580 Panel Data Methods
utilization of services for the Ontario Health Insurance Plan (OHIP) for the survey
year and the five previous years. Linear probability models are used to analyze the
probabilities of false negatives and false positives in the self-reported data. This
shows that reporting errors are associated with individual characteristics. The data
reveal a large number of false negatives, although the probability declines for those
with more recorded medical treatments, suggesting that this reflects undiagnosed
conditions and lack of information among respondents. The number of false posi-
tives is much smaller, but there is some evidence of “justification bias”: those not in
work are more likely to report false positives for conditions, such as hypertension,
ulcers and bronchitis.
A more favorable view of self-reported measures emerges in the work of Benitez-
Silvaet al.(2004). They take the relatively small sub-sample of respondents to the
first three waves of the US Health and Retirement Survey (HRS) who had applied for
disability benefit from the Social Security Administration (SSA) and compare their
self-reported disability to the outcome of the SSA decision. In this case the SSA
decision to award benefits is used as an objective indicator to assess the reliability
of the self-reported data on limitations that prevent work. Conditional moment
tests for whether self-reported disability is an unbiased indicator of the SSA decision
suggest that a large fraction of this population report their health accurately. Unlike
Bakeret al.(2004), this study relies on information collected within the HRS rather
than matching the survey data with administrative records. McGarry (2004) adopts
another strategy to get around the problem of “justification bias.” Rather than using
data on actual retirement, she uses information from the HRS on the subjective
expected probability of retirement by age 62, which is collected while people are
still in work. Using this measure she finds strong effects of health on expected
retirement age.
It is sometimes argued that the mapping of health into SAH categories may vary
with respondent characteristics. This source of measurement error has been termed
“state-dependent reporting bias” (Kerkhofs and Lindeboom, 1995), “scale of ref-
erence bias” (Groot, 2000) and “response category cut-point shift” (Murrayet al.
2001; Sadanaet al.2000). Regression analysis of SAH is often done by specify-
ing an ordered probability model, such as the ordered probit or logit. Then the
symptoms of measurement error can be captured by making the cut-points depen-
dent on some or all of the exogenous variables used in the model and estimating
a generalized ordered model. This requires stronga priorirestrictions on which
variables affect health and which affect reporting in order to separately identify
the influence of variables on latent health and on measurement error. Attempts to
surmount this fundamental identification problem include modeling the reporting
bias based on more “objective” indicators of true health (Kerkhofs and Lindeboom,
1995; Lindeboom and Van Doorslaer, 2004) and the use of “vignettes” to fix the
scale (Das and Hammer, 2005; Murrayet al.,2001). Lindeboom and Van Doorslaer
(2004) analyze SAH in the Canadian National Population Health Survey and use
the McMaster Health Utility Index (HUI-3) as their objective measure of health.
They find evidence of reporting bias with respect to age and gender, but not for
income, education or linguistic group.