesting hypotheses to test - designed experiments vs. correlational studies

John Tukey
• Confirmatory Data Analysis - causal effects vs. associations

- data are gathered & analyzed to evaluate specific a priori hypotheses • descriptive vs inferential statistics

- example: clinical drug trials - pop vs sample

• Important not to confuse two types of analyses - params vs statistics

- replication crisis in Psychology related to confusion about two types of research • exploratory vs confirmatory statistics