Generating and validating synthetic unbalanced panel data: evidence from Italian firm microdata

This study develops and evaluates a framework for generating synthetic microdata for the Survey of Industrial and Service Firms (INVIND), a large dataset of Italian firms observed between 1993 and 2024. We extend synthetic-data methodologies, traditionally designed for cross-sectional settings, to panel data and compare a sequential modelling approach based on regression and classification trees with a joint modelling strategy employing Gaussian copulas. The procedure differentiates between short and long firm histories to accommodate heterogeneous time spans. We synthesize some key variables, including geographic area, sector, employment, and turnover, and assess the resulting dataset along multiple dimensions: reproduction of univariate and conditional distributions, stability of structural relationships through fixed-effects regressions, and disclosure risk. While both modelling approaches ensure strong fidelity to the original data distributions, re-identification risk is also sizeable. To address this trade-off, we introduce controlled measurement errors, which significantly reduce risk while preserving essential statistical properties.