Identification Using Border Approaches and IVs

We document that recent quasi-experimental strategies for identifying advertising effects can be derived from a model in which ad decisions are made at a more aggregate level than conversion is measured. Next, we show that the identifying variation in one of these strategies, the Border Approach, is conceptually similar to what are commonly known as Waldfogel IVs. We compare these, as well as supply-side instruments and fixed effects, in a data set on advertising in US presidential elections. Both border approaches and IVs are known to sacrifice statistical power and they do, but not by enough to affect statistical significance, in this application. The Waldfogel IVs are much more powerful than the supply-side IVs and, when combined, the standard errors are substantially reduced. Each IV estimator has the potential to produce a local average treatment effect that weights aggregate markets differently. Estimates suggest differences may exist, but they are not significant. When both IVs are combined, the point estimate is identical to a fixed effect estimate that is likely to be unbiased. The Border Approach can produce local effects at the disaggregate level when border and non-border regions differ. We find evidence of a statistically signficant difference when analysis is restricted to those counties where identifying assumptions are more plausible. The point estimate drops to nearly zero and becomes insignificant despite a standard error that is as small as the lowest IV standard error. We suspect local estimate concerns are greater for the Border Approach because it identifies advertising effects that exclude the high population counties in all markets, whereas IVs may weight each market differently but include counties of all types within each market