We use the ‘oaxaca’ command in Stat

We use the ‘oaxaca’ command in Stata, developed by Jann (2008), with the pooled option, which utilises the coefficients of a joint regression model of both years (including a year dummy) as reference coefficients. While decomposition analysis was originally developed for use in conjunction with linear regression, a non-linear regression model is required in our case given the outcome variable is categorical. We therefore estimate a series of binary logistic regression models where the outcome is the likelihood of working in one of the employment types relative to working in any of the other employment types,5 and apply the decomposition methodology for non-linear models proposed by Yun (2004). Given the non-linear trend in the prevalence of different employment types over the 17-year period, we present three decomposition models for each employment type: The first model analyses changes over the entire period by comparing 2001 with 2017, and the second and third models look at changes within sub-periods by comparing 2001 with 2008 and 2008 with 2017, respectively. The analysis is weighted using crosssection responding person weights, and standard errors are clustered on the individual. The models include all worker and job characteristics presented in Table 2, but to avoid potential collinearity, neither firm size nor public sector are included in the model for self-employed workers. We apply the deviation contrast transform to the categorical predictor variables in our models to ensure the results are not dependent on the choice of reference category (Jann, 2008). When presenting results, the contribution of sets of variables is summarised in a single coefficient. However, when discussing our results, we also make use of the estimated coefficients on the individual variables.