Simultaneity between Trade and Conflict: Endogenous Instruments of Mass Destruction

Abstract

The classical liberal belief is trade, which economically benefits countries, creates ties binding the interests of countries and reduces conflict. While the vast majority of the empirical literature supports this view, recent research questions these findings by also considering the reciprocal relationship between trade and conflict. If conflict also influences trade, then trade is an endogenous right hand side regressor and previous estimates which ignore this are inconsistent. This article determines when one uses appropriate instruments for the endogenous regressors that trade reduces conflict and conflict reduces trade. Failure to use such instruments results in inconsistent estimates and can lead to the spurious conclusion that trade increases conflict. The lesson is the use of inappropriate instruments can be worse than not using them at all. 4 Use of a probit or logit model in the first stage should not be used. See Kelejian (1971) and Angrist & Krueger (2001) for further discussion. 5 The standard errors of this second stage will be incorrect as they are based on the fitted values of conflict rather than the actual values. See Wooldridge (2002: 95) for the correction. Stata's ivreg2 command automatically corrects the standard errors and is used below to produce the estimates of the linear trade model.

Mission accomplished: A reply to Reuveny and Keshk

“…This latter distinction identifies the models as limited-or full-information methods. The estimation techniques used by Goenner (2011) and Keshk et al (2004) are limited-information methods, where the real difference is in their choice of instruments. 1 The choice of instruments needs to be motivated by theory, as theory can help to identify potential instruments that influence the endogenous regressor(s) and do not influence the dependent variable.…”

The build-up of coercive capacities

“…Similarly, as detailed out in Appendix Section A.2.4, the core assumption of the simultaneous equation model implies an exclusion restriction : Variables that only enter the import equation (vector Z im ) may statistically not influence the (perceived) likelihood of conflict onset, other than through increases in weapons inflow, and vice versa (for vector Z on ). As highlighted by Goenner (2011), this is a critical assumption that needs to be justified, similar to a standard instrumental variables approach. In line with Reuveny & Keshk (2013), we predominantly rely on a theoretical justification for our im-port and the onset equations; still, the instruments used may themselves be endogenous.…”

“…As reported in Appendix Section A.2.5, a linear probability model would require less modeling assumptions (Angrist & Pischke, 2009;Goenner, 2011) compared to Maddala's SEM or even an instrumental variable probit model. As can be seen from Panel D of Table A.2, results are very similar, as both the SEM, IV probit and IV 2SLS approaches in the end invoke similar assumptions for identifying the average effects of imports.…”

“…This is a critical assumption that needs to be justified, similar to a standard instrumental variables approach (as e.g. highlighted by Goenner, 2011). In line with Reuveny & Keshk (2013), we predominantly rely on a theoretical justification for our import and onset equations; still, the instruments used may themselves be endogenous.…”

“…Given the trade-off in adequately addressing the simultaneity problem and the modeling assumptions necessary, we chose both an instrumental variables probit approach (as argued for e.g. by Goenner, 2011) and the SEM approach of . Alternatives to the IV probit would have been a two stage estimation of a linear probability model (which is less dependent on functional form assumptions (Angrist & Pischke, 2009: 198ff.)).…”