Television advertisers have traditionally cared about brand awareness—reaching the right audience at the right frequency and cost—but the industry is under growing pressure to demonstrate measurable treatment effects on conversion outcomes. This talk examines why that shift is so difficult to deliver on in practice. Unlike digital advertising, where randomized holdout experiments and persistent user identities enable credible causal inference, TV measurement involves a confluence of challenges: self-selected viewership with no auction mechanism to exploit, a patchwork of non-representative data sources, ecological inference challenges from household-level observation of individual behavior, and very small effect sizes. This talk traces the arc of academic progress in digital advertising measurement and explains the challenges of adapting them for TV. I will also discuss possible directions for future research at the intersection of statistics and causal inference where statisticians and causal inference researchers can make meaningful contributions.

When a consumer makes a choice, what can a statistician reveal that conventional wisdom might miss? This talk presents three empirical studies — spanning the US pharmaceutical market, the 2010 UK General Election, and India's 2016 demonetization — each using behavioural trace data recorded at the moment of decision rather than recalled afterwards. Together, they form a methodological arc: from joint Bayesian models that uncover hidden structure in drug prescribing, to dynamic choice models that reveal how media sentiment outlasts volume as elections approach, to causal identification at a national scale that surfaces $1.5 billion in retail behavior that aggregate policy analysis had not anticipated. In each setting, a natural intuition that requests drive prescriptions, that louder campaigns win, and that policy shocks affect all consumers uniformly proves incomplete. Statistical modeling does not contradict common sense so much as sharpen it, recovering structure that simpler approaches leave hidden. Because consumer behavior is not noise. It is the signal.