Files
Abstract
Panel (or longitudinal) data often provide an understanding of the dynamic
behavior of individual households not possible with cross-sectional or time-series
information alone. However, a disturbing feature of this type of survey in both developed
and developing countries is that there is often substantial, nonrandom attrition. Therefore,
an important concern is the extent to which attrition inhibits inferences made using the
data. This note examines attrition in the KwaZulu-Natal Income Dynamics Study (1993–
1998) and assesses the extent of attrition bias for a specific empirical example. The
analysis shows that 1993 first round nonresponse is largely unrelated to observable
characteristics of the communities other than indicators of migration activity.
Multivariate regressions are then used to describe the characteristics of the households
attriting in 1998, revealing the importance of distinguishing between two types of
attriting households, those that moved and those that apparently moved but left no trace.
For example, increased household size reduced the probability of either type of attrition,
whereas measures of higher quality of fieldwork in the 1993 survey only reduced the
probability that a household left no trace. While observable differences between attritors
and non-attritors indicate attrition is nonrandom, it does not necessarily follow that
estimated relationships based on the non-attriting sample suffer from attrition bias.
To more directly explore attrition bias, which is by its nature model specific, this
analysis estimates household-level expenditure functions correcting for attrition bias
using standard Heckman selection procedures and a quality of 1993 interview variables as identifying instruments. There is positive selection, and although many of the other
parameter estimates are quite similar, a Hausman test rejects the equality of coefficients
between the corrected and uncorrected models. Therefore, this study concludes, at least
for this simple case, that attrition does appear to bias the “behavioral” coefficients. These
results are in contrast to other work using these data that suggests little attrition bias for
different estimated models, highlighting that attrition is indeed model specific. Large
levels of attrition do not always lead to attrition bias; however, sometimes they do. Since
it is typically difficult to determine the bias for a particular analysis a priori, it behooves
researchers using panel data not to avoid using panel data when there is attrition, but to
always evaluate the effect of such bias on the analysis at hand.