When attrition is predictable: machine learning and the estimation of corporate leverage dynamics

Firm-level data sets suffer from panel attrition, and it can also be a problem with dynamic panel estimates. In this paper, we explore the question of whether the leverage estimates of a firm are meaningfully impacted by a predictable exit event, after adjusting for firm and time characteristics. We use a cross-fitted machine-learning model to estimate survival probabilities for an unbalanced panel of firms from 1999 to 2024, using lagged financial characteristics. Predictions show almost perfect separation between the groups of firms that survive and the group of firms that exit, suggesting little common support. We build stabilized inverse probability weights and re-estimate dynamic leverage models with firm and year fixed-effects. The weighted estimates suggest there is less leverage persistence and more speed of adjustment to the target leverage than conventional fixed-effects estimates, and the core qualitative relationships are largely unchanged. Therefore, the direction of estimated leverage dynamics is not really affected by predictable attrition, as it is the magnitude that is affected. The weighted estimates are only interpreted as sensitivity evidence due to limited overlap; there are no globally identified causal effects.