Who overuses sickness insurance? Evidence from a randomised experiment

Author: Adrian Adermon, And Yaroslav Yakymovych, And
Dnr: 213/2020

The aim of this project is to use causal forests, a machine learning approach, to study what characterises individuals who are prone to stay longer on sick leave if not monitored by medical professionals. This is important for a better understanding of the issue of moral hazard in the social insurance system.

The researchers analyse an experiment which took place in 1988 in Gothenburg city and Jämtland county. Individuals born on odd dates needed to provide medical certificates after they had been absent from work for seven days, as normal, while individuals born on even dates could be on leave for 14 days without needing to provide a certificate. The change in medical certificate requirements is very similar to that enacted during the Covid-19 pandemic.

Differences between workers are analysed based on a large number of possible determinants of sickness absence, including previous sickness absence history, family situation, income and education level, sector, industry as well as colleagues’ and neighbours’ sickness absence behaviour.

The causal forest algorithm has many advantages relative to traditional methods for analysing the effects of experiments across different subgroups. It cycles in turn through all variables of interest and all possible splits according to these, choosing the split which maximises the differences in estimated treatment effects.