Scholars in international relations (IR) are increasingly using time-series cross-section data to analyze models with a binary dependent variable (BTSCS models). IR scholars generally employ a simple logit/probit to analyze such data. This procedure is inappropriate if the data exhibit temporal or spatial dependence. First, we discuss two estimation methods for modelling temporal dependence in BTSCS data: one promising based on exact modelling of the underlying temporal process which determines the latent, continuous, dependent variable; The second, and easier to implement, depends on the formal equivalence of BTSCS and discrete duration data. Because the logit estimates a discrete hazard in a duration context, this method adds a smoothed time term to the logit estimation. Second, we discuss spatial or cross–sectional issues, including robust standard errors and the modelling of effects. While it is not possible to use fixed effects in binary dependent variable panel models, such a strategy is feasible for IR BTSCS models. While not providing a model of spatial dependence, Huber's robust standard errors may well provide more accurate indications of parameter variability if the unit observations are intra-related. We apply these recommended techniques to reanalyses of the relationship between (1) democracy, interdependence and peace (Oneal, Oneal, Maoz and Russett); and (2) security and the termination of interstate rivalry (Bennett). The techniques appear to perform well statistically. Substantively, while democratic dyads do appear to be more peaceful, trade relations, as measured by Oneal, et al., do not decrease the likelihood of particpation in militarized disputes, Bennett's principal finding regarding security and rivalry termination is confirmed; his finding on common external threats, however, is not; his results on the influence of issue salience are even more robust.
Working Paper 97–08, Weatherhead Center for International Affairs, Harvard University, November 1997.