Abstract

Nonresponse and attrition are among the most important issues in panel surveys. Besides adversely affecting the usability and analytic value of the data, both nonresponse and attrition result in estimation efficiency losses and increased risks of bias.
Although nonresponse and attrition are sharply distinct phenomena conceptually, separating them rigorously can sometimes prove impossible in practice. To analyze in depth and disentangle these two phenomena, we explored the application of survival analysis to panel data.
The literature in this field typically treats as “panel attrition” the earliest nonresponse event recorded along the panel waves, disregarding the fact that non-respondents often resume participation in later waves, thus actually mixing attrition with interim nonresponse.
To overcome this limit and better capture the phenomenon of permanent withdrawal of units from the panel, we propose an operational definition of attrition that strives to characterize attriters as units with a persistent non-respondent state over time.
We then leverage existing longitudinal data to model our definition of attrition by means of survival analysis techniques. The first goal of our work is, indeed, to devise a well performing survival model of attrition and fit it to available panel data. To this end, we analyzed the first 11 rounds of the panel study “High Frequency Phone Surveys of Households” that was conducted in Ethiopia in 2020 and 2021 to monitor the impact of the Covid-19 pandemic.
The ultimate aim of our research is to exploit the fitted model, after out-of-sample validation, to predict attrition hazards in real time for ongoing panels, so as to improve and better target survey management decisions during data collection. For instance, costly survey efforts, like additional follow-up attempts and incentives, could be devoted to units whose attrition hazards predicted by the model turn out to be substantial but not too high. On the contrary, units whose propensity towards panel participation is predicted by the model to be “compromised” could form an ideal target for substitutions, if allowed by the survey protocol, or for capping/stopping further contact attempts.