Attrition Bias in Panel Data: A Sheep in Wolf's Clothing? A Case Study Based on the Mabel Survey

Abstract

This paper investigates the nature and consequences of sample attrition in a unique longitudinal survey of medical doctors. We describe the patterns of non-response and examine if attrition affects the econometric analysis of medical labour market outcomes using the estimation of physician earnings equations as a case study. We compare the econometric estimates obtained from a number of different modeling strategies: balanced versus unbalanced samples; an attrition model for panel data based on the classic sample selection model; and a recently developed copula-based selection model. Descriptive evidence show that doctors who work longer hours, have lower years of experience, are overseas trained, and have changed their work location are more likely to drop out. Our analysis suggests that the impact of attrition on inference about earnings of General Practitioners is small. For specialists, there appears to be some evidence for an economically significant bias. Finally we discuss how the top-up samples in the MABEL survey can be used to address the problem of panel attrition.

Drug therapy adherence and health outcomes in the presence of physician and patient unobserved heterogeneity

“…Nevertheless, a comparison of the estimates with those reported in column 3 of Table 4 (i.e., without attrition adjustment) reveals that these estimates are very similar in magnitude. As in Cheng and Trivedi (2015), this evidence suggests that despite the missing-at-random assumption is strongly rejected, withdrawal attrition does not have a significant impact on the estimates of the TC/HDL ratio equation for the selected incident patients, corroborating the results described above (Table AI). …”

“…As in Cheng and Trivedi (2015), instead of estimating t by a sequence of probits where, in each quarter, the estimation sample uses the patients still in the sample in quarter t − 1, we estimate a pooled model where the sequential response function of each quarter t is pooled across quarters 2002q4, … , 2009q4 to maximise statistical power using covariates observed at quarter t − 1. The set of covariates included in w ijt−1 are those included in our main specification plus a dummy for females and a set of macro-area of residence dummies (using South and Islands as base category).…”

Periodontal Disease Is Associated with Insomnia among Victims of the Great East Japan Earthquake: A Panel Study Initiated Three Months after the Disaster

“…Indeed, as indicated by the largest number of participants in the first survey, panel studies typically suffer from attrition, which pretend the higher response rate in the initial survey as biased inferences (Cheng and Trivedi 2015). Therefore, we cannot rule out the possibility of a selection bias.…”

Punto Seguro: A Randomized Controlled Pilot Using Conditional Economic Incentives to Reduce Sexually Transmitted Infection Risks in Mexico

“…Second, and more generally, the experimental comparison between the different type of conditionality and the level of incentives contributes to the debate in the health and development fields [41,43,45]. Third, in terms of methods, the issue of attrition is common in both experimental and non-experimental methods, yet rigorous techniques to address it, such as inverse probability weighting, have not yet been fully utilized in the context of economic incentives to reduce HIV/STI risks, with only a few applications in health economics [46,47]. …”