zeropaper, adapted for marketing

Consumer packaged goods firms routinely pair pre-launch conjoint studies with home-use product trials to forecast new-product demand, yet giving respondents free samples can inflate their stated willingness to buy independently of anything they learn about product quality. Because a new brand has no baseline awareness, standard pretests cannot separate genuine quality learning from trial-induced demand inflation. We isolate this bias experimentally using a three-arm randomized home-use trial conducted for Müller Yogurt's US launch, in which one control arm received a familiar, category-leading incumbent (Yoplait) where quality was already known. Merely receiving and trying the familiar incumbent inflates stated choice share by 16 to 19 percentage points—roughly twice the magnitude of typical hypothetical bias in consumer goods. Validating against post-launch scanner data, we show when debiasing improves volume forecasts and provide a simple pre-launch calibration benchmark.

Major household transitions—marriage, childbirth, divorce, retirement, and health shocks—trigger broad reallocations of household budgets, yet marketing research typically studies one life event or product category in isolation. Using twenty years of Panel Study of Income Dynamics data (1999–2019), we estimate a unified causal substitution matrix capturing how eight lifecycle transitions shift expenditure shares across nine consumption categories, and translate these shifts into lifetime-budget counterfactuals for targeting. A first birth, for example, shifts between $2,000 and $39,500 in lifetime household spending toward childcare while crowding out $3,000 to $61,500 in recreation. Our cross-event design also uncovers sharp demographic heterogeneity masked by pooled averages, including a 2.3 percentage-point reversal in out-of-pocket healthcare spending at the age-65 Medicare eligibility threshold and an income-tier reversal in childcare expenditure.

When evaluating staggered policy rollouts or promotional subsidies for durable goods, managers need to know whether the incentive generated net new adoption or merely pulled forward purchases that would have occurred later. Standard staggered difference-in-differences estimators cannot answer this question because a single treatment effect bundles permanent demand creation with forward-looking timing shifts—producing opposite-signed estimates on the same residential solar subsidy panel (NY-Sun, 122,195 applications across 913 ZIP codes). We develop Structural Imputation Difference-in-Differences, an estimator that embeds a forward-looking dynamic adoption model inside an imputation framework to separate demand activation from anticipation. We prove that while anticipation is point-identified from pre-adoption dips, net demand creation is fundamentally partially identified in finite time panels. Applied to NY-Sun, our decomposition shows that demand creation outweighs purchase acceleration by at least two to one.

Hierarchical Bayesian demand models generate personalized pricing policies that substantially outperform uniform pricing, but the resulting price schedules depend on thousands of individual- or market-specific parameters that field managers cannot interpret and regulators cannot audit. We ask how much of the profit premium from personalized pricing survives when a high-dimensional Bayesian pricing policy is distilled into a single closed-form equation using symbolic regression. Across NielsenIQ scanner data for cereal, yogurt, and fluid milk, a five-parameter equation at the three-digit ZIP-code level captures 30 to 62 percent of the profit gain from full Bayesian personalization in cereal and yogurt—matching regression and decision-tree benchmarks while fitting on a single line. We also identify two clear boundaries where symbolic compression fails: finely disaggregated retail markets with noisy estimates, and categories with extreme, tail-driven price sensitivity.