📌 The Problem With Current Binary Panel Models
Logit and probit (L/P) models are the standard tools for binary time-series cross-sectional (BTSCS) data, and researchers typically add cubic splines or time polynomials to capture temporal dynamics. However, L/P models struggle with three other temporal features that commonly arise in BTSCS data:
- whether covariate effects vary depending on time,
- whether the underlying process is causally complex, and
- whether the assumed functional form for time is correct.
Failing to account for any of these features produces misspecification bias and threatens the validity of inferences.
📊 How the Models Were Compared (Monte Carlo Evidence)
- Monte Carlo simulations compare Cox duration models to standard logit models across a range of BTSCS settings.
- Simulations include both a basic BTSCS scenario and more complex situations that introduce time-conditional effects and causal complexity.
- Assessment focuses on the ability to test the same hypotheses, estimator performance, and susceptibility to misspecification bias.
🔍 Key Findings
- Cox duration models create fewer opportunities for the kinds of misspecification bias that afflict L/P models with temporal complexity.
- In basic BTSCS settings, Cox models perform about as well as logit models and sometimes better.
- In more complex BTSCS situations, Cox models perform considerably better than logit models.
📈 A New Way to Read Cox Results
- Transition probabilities are introduced as an interpretation technique for Cox models to make coefficients and effects more readily interpretable for practitioners.
🧭 Applied Example
- An application to interstate conflict illustrates the practical differences in inference and interpretation between Cox and L/P approaches.
⚠️ Why It Matters
- For BTSCS research, Cox duration models offer a viable alternative that reduces misspecification risk while allowing researchers to address the same substantive hypotheses as logit/probit models.
- Using transition probabilities improves the accessibility of Cox-model results for applied audiences.
