A Bayesian approach to labour market modelling in dynamic microsimulation

The paper presents a novel approach to modelling labour market processes in dynamic microsimulation. The method combines and integrates Bayesian simulation based estimation and simulation of the dependent variables. The approach is applied to a dynamic panel model for hourly wage rates for Danish employees using a large panel data set with 17 years of data for 1995 to 2011. The wage rate model and a parallel model for annual work hours are currently being implemented in SMILE (Simulation Model for Individual Lifecycle Evaluation), a new dynamic microsimulation model for the Danish household sector. The application benefits from the richness of Danish administrative panel data. Nevertheless, the results and the approach have several features that should be of interest to micro-simulators and others. Indeed, the model features both an extraordinarily comprehensive list of dependencies and a rich dynamic structure. Together, these features contribute to ensure that simulations produce realistic cross-sectional distributions and interactions as well as inter-temporal mobility – the key determinants of the quality of a dynamic microsimulation model. In addition to the 'usual' socio-demographic variables (gender, age, ethnicity, experience, education etc.), the dependencies include a more novel set of variables that represent a person’s labour market history, secondary school grade and social heritage (represented by the parent’s education level). The dynamic model structure includes a lagged dependent variable, an auto-correlated error term with a mixed Gaussian distribution for the white noise component, an individual random effect with a mixed Gaussian distribution and permanent effect of a person’s first wage after leaving the education system. The estimation sample is identical to the simulation sample, which allows us to use the same historical detail as well as estimated individual effects – i.e. random effect components – for the simulation of future wage rates. The Bayesian estimation method handles missing observations for the dependent variable – due to either non-employment or temporary non-participation – by treating missing observations as latent variables that are simulated alongside the Bayesian iterations. As a byproduct, the estimations produce model consistent latent wage rates for the unemployed that are useful for labour supply analysis.