A common critique I have heard levied against conflict studies (research examining the causes, consequences, and solutions to violence such as civil war, terrorism, etc.) is the problem of selection bias. To qualify this, much of this research consists of non-random samples that are basically entire populations within a specified time frame (all countries experiencing conflict from 1989-2020, for example). Dependent on the author's research question, a sample may include 1) only conflict-level countries, 2) only post-conflict level countries, 3) both conflict and post-conflict-level countries, 4) conflict-level dyads (where each observation is a dyadic pairing between two armed actors), or, in the rare case, 5) all countries, regardless of conflict status.

After reviewing Cinelli et al. 20222, I am left thinking about how the points from this article apply to common practices in conflict studies. For example, suppose I am examining the effect of mediation on levels of civil violence (X $\rightarrow$ Y). In such studies, it is common to only include countries that are experiencing conflict since mediation does not occur in non-conflict cases. Therefore: X $\leftarrow$ Conflict $\rightarrow$ Y since Conflict creates an opportunity for mediation to occur and for levels of violence to increase. In this first scenario, is simply analyzing instances of conflict-level countries sufficient? Or, given that conflict represents a fork, should conflict (a dummy) be included directly in a model as a control variable?

Additionally, the non-random selection of only conflict-level countries feels like selection bias. Although, it does not follow the DAG of selection bias that Cinelli et al. 2022 outline: X $\rightarrow$ Z $\leftarrow$ Y. I am also confused on this point as well.


1 Answer 1


In your example it sounds like conflict is a confounder (not a collider), so it should be controlled for. Cinelli 2022 seem to argue that controlling for a collider would lead to "collider stratification bias", which they also refer to as "selection bias". Regardless of the naming, as soon as you have a graph you can rely on finding correct adjustment sets algorithmically, so what to control for should no longer be a problem. I imagine the real difficulty is justifying the directionality for some edges. Since the variables in conflict studies appear to be broad concepts, this is likely not an easy task with clear answers (economists often have similar problems, which is why instrumental variable methods are very popular).


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.