Toolkit Note

Testing and adjusting for attrition in household

panel data
Bob Baulch (CPRC)
Agnes Quisumbing (IFPRI)

Abstract
This note describes the use of a simple procedure to correct for attrition due to observables
in household panel survey: inverse probability weights. The procedure involves estimating
two probit regressions, one with and one without variables that are significantly associated
with attrition, and using the ratio of predicted probabilities from these regressions to reweight
the observations. The procedure is illustrated in Stata using data from part of the CPRCDATA-IFPRI panel in rural Bangladesh.

Introduction
Attrition has been described as the panel researchers nightmare (Winkels and Withers,
2000). This is because if the members who drop out of a panel differ systematically from
those who stay in it, then the dataset of continuing members is no longer representative of
the original population. So results based on data in which only continuing panel members
are included may be seriously affected by attrition bias. Fortunately, a number of studies
suggest there is a simple method of adjusting for sample attrition (at least when it is based
on observable characteristics) known as inverse probability weights (Fitzgerald et al., 1988;
Wooldridge, 2002). This Toolkit note aims to provide a simple introduction and illustration of
how to test and adjust for attrition bias in panel data using the statistical software Stata.
To fix ideas consider a household panel consisting of i.N households who have been
surveyed in two different years (t=1, 2). Denoting the outcome of interest for household i in
the second year by yi2, household variables in the first year by xi1, and additional instrumental
variables that only affect attrition by zi1, we may write:
(1)

yi 2 xi1 i

(2)

A* xi1 zi1 vi

yi2 observed if A*>1

CPRC Toolkit Note

This set-up resembles a standard one-period selection model except that the outcome
variable is measured in the second period. In practice, however, the probability of attrition,
A*, is not observed and is replaced by an attrition dummy, A, which takes the value 0 when
both yi1 and yi2 are observed, and one when yi2 is not observed. So one possible solution to
sample attrition is to estimate a selection model, which relies on identifying a set of
instrumental variables, zi , which are correlated with attrition but not with i (Heckman,
1979).1 This is often referred to the case of selection on unobservables. However, it is often
difficult to identify suitable instrumental variables for selection models. A second solution to
sample attrition is to estimate inverse probability weights, which relies on an auxiliary
variable(s) which can be related to both attrition and the outcome variable (Fitzgerald et al,
1998). This is the case of selection on observables, and requires a much weaker condition
for the z variables: that i and i are uncorrelated. To estimate such inverse probability
weights, equation (2) is respecified as a probit:
(3)

A xi1 ai1 vi

where A=0 for households who remain in the sample and A=1 for attritors, and ai1 are the
auxiliary variables in the first period. Next, a restricted version of equation is re-estimated
without the auxiliary variables:
(4)

A xi1 i

The ratio of the predicted values from equation (4) and equation (3) give the inverse
probability weights:

(5)

Wi

pr
pu

The intuition behind this procedure is that it gives more weight to households who have
similar initial characteristics to households that subsequently attrit than to households with
characteristics that make them more likely to remain in the panel.
The question that then arises is what variables are suitable for inclusion in zi or ai? Clearly
these variables must be observed for both panel households and attritors, and be correlated
with the probability of attrition. In selection models, lagged values of the outcome variable
are often used as instrumental variables, but this typically requires that at least three waves
of panel data are available. Measures of the quality of the interview are often included
among the z variables (Maluccio, 2004) as these seem likely to be related to the probability
of attrition but are not necessarily to the outcome variable. The auxiliary variables used to
calculate inverse probability weights can also include household demographic variables,
community level variables, shocks or treatment variables (for example, whether a household
1

Note that i and vi are not uncorrelated in this case.

CPRC Toolkit Note

receives a transfer payment). Attrition is often found to be related to the age of the
household head, or the demographic composition of the household and also to shocks or
treatment variables. As these variables are usually correlated with the outcomes,
demographic, community and shock variables cannot be used in the selection model in
equations (1) and (2) which requires z to be instrumental variables which are correlated
with attrition but not the outcome variable. However, they can be used in the ai that are used
when calculating inverse probability weights. In our application below, we include a variable
for community level shocks and households technology adoption status in year 1 as well as
the household demographic and thana (sub-district) dummies.

Testing whether attrition is random

Prior to calculating inverse probability weights, it is first essential to test whether attrition in a
panel data model is random. There may be situations in which attrition is entirely random
and, in this fortunate situation, it is not necessary to do anything further.
A number of tests have been proposed for whether attrition in a panel is random, including
attrition probits (Fitzgerald et al, 1998) and pooling tests, in which the equality of coefficients
from the baseline sample with and without attritors are equal (Becketti, Gould, Lillard and
Welch, 1988). We implement both of these tests in this note.2
The simplest test for whether attrition is random is to estimate a probit in which the
dependent variables takes the value one for households which drop out of the sample after
the first wave (attrit) and zero otherwise. Explanatory variables are baseline values for all
variables that are believed to affect the outcome variable of interest plus any available
variables which characterise the interview process. It is usual to include lagged values of the
outcome variable in such attrition probits. As pointed out by Outes-Leon and Dercon (2008),
it is also useful to examine the pseudo R-squared from attrition probits, as they can be
interpreted as the proportion of attrition that is non-random.
Another commonly used test for whether attrition is random is the pooling test due to
Becketti, Gould, Lillard and Welch (1988). The Becketti, Gould, Lillard and Welch (hereafter
BGLW) test involves regressing an outcome variable from the first wave of a survey on
household and community variables, an attrition dummy, and the attrition dummy interacted
with the other explanatory variables. An F-test of the joint significance of the attrition dummy
and the interaction variables is then conducted to determine whether the coefficients from
the explanatory variables differ between households who are retain or attrit from the panel.

Other tests for attrition include parametric selectivity models (Falaris, 2003) and examining the significance of
lagged dependent variables.

CPRC Toolkit Note

It is important to understand that these tests are model specific and needs to be repeated for
each outcome variable of interest. Thus in our application to data from rural households in
Bangladesh, we calculate separate tests for expenditures and assets.

Application
To illustrate the calculation of inverse probability weights we use the agricultural
technologies portion of the CPRC-DATA-IFPRI panel from rural Bangladesh. This panel
spans the ten year period from late 1996 to late 2006/early 2007, and contains just over
1,300 households located in four of Bangladeshs 64 districts: Manikganj, Kishoreganj,
Jessore, and Mymensingh. The survey was clustered at the village level, and there are 47
villages included in it (see http://www.ifpri.org/dataset/chronic-poverty-and-long-term-impactstudy-bangladesh for further details of this dataset, which is publicly accessible).
The Stata dataset Bangladesh_example.dta contains observations for 965 households in
1996 (the baseline), 47 of whom drop out of the sample between 1996 and 2006. Thus
attrition from the panel at the household level is just under 5%. There are also 10
households for whom we do not have expenditure data in 1996, who are included among the
attritors in the expenditure model below.3 In addition to the usual variables on the
households demographic characteristics, age and education of the household head, asset
ownership and location (thana) variables, we have detailed information on four variables that
may be correlated with attrition. These are: (i) lagged values of the dependent variable (per
capita expenditures or the value of assets owned by the household); (ii) the percentage of
households in the village experiencing a flood between 1996 and 2006; (iii) the village level
attrition rate during the four rounds of the survey conducted in 1996, which is taken as a
indicator of interview quality;4 and, (iv) the adoption status of household with respect to the
agricultural technologies (introduced vegetables, individual and group fishponds) that were
the focus of the original study. All these variables are observed for both attritors and
households that remain in the sample. Note that with the exception of the village attrition
rate, these variables could not be included in a selection type model, as they are correlated
with both attrition and the outcome variable
Each of these variables are included in the attrition probit for expenditures produced by the
following Stata command and reported in Table 1 below:

Note that households who split (sub-divided) between 1996 and 2007 have been excluded from the dataset.
Note that the 1996 wave of the Bangladesh panel contained four rounds, over which an intra-annual village
level attrition rate can be calculated. This will not be feasible for many panel surveys.
4

CPRC Toolkit Note

#delimit ;
xi: probit A $headchar hhsize $demog ownland lvasset96 i.thana
lpcx96 perfloods9607 villattrate i.categ, robust cluster(village) ;
#delimit cr

Table 1: Attrition Probit for Consumption Expenditures

Probit regression

Number of obs =

Log pseudolikelihood = -162.38681

(Std. Err. adjusted for 47 clusters in village)

Pseudo R2

Variable (1996 values)

Age of household head
Age of household head squared

Coefficient

Robust
Std Err

954
0.1331

Pvalue

0.003
0.001

0.008
0.000

0.330
2.040

0.743
0.042

0.017
-0.074

0.020
0.066

0.830
-1.120

0.408
0.262

% of household members aged

0-4 years

0.013

0.006

2.090

0.037

5-14 years
15-19 years

0.010
0.003

0.005
0.006

2.030
0.490

0.042
0.621

35-54 years

0.015

0.007

2.230

0.026

55 and older
Total land owned (decimals)

0.012
0.000

0.007
0.001

1.860
-0.030

0.062
0.977

-0.108
0.842

0.100
0.218

-1.080
3.860

0.281
0.000

0.532
0.321

0.270
0.241

1.970
1.330

0.049
0.182

-0.128

0.219

-0.580

0.560

0.002

0.003

0.690

0.487

0.013

0.029

0.430

0.665

-0.077

0.175

-0.440

0.661

0.114
0.373

0.225
0.208

0.500
1.790

0.614
0.073

-1.977

1.340

-1.480

0.140

Education of household head

(years)
Household size

(Log) Value of Assets

Pakundia thana
Gaffargao thana
Jessore thana
(Log) Per Capita Expenditure
% of households in village
experiencing floods
Village Attrition Rate in 1996
Adoption Status in 1996
B (adoptor, comparison village)
C (likely adoptor, comparison)
D (non-adopter, comparison village)
Constant

The pseudo R-squared from the attrition probit in Table 1 suggests that baseline variables
and village attrition explain about 13% of panel attrition between 1996 and 2006/07. While
this is relatively high explanatory power for attrition probit, note that it still leaves some 87%
of attrition as unexplained. The z-statistics and P-values in the middle two columns of the
table show only six of the 22 variables in the attrition probit are statistically different from
zero at the 5% level, although two more are statistically different from zero at the 10% level .
Variables that are significant predictors of attrition including the age of the household head

CPRC Toolkit Note

squared, selected household demographic variables and one thana (sub-district) and one
adoption status dummy.
Using the Stata test command we perform a Wald test of whether these groups variables are
jointly equal to zero using the command:
#delimit ;
test $headchar $demog lpcx96 perfloods9607 villattrate
_Icateg_96_2 _Icateg_96_3 _Icateg_96_4 _Ithana_2 _Ithana_9 _Ithana_
;
#delimit cr
The resulting Chi-squared statistic of 85.00 with 17 degree of freedom indicates these
variables are jointly statistically different from zero at the highest level of significance (the Pvalue is 0.000), so we can conclude these are significant predictors of attrition. Notice that
the characteristics of the household head and demographic composition variables are
among the variables which are jointly able to predict attrition. None of the seven groups of
variables are, however, individually different from zero at conventional levels of significance.
The BGLW test for attrition is also implemented by creating variables with the interactions
between the attrition variable (A) and all other variables using Statas xi command.
#delimit ;
xi i.A*agehh i.A*agesqr i.A*educ_h i.A*p0_4 i.A*p5_14 i.A*p15_19 i.A*p35_54
i.A*p55p i.A*ownland i.A*lvasset96 i.A*i.thana
i.A*perfloods9607 i.A*i.categ, prefix(I) ;
#delimit cr

A (clustered) regression is then estimated, with the log of per capita expenditures in 1996 as
the dependent variable, and the household and auxiliary variables plus their interactions with
the Attrition variable (denoted by IAX*) as the explanatory variables:
#delimit ;
xi: reg lpcx96 hhsize agehh agesqr educ_h
p0_4 p5_14 p15_19 p35_54 p55p ownland lvasset96 i.thana
perfloods9607 villattrate i.categ A IAX*, robust cluster(village) ;
#delimit cr

Statas testparm command is then used to test for whether the attrition dummy and all the
interactions are jointly equal to zero:
#delimit ;
testparm A IAXagehh_1 IAXagesq_1
IAXp15_1_1 IAXp35_5_1 IAXp55p_1

IAXeduc__1 IAXp0_4_1 IAXp5_14_1

CPRC Toolkit Note

IAXownla_1 IAXlvass_1 IAXtha_1_2 IAXtha_1_9 IAXtha_1_47 IAXperfl_1

IAXcat_1_2 IAXcat_1_3 IAXcat_1_4 ;
#delimit cr

The F-statistic of 24.67 is able to reject the null hypothesis that attrition is random at the
highest levels of significance.
Given that both the standard tests indicate that attrition for the expenditure model is nonrandom, we proceed to calculate inverse probability weights for this model. To do this we
first calculate the predicted probabilities from the unrestricted attrition probit in Table 1, and
then re-estimate it excluding seven groups of auxiliary variables, which include the
characteristics of the household head, demographic composition and per capita income of
the household in the initial period, as well as floods, the thana and treatments dummies, and
the treatment dummies. After calculating the predicted probabilities from the restricted
attrition probit, the inverse probability weights are calculated straightforwardly by taking the
ratio of the restricted to unrestricted probabilities. These steps are accomplished using the
following Stata commands:
#delimit ;
xi: probit A $headchar hhsize $demog ownland lvasset96 i.thana
lpcx96 perfloods9607 villattrate i.categ, robust cluster(village) ;
#delimit cr
gen sample=e(sample)
predict pxav
xi: probit A hhsize ownland lvasset96 if sample==1, robust cluster(village)
predict pxres
gen attwght=pxres/pxav

(In the example do-file, attrition_weights.do, there are also some additional capture drop
commands to ensure that existing predicted values which may be in the memory are
deleted).
The inverse probability (or attrition) weights produced vary from .09 to 31.53, with a mean
value of 1.48. In general, the inverse probability weights give more weight to households that
have remained in the panel than an unweighted regression would, although in some cases
households whose characteristics make it very unlikely that they attrit are weighted
downwards.

CPRC Toolkit Note

Tables 3 and 4 below show the relatively minor impact that applying these inverse probability
weights to standard poverty transition matrices has. Without weighting, households moving
out of poverty account for around 50.7% of panel households, while with weighting these
households account for 50.5%. Similarly, the number of chronically poor households
(households that are poor in both periods) falls from 11.2% without weighting to 9.7% of
households with attrition weights.
Table 2: Poverty Transition Matrix Without Attrition Weights
Poor 2007
Poor
Non-Poor
Poor
11.23
50.66
Non-Poor
1.76
36.34
Total
13.00
87.00
Poor 1996

Total
61.89
38.11
100.00

Table 3: Poverty Transition Matrix With Attrition Weights

Poor 1996

Poor 2007
Non-Poor
9.71
49.72
2.00
38.56
11.72
88.28

Total

Poor
Poor
Non-Poor
Total

59.43
40.57
100.00

Note that these transition matrices are calculated using the Bangladesh Bureau of Statistics
upper poverty line for 2005 adjusted to 1996 and 2007 terms by the food and non-food
consumer price indices.
Expenditure regressions (not shown, but see the code included in attrition_weights.do) show
that whether inverse probability weights are applied makes a fairly small difference to semilog expenditure regressions.
When the value of household assets is the outcome variable of interest, we have a slightly
larger sample of 963 households of whom 57 attrit between 1996 and 2007. Table 4 shows
the an attrition probit for asset attritors, in which we find limited evidence of non-random
attrition, with just three of our independent variables being significantly different from zero at
the 5% level. The only one of auxiliary attrition variables which is significantly different from
zero at this level is the (natural) logarithm of the value of assets in 1996. A Wald test of the
joint significance of the variables related to the household head, the demographic
composition of the household, the auxiliary variables and the thana dummies has a Chisquared value of, 96.92 and a P-value of 0.00, so the null hypothesis of random attrition can
be easily rejected. The BGLW pooling test has a F value of 7.35 and a P-value of 0.000,
confirming that asset attrition is not random.

CPRC Toolkit Note

Table 4: Attrition Probit for Value of Household Assets

Probit regression
Log pseudolikelihood = -188.15118
(Std. Err. adjusted for 47 clusters in village)

Number of obs =
Pseudo R2
=

Variable (1996 values)

Coefficient
Age of household head
-0.007
Age of household head squared
0.001
Education of household head (years)
0.017
Household size
-0.009
% of household members aged
0-4 years
0.009
5-14 years
0.005
15-19 years
-0.001
35-54 years
0.011
55 and older
0.014
Total land owned (decimals)
0.000
Attrition Variables
(Log) Value of Assets in 1996
-0.272
% of households in village experiencing
floods
0.000
Village Attrition Rate in 1996
0.036
B (adoptor, comparison village)
-0.213
C (likely adoptor, comparison)
0.103
D (non-adopter, comparison village)
0.302
Pakundia thana
0.646
Gaffargao thana
0.237
Jessore thana
-0.089
Constant
-1.741

Robust
Std Err
0.006
0.000
0.018
0.063

963
0.1306

z
P-value
-1.060
0.290
1.950
0.051
0.930
0.351
-0.140
0.886

0.007
0.005
0.007
0.007
0.006
0.001

1.240
0.970
-0.140
1.670
2.590
0.770

0.215
0.330
0.889
0.094
0.010
0.442

0.106

-2.570

0.010

0.003
0.030
0.187
0.152
0.198
0.237
0.275
0.237
0.392

0.000
1.180
-1.140
0.680
1.530
2.730
0.860
-0.380
-4.440

0.999
0.237
0.256
0.499
0.127
0.006
0.389
0.707
0.000

Accordingly the groups of variables which predict attrition were dropped and the restricted
probit was estimated. The ratios of the predicted values of the restricted to the unrestricted
model were calculated, producing inverse probability weights which range from 0.15 to
12.81.
Table 5 report linear panel regressions for the (natural) logarithm of household assets with
and without inverse probabilities weights. The regressors are the same basic variables as in
attrition probits but exclude the auxiliary variables. These two regressions are produced by
the following Stata commands:

CPRC Toolkit Note

* without inverse probability weights

#delimit ;
xi: reg lvasset07 $headchar1 hhsize $demog2 $land i.thana, robust
cluster(thana) ;
*with inverse probability weights
xi: reg lvasset07 $headchar1 hhsize $demog2 $land i.thana
robust cluster(thana);
#delimit cr

[pw=attwght2],

Table 5: Linear Regressions for Log of Household Assets, 2006-07

Variable
(1996 values)
Age of household head
Age of household head squared
Education of household head (years)
Household size
% of household members aged
0-4 years
5-14 years
15-19 years
35-54 years
55 and older
Total land owned (decimals)
Pakundia thana
Gaffargao thana
Jessore thana
Constant

Number of obs = 906

Number of obs = 906
R-squared = 0.273
R-squared = 0.254
Root MSE
= 1.1044
Root MSE
= 1.0387
(Std. Err. adjusted for 47 clusters in village)
Without Attrition Weights
With Attrition Weights
Robust
Robust
Coefficient Std Err
P-value
Coefficient Std Err
P-value
0.011
0.005
0.387
0.005
0.006
0.033
0.000
0.000
0.168
-0.001
0.000
0.622
0.111
0.011
0.000
0.098
0.013
0.000
0.005
0.021
1.000
0.000
0.021
0.817
-0.014
-0.003
-0.006
-0.011
-0.015
0.001
-0.681
-0.726
-0.541
9.611

0.004
0.003
0.004
0.004
0.005
0.000
0.156
0.177
0.145
0.213

0.008
0.229
0.179
0.021
0.123
0.000
0.000
0.000
0.001
0.000

-0.014
-0.004
-0.007
-0.009
-0.007
0.001
-0.570
-0.801
-0.503
9.967

0.005
0.003
0.005
0.004
0.005
0.000
0.120
0.180
0.146
0.265

0.001
0.346
0.144
0.006
0.007
0.000
0.000
0.000
0.001
0.000

Inspection of the left and right hand sides of Table 5 reveals that the coefficients and
significance of individual coefficients are fairly similar. However, using the usual 5% level of
significance, there are two variables (the age of the household head, and the percentage of
household members over 55 years of age) which is not significantly different from zero in the
regression without attrition weights but significant when weights are applied. A Hausman
test of the equality of the coefficients of the weighted and unweighted models is firmly
rejected (Chi-squared(13) value of 197.33, P-value of 0.000).

CPRC Toolkit Note

Conclusions and caveats

Inverse probability weights are one of two methods in common use for correcting for attrition
bias (the other being the estimation of a Heckman type selectivity model).
The great advantage of inverse probability weights is their simplicity but a number of caveats
must be borne in mind about their use.
First, the tests and adjustment for attrition presented above assume that attrition is based on
observables only. When attrition also depends on unobservable factors, as will often be the
case, other methods (such as a selectivity model) need to be used.
Second, the attrition tests and adjustments described above are model specific, and must
therefore be repeated for each outcome variable of interest.
Third, only one type of attrition considered here (permanent unit non-response) has been
considered in this note. With multiple wave panels some members (in particular when panel
units are individuals) may be missing from one wave of a panel only to reappear at a later
date. While inverse probability weights can also be used to adjust for temporary attrition,
and also item non-response (when particular questions are not answered), some
modifications to their derivation are necessary.
Fourth, while it is possible to correct for attrition bias, it is always wise to try and minimise
attrition at the data collection stage. Some useful strategies for reducing attrition in panel
data are described in Hill (2001).
Finally, it should be remembered that many significant factors of the poverty experiences of
individuals and households are suppressed by the construction of panels, although they
are informative in their own right. Qualitative and participatory studies, for example, suggest
that extreme poverty often leads to the migration of household members, the dissolution of
households, and in the most extreme and heart-rending cases, the death of unsupported
individuals.

CPRC Toolkit Note

References
Alderman, H., Behrman, J., Kohler, H.P., Maluccio, J. and Cotts Watkins, S. (2001) Attrition in
longitudinal household survey data. Demographic Research 5(4): 79-124.
Becketti, S., Gould, W. Lillard, L., and Welch, F. (1988). The Panel Study of Income Dynamics after
Fourteen Years: An Evaluation. Journal of Labor Economics 6: 472-92.
Outes-Leon, I. and Dercon, S. (2008). Survey Attrition and Bias in Young Lives. Young Lives
Technical Note 5. Oxford: University of Oxford.
Falaris, E. (2003). The effect of survey attrition in longitudinal surveys: evidence from Peru, Cote
DIvoire and Vietnam. Journal of Development Economics, 70: 133-158.
Fitzgerald, J., Gottschalk and Moffit, R. (1998). An analysis of sample attrition in panel data. Journal
of Human Resources, 33(2): 251-299.
Heckman, J, (1979) Sample selection basis as a specification error. Econometrica, 47: 153-161.
Hill, Z. (2001). Reducing attrition in panel studies in developing countries. Young Lives Working
Paper 5.
Mallucio, J. (2004). Using quality of interview information to assess non-random attrition bias in
developing country panel data. Review of Development Economics, 8(1): 91-109.
Winkels, J. and Withers S. (2000). Panel attrition in Rose, D. (ed) Researching Social and Economic
Change: The Uses of Household Panel Studies. London: Routledge.
Wooldridge J. (2002). Econometric Analysis of Cross-sectional and Panel Data, Cambridge MA: MIT
Press.