Political Science 867

Event History

Professor Janet M. Box-Steffensmeier                              Phone: 292-9642 (office) & 326-2533 (home).

Ohio State University                                                         E-mail: steffensmeier.2@osu.edu

Winter 2005                                                                       

Course Assignments and Due Dates (PDF)

Lab Exercise I: The purpose of this assignment is to reinforce the rationale behind duration modeling and to begin looking at model interpretation. One of the reasonably regular empirical occurrences in International Relations is that democracies do not fight each other. Scholars have argued that in addition to democracy, there are a number of other factors that compound the likelihood that conflict will not exist within a pair of states. Using a modified version of the Russett and Oneal (1997) International Studies Quarterly data set (download here) in which the missing data have been eliminated, (see Beck et al. 1998), evaluate the argument that democracy limits conflict. This is a time-varying data set on 827 “politically relevant” dyads in the international system. Each dyad has one observation per year from 1950-1985. Omitting observations with ongoing conflicts results in an N=20,448. The variables, in order, are:

            DYADID                    The dyad identification number

            YEAR                         The year indicator

            DISPUTE                   1 if a militarized interstate dispute occurred between the members of that dyad in that year, 0 otherwise

            START                       The “starting” counter variable

            FOLLOWUP              A “follow-up time” variable

            DURATION               The duration variable

            DEMOC                     Rescaled POLITY democracy variable

            GROWTH                  Lagged measure of growth, as a proportion of GDP

            ALLIES                      1 if the dyad members are allied, 0 otherwise

            CONTIG                    1 if the members of the dyad are geographically contiguous, 0 otherwise

            CAPRATIO                The natural log of the ratio of the two states’ military capacities, as measured by the Correlates of War (COW) data

            TRADE                      The ratio of bilateral trade to GDP, in constant US dollars

1. Replicate the traditional analysis of this question (hint: run a logit).

2. Graph and describe the underlying hazard rate.

3. Estimate a Cox proportional hazards model using the six covariates provided. Interpret your results and compare them across the two sets of analyses. Explain why the event history approach is superior? How can you demonstrate its superiority from the above analysis?

DUE: Monday, February 7

Lab Exercise II: This assignment focuses on estimation and interpretation of parametric and semiparametric duration models. You will be using “famous” data on 314 cabinet durations in 15 countries (download here). The data have been used by (among others) Alt and King 1994; Strom 1985; King et al 1990; and Warwick 1992, 1994. Your mission is to replicate the analyses of cabinet durations presented in Alt and King (Comparative Political Studies 1994), Table 1.

            DURATION                           Cabinet duration, in months.

            CENSOR                                An indicator variable (1 if the cabinet “failed”, 0 if it did not, or did not do so within the 12-month “constitutional interelection period” (CIEP).

            INVESTITURE                      Whether (=1) or not (=0) there was an investiture requirement.

            POLARIZATION                  A measure of political polarization in the cabinet

            FRACTIONALIZATION      A measure of fractionalization.

            NUMERICAL STATUS        Whether (=1) or not (=0) the cabinet was a majority.

            FORMATION ATTEMPTS  The number of formation attempts.

            POSTELECTION                  Whether (=1) or not (=0) the cabinet was formed following an election.

            CARETAKER GOVT            Whether (=1) or not (=0) the government was a caretaker.

1.   Replicate the parametric analyses in Table 1 of King & Alt (1994), in particular the exponential and Weibull distributions. Estimate other distributions as you desire, such as the log-logistic. Discuss which model(s) provide the best “fit” and how you reached that conclusion. What else could be done to determine which parametric model to use? Present your results for the coefficient estimates and discuss them in substantive terms. Be sure to address any differences across models and speculate (briefly) on the reasons for those differences.

2.   Generate and discuss plots of the estimated hazards for each of the models you estimate, holding the independent variables constant. What do these plots suggest (or assume) about the shape of the hazard?

3.   Replicate the Cox results from Table 1 as well. Briefly interpret those results, in terms of odds ratios and the like. In addition, plot the estimated baseline hazards, and discuss its shape in both substantive and statistical terms.

4. Discuss the relative strengths and weaknesses of the parametric versus semiparametric approaches. What else could be done to improve the analysis?

Due: Monday, February 28

Lab Exercise III: The purpose of this assignment is to think about diagnostics and model estimation. You are encouraged to assemble and analyze your own duration data set. Alternatively, you can use the data set from Exercise I on International Conflict.

1.   Fit a Cox proportional hazards model with the Efron method for ties and robust standard errors. This is your base model.

2.   Assess the proportional hazards assumption for the Cox model, with the emphasis on statistical methods. In practice, about 1 out of every 7 or 8 variables violate the proportional hazards assumption. Based on your conclusions about whether the effect of the various covariates are proportional over time, reestimate the base model with whatever time interactions (if any) you think are appropriate to address the nonproportionality. Interpret any nonproportionality you find (i.e., what do the results mean?).

3. Assess the prevalence of tied survival times in the data (i.e., tell how many there are in the data). Reestimate the model, this time using the Breslow method and exact partial likelihood correction for ties. (Note: the latter may take awhile to converge.) Discuss the differences between the models, if any, and suggest one or more rationales for why the differences do or do not exist.

Due: Monday, March 14

Discussion papers: Everyone needs to have chosen a paper and informed Jan of what paper they want to discuss and where the paper can be found by Monday, January 24th. If it is on JSTOR or PROceedings, everyone can easily get a copy. If it is a working paper we can make it available on the website or via email.

Other things we’d do if we had more time, so much to do & so little time/things you might do for fun in your spare time.

Lab Exercise IV: More residual diagnostics! Look at the functional form of a covariate, for influential observations, etc.

Lab Exercise V: Again using a data set of your own or the King, Alt, Burns, and Laver (1990) data set from Exercise 2, run a competing risks model. See Diermeier, Daniel, and Randy T. Stevenson. 1999. “Cabinet Survival and Competing Risks.” AJPS 43(4): 1051-68 for assistance with using the cabinet data with a competing risks set up.

Lab Exercise VI: Again using a data set of your own or the Oneal and Russett data set from Exercise 1, show how the data would be set up for running a conditional gap time model. Run the model and interpret the results. Estimate a frailty model. Discuss how the conditional gap time model differs from a frailty model. What are the advantages or disadvantages compared to one another.