Why SUTVA fails in marketing
SUTVA says one unit’s outcome depends only on its own treatment. In marketing panels that is often false: a loyalty program in one store affects nearby stores, ads in one market spill over into adjacent markets, and competitor responses propagate through the network.
When SUTVA fails, the estimand changes. The question becomes: what is the effect of own treatment combined with neighbors’ treatment exposure?
Exposure mappings make spillovers explicit
Instead of $Y_{it}(d)$, outcomes can depend on an exposure mapping $h_i(D_{-i,t})$:
$$ Y_{it}(d, h_i(D_{-i,t})). $$This makes the causal estimand explicit about both own treatment and spillover dose.
Design choices that mitigate spillovers
- Cluster randomization: randomize at a cluster level to internalize within-cluster spillovers.
- Buffer zones: separate treated and control regions to reduce contamination.
- Explicit exposure measures: use geography or network links to define $h_i(\cdot)$.
These are not afterthoughts. They are part of the assignment mechanism.
Why estimands shift
If you randomize individual stores but ignore cross-store advertising spillovers, the estimated effect is a mixture of direct and indirect effects. The correct estimand in the presence of interference depends on both own treatment and the exposure pattern in neighbors.
Practical implications
- Diagnose plausible spillover paths before choosing a design.
- Align cluster definitions with the likely interference structure.
- Interpret “direct effects” cautiously when treatment can propagate.
Takeaway
SUTVA violations are the norm, not the exception, in marketing panels. Credible designs either internalize spillovers through clustering or model them through exposure mappings. Otherwise, the estimand is mis-specified and the interpretation becomes unstable.
References
- Shaw, C. (2025). Causal Inference in Marketing: Panel Data and Machine Learning Methods (Community Review Edition), Section 3.2.6.