How and When to Use the Political Cycle to Identify Advertising Effects

Abstract

A central challenge in estimating the causal effect of TV advertising on demand is isolating quasi-random variation in advertising. Political advertising, which topped $14 billion in expenditures in 2016, has been proposed as a plausible source of such variation and thus a candidate for an instrumental variable. We provide a critical evaluation of how and where this instrument is valid and useful across categories. We characterize the conditions under which political cycles theoretically identify the causal effect of TV advertising on demand, highlight threats to the exclusion restriction and monotonicity condition, and suggest a specification to address the most serious concerns. We test the strength of the first stage category-by-category for 274 product categories. For most categories, weak-instrument robust inference is recommended, as first-stage F-statistics are less than 10 for 221 of 274 product categories in our benchmark specification. The largest first-stage F-statistics occur in categories that typically advertise locally, such as automobile dealerships and restaurants. Failure to use the suggested specification leads to results that suggest violations of exclusion and monotonicity in a significant number of categories. Finally, we conduct a case study of the auto industry. Despite a very strong first stage, the IV estimate for this category is imprecise.