Onsets of binary events—such as regime change, civil war onset, or treaty signing—are central to many political-science questions. A common empirical shortcut is to set ongoing post-onset years to zero to create an onset indicator. That transformation is intuitive but introduces serious inferential problems.
🔎 The Core Problem
- Setting ongoing years to zero creates two qualitatively different reasons for a zero outcome (pre-onset absence versus post-onset continuation) that estimators cannot distinguish.
- The transformation also masks the possibility that covariates affect the timing of onsets differently from how they affect the duration of events.
🧪 How This Was Tested
- Analytical derivation demonstrating the bias that arises from the transformation.
- Monte Carlo experiments quantifying how the bias varies across simulated data-generating processes.
- A sensitivity analysis of determinants of civil war onset that applies the transformation and compares substantive inferences.
📌 Key Findings
- The transformation analytically induces bias in estimated coefficients because zeros acquire mixed meanings.
- Monte Carlo results show this bias can be substantial, with magnitude and direction depending on the data-generating process and covariate dynamics.
- The civil war sensitivity analysis reveals considerable differences in coefficient sizes and in whether particular covariates are deemed robust determinants of onset when the problematic transformation is used.
⚖️ Recommendations and Why It Matters
- Avoid conflating pre-onset non-occurrence with post-onset continuation by relying on the zeroing shortcut.
- Prefer empirical approaches that separate onset and duration processes or otherwise model duration dependence explicitly so that covariate effects on timing and persistence can differ.
- Conduct sensitivity checks to assess how inference about onset determinants changes with outcome construction.
The issues identified affect a wide range of binary-event studies in comparative politics and international relations; careful outcome construction and modeling choices are necessary to obtain reliable substantive inferences.
