Verifying the existence of maximum likelihood estimates for generalized linear models

TL;DR

This paper investigates the conditions for the existence of maximum likelihood estimates in generalized linear models, especially with high-dimensional parameters, and demonstrates their importance through empirical examples.

Contribution

It clarifies the conditions under which GLM estimates exist and provides methods to verify these conditions in complex models with many fixed effects.

Findings

Some GLM estimators remain consistent even when conditions for existence fail.

Verifying existence conditions is crucial to avoid misleading estimates in high-dimensional models.

Applying these methods to a gravity model reveals potential nonexistence issues affecting results.

Abstract

A fundamental problem with nonlinear models is that maximum likelihood estimates are not guaranteed to exist. Though nonexistence is a well known problem in the binary response model literature, it presents significant challenges for other models and is not as well understood in more general settings. These challenges are only magnified for models that feature many fixed effects and other high-dimensional parameters. We address the current ambiguity surrounding this topic by studying the conditions that govern the existence of estimates for (pseudo-)maximum likelihood estimators used to estimate a wide class of generalized linear models (GLMs). We show that some, but not all, of these GLM estimators can still deliver consistent estimates of at least some of the linear parameters when these conditions fail to hold. We also demonstrate how to verify these conditions in models with high-dimensional parameters, such as panel data models with multiple levels of fixed effects. Applying our methods to a gravity model with heterogeneous free trade agreement effects, we show that failing to detect nonexistence can produce misleading numerical estimates.