Hybrid methods are particularly well suited to marketing problems where treated units are few, pre-treatment periods are short, adoption is staggered and decision-makers care about both transparency and robustness. This section sketches four stylised applications drawn from common marketing settings. The examples are illustrative rather than empirical case studies. They show how hybrid designs are put together in practice and what credible findings typically look like when the methods are used carefully.

Application 1: Multiple-Market Loyalty Programme Rollout Suppose a retailer pilots a loyalty programme in ten flagship stores, chosen for their size and affluent urban locations, with forty comparable stores never receiving the programme during the study window. Quarterly revenue is observed for three years before the pilot and two years after. The goal is to estimate the average effect on quarterly revenue for the treated stores. Flagship stores are larger and more urban than most donors, so a hybrid that can correct systematic covariate imbalances is natural. ASCM fits this structure well [Ben-Michael et al., 2021]. For each flagship, the analyst constructs a synthetic control using pre-treatment revenue trajectories and store characteristics such as floor space, local income and urbanisation. A regularised outcome regression then adjusts for any remaining level differences linked to these covariates. Diagnostics show that ASCM tracks pre-treatment revenue more closely than standard SC: pre-period RMSPE is noticeably smaller relative to revenue volatility, covariate imbalances on store size and urbanisation are substantially reduced, and weights are spread across a moderate number of sensible donor stores rather than concentrating on one or two idiosyncratic outlets. Leave-one-donor-out checks indicate that no single donor dominates the estimates. When the pilot goes live, the hybrid counterfactual reveals an uplift in revenue that builds over the first few quarters as customers enrol and then levels off. The magnitude is economically meaningful but not explosive — think of a mid-single-digit percentage increase on a quarterly basis, with intervals that suggest a positive effect while still reflecting meaningful sampling noise. In-space placebo analyses, treating donor stores as if they had launched loyalty, yield much smaller gaps on average, suggesting that the observed pattern is not easily reproduced by chance. The retailer uses these results, together with cost information, to form ROI projections for a wider rollout.

Application 2: Staggered Advertising Campaign Launch Consider a brand that rolls out an advertising campaign in thirty regional markets over six quarters. Ten markets adopt in Q1, ten in Q3 and ten in Q5. Twenty markets never receive the campaign and serve as

7.11 Marketing Applications donors. The brand wants to understand how effects evolve over event time and whether early and late adopters respond differently. The staggered structure and potential differences in pre-treatment trends across cohorts make SDID a natural choice [Arkhangelsky et al., 2021]. For each adoption cohort, the analyst applies SDID using notyet-treated and never-treated markets as donors, estimating unit weights that align markets with similar pre-trends and time weights v̂t that emphasise pre-treatment periods where treated and donor markets move together. The cohort-time effects $ au$ (g, t) are aggregated into event-time effects $ heta$̂k using the framework from Chapter 5. Diagnostics show that, for each cohort, the SDID synthetic paths match pre-treatment sales reasonably well, with time weights v̂t concentrating on the few quarters just before adoption — a pattern that fits the idea that recent history is most informative. When seasonality is strong, you should check that the time weights are not simply picking one season and ignoring the rest. Event-time estimates tell a coherent story: effects are small on impact, build over a few quarters as awareness accumulates and then settle at a roughly constant lift. Pre-treatment leads in the event-time plot hover close to zero, supporting weighted parallel trends, and cohort-specific profiles look broadly similar rather than wildly heterogeneous. From a marketing perspective, the key insights are about timing and persistence: the campaign appears to ramp up over a few quarters and then sustain a modest but non-trivial sales lift, with little evidence that early or late adopters behave differently once you condition on pre-trends.

Application 3: Overlapping Loyalty and Promotional Programmes Now imagine a retailer that introduces two interventions over the same horizon: a loyalty programme launching in Q1 and a new promotional strategy starting in Q4. Ten markets receive both, ten receive only loyalty, ten only promotions and twenty receive neither. Management wants to know not just whether each programme works, but also whether they interact. The overlapping treatment structure creates identification challenges. For markets that receive both programmes, the relevant comparison is markets that receive neither, not those receiving only one programme. A hybrid extension of ASCM can help by constructing synthetic controls separately for each treatment sequence group, using never-treated markets as donors and including programme indicators in the augmentation model. In practice, the analyst would first check that each treatment-sequence group can be matched reasonably well in the pre-period using donors that never receive any programme. Augmentation then helps correct remaining level differences linked to store and market characteristics. This approach effectively treats “treatment sequence group” as the exposure, requiring that sequence assignment is not driven by unobserved shocks to untreated outcomes (beyond the no-anticipation and no-carryover conditions already assumed). Under the same identification assumptions that would justify a richer DiD or event-study model — no unobserved shocks tied systematically to the timing of loyalty or promotions — ASCM can help disentangle the average effect of each programme and their overlap.

A stylised outcome might be that loyalty produces a fairly stable, medium-sized sales lift that accumulates over several quarters, while promotions produce shorter-lived spikes around the launch period. Markets with both programmes might show combined effects that are smaller than the sum of individual effects, suggesting some substitution or saturation. The value of the hybrid here is in sharpening the counterfactuals for each group and making these patterns visible, not in eliminating the need for strong assumptions about how overlapping treatments are assigned.

Application 4: Flagship City Launch with Potential Spillovers Finally, consider a platform that launches a new service in a flagship city, with concern that neighbouring cities may be indirectly affected through word-of-mouth, competition or changes in supply. Five neighbouring cities are plausibly exposed to spillovers. Thirty more distant cities form a candidate donor pool. The platform wants to estimate both the direct effect in the flagship and any spillovers among neighbours. For the direct effect, ASCM applied to the flagship city with only distant donors in the pool is a natural starting point. The key design step is donor curation: neighbours are excluded from the donor set because they may be contaminated by spillovers, while distant cities are retained if there is no compelling reason to think they are affected by the launch. As in earlier examples, diagnostics focus on pre-treatment fit, balance on relevant city characteristics and weight stability. Estimating spillovers requires adopting the exposure-mapping framework from Chapter 11. One simple mapping treats neighbours as “exposed” (using the exposure-mapping notation hi (D−i,t ) from Chapter 11) once the flagship launches. For each neighbour, the analyst constructs a synthetic control from distant cities and includes exposure indicators in an augmentation model that captures how similar the neighbour is to the flagship on demographics and baseline adoption. Under the assumption that distant donors are unaffected and that exposure captures the main channel linking the flagship to neighbours, differences between neighbours and their synthetic controls after launch can be interpreted as spillover effects. In a plausible outcome, the flagship city shows a relatively large, sustained jump in adoption following launch, while neighbours exhibit smaller, delayed increases that are stronger for nearer cities than for those further away. The platform can then form a regional impact measure that combines the direct effect in the flagship with spillover effects in neighbours, weighted by population. That informs decisions about where to launch next: cities with dense clusters of nearby markets may generate more total value than isolated cities with similar own-market potential.

Summary These stylised applications show how hybrid methods can be woven into realistic marketing analyses. The common elements are straightforward. You choose a hybrid that matches the structure of the problem — ASCM when covariate imbalance is central, SDID when staggered timing and differential trends matter, more

7.11 Marketing Applications complex factor-based hybrids only when the data and team can support them. You curate donors carefully to respect comparability and spillovers. You lean heavily on diagnostics — pre-treatment fit, covariate balance, weight patterns, placebos and sensitivity checks — to assess whether the resulting counterfactuals are credible. Quantitatively, credible marketing findings from these designs often tend to be modest in size (single- to low-double-digit percentage lifts), with uncertainty intervals that acknowledge real sampling noise and with patterns that are consistent across neighbouring periods and alternative estimators. Used in this way, hybrids provide a disciplined way to extract causal insight from complex marketing panels without promising more precision than the data can support.

References

  • Shaw, C. (2025). Causal Inference in Marketing: Panel Data and Machine Learning Methods (Community Review Edition), Section 7.11.