Event History Turnover Analysis

A logistic regression on a stayed-or-left indicator ignores when people left and mishandles censoring. It treats someone who quit in the second month the same as someone who quit in the eleventh, and it has no principled way to deal with employees hired late or still present when the study ends. Event history analysis models the hazard — the instantaneous risk of leaving given survival so far — using both the duration and a censoring indicator, so it uses the timing information and lets censored cases contribute exactly the information they carry. As Morita, Lee, and Mowday showed, this recovers effects on the timing of turnover that binary methods miss.

An observation is right-censored when you know the employee survived up to a certain point but did not observe their turnover — because the study ended, they were hired late, or they exited for a reason outside your event of interest. Allison emphasizes that the defining strength of event history analysis is handling this correctly: the likelihood credits each censored case with having survived to its censoring time rather than dropping it or coding it as a non-leaver. Mishandling censoring is one of the most common and damaging errors in turnover research, since it discards real information and biases the estimated effects of predictors and of tenure itself.

In a proportional hazards model the hazard for an employee is a baseline hazard multiplied by the exponential of the covariates, so each coefficient exponentiated gives a hazard ratio — the multiplicative effect of a one-unit change in a predictor on the instantaneous risk of leaving. A hazard ratio of 1.5 means a 50 percent higher risk of quitting at any given moment, implying faster turnover; a ratio below 1 means lower risk and longer retention. The interpretation assumes proportional hazards, meaning the ratio is constant over tenure, which should be tested; when it fails, time-varying coefficients or stratified models are used instead.