Consequences (and Solutions) of Violating SUTVA in a DID Design

Question

Using a difference-in-differences design, what are the consequences of violating the Stable Unit Treatment Value Assumptions (SUTVA)? I ask this because, in my specific field of research (conflict and conflict management), succesful treatments are strongly recognized as geographically "contagious". As a result, it seems like SUTVA will inherently be violated in any form of causal analysis in this area of study. Can the consequence be stated as simply you will have a biased causal estimate if SUTVA is violated? If so, does violating SUTVA using DID bias the estimated causal effect by over- or underestimating the causal effect?

As for solutions, this Wikipedia article states that inverse probability weighting (IPW) can be used to address spillover concerns. However, the article does not provide any citations for this and, after reading the page, I am still a bit confused of how I would apply IPW to a DID design. So this really leaves me with these questions

• How exactly does IPW adjust for the violation of SUTVA?
• Does the IPW solution work with DID?

Answer 1

The IPW allows you to build a model of how units select into treatment. For example, adoption probability can be a declining function of geographic distance from the pilot regions. Then these weights down-weight untreated units with very different probability of adoption when you do the DID comparison. The vanilla DID weights all comparison control units equally.

There is R code, examples, and papers cited here.

The fact that only successful treatments are contagious may still present a problem, but it’s hard to say more without lots of detail about the setting and how treatment is assigned. It may work if you can control for lagged outcomes and distance to pilot regions in the assignment equation.