Reweighting in panel surveys: machine learning techniques for the Health Care and Social Survey

BACKGROUND: Healthcare statistical services have used probability surveys to respond to such information needs. The Health Care and Social Survey (ESSOC) research project arises from the need to provide data on the evolution of the COVID-19 impact. This survey has an overlapping panel survey design with 4 measurements throughout 1 year with random samples. OBJECTIVE: To develop a new reweighting method for overlapping panel surveys affected by non-response. METHODS: Each ESSOC measurement is composed of two samples: a longitudinal sample from previous measurements and a new sample at each measurement. In each measurement, missing units are substituted by new surveyed units, allowing the obtention of cross-sectional and longitudinal estimates. The advantage of this design is that, in addition to being able to obtain longitudinal estimates, cross-sectional estimates are more accurate because of the larger sample size. However, the problem of non-response is particularly aggravated in the case of panel surveys, due to the fatigue of the population to be repeatedly surveyed. RESULTS: Considering the design, timing and objectives of this survey, our reweighting method produces suitable estimators for both cross-sectional and longitudinal samples. The weights are the result of a two-step process: the original sampling design weights are corrected during a 1st phase by modelling the non-response with respect to the longitudinal sample obtained in a previous measurement using machine learning techniques. Then, during a 2nd phase, they are calibrated using the auxiliary information available at the population level. The proposed method is applied to the estimation of totals, proportions, ratios, and differences between measurements as well as gender gaps. CONCLUSIONS: For addressing future health crises like COVID-19, it is therefore necessary to reduce potential coverage and non-response biases in surveys by means of utilising reweighting techniques as proposed in this study. KEY MESSAGES: • It is therefore necessary to reduce potential coverage and non-response biases in surveys. • This study proposes the utilisation of reweighting techniques based on calibration and XGBoost.