Using Pooled Regression to Model Different Regions in Marketing Mix Modeling

How pooled regression measures sales across regions, stores, and retailers in Marketing Mix Modeling, why normalization keeps the results honest, and where hierarchical Bayesian methods now fit alongside pooled OLS.

National models cannot answer regional questions

Marketing Mix Modeling at the national level treats sales as a single series and asks one question: what moved the total. That works until the question becomes regional. A retailer wants to see sales store by store. A consumer goods brand wants to compare performance region by region. Once measurement has to happen below the national line, a single national equation cannot answer it, and pooled regression is the technique that can.

For a marketing leader, this is not an academic distinction. A national model can confirm that television worked. It cannot tell you which regions to move budget into, or defend that reallocation to a finance team that wants accountability rather than conviction. Regional measurement is what turns a model result into a decision you can take to the board, granular enough to act on and specific enough to stand up to scrutiny on return on investment.

This article explains how pooled regression models many regions in one equation, why untransformed data can produce results that point the wrong way, and how the same idea extends to the hierarchical Bayesian methods now common in geo-level MMM.

Pooled regression combines cross-sectional and time-series data

Pooled regression analyzes data that has two dimensions at once. The cross-sectional dimension captures differences between entities at a point in time. The time-series dimension captures change within a single entity over time. Combining the two gives a fuller view of how variables relate to each other, both across regions and across periods.

Under the hood it is still ordinary least squares, so the estimator fundamentals in regression analysis for Marketing Mix Modeling all still apply. It belongs to the broader family of panel-data methods. Pooled ordinary least squares (pooled OLS) is the simplest member of that family: it stacks every entity into one dataset and fits a single equation. Fixed-effects and random-effects estimators are its more elaborate relatives, and they become relevant the moment regions differ in ways the pooled model has to account for. In MMM, a cross-section is whatever unit you want to measure below the national level:

• Market. a Designated Market Area, such as Los Angeles, Dallas, Washington, or Philadelphia
• Media. a media region, such as a BARB TV region
• Store. a store, such as two separate Walmart branches
• Retailer. a retailer, such as Walmart, Walgreens, Costco, or Target
• Product. a sub-product, split by pack size or flavor
• Grouping. a chosen aggregation, such as grocery stores against pharmacy stores

Choosing the right level of detail is a discipline in its own right, which we set out in the granularity trinity in Marketing Mix Models. Pooled regression is how you measure that detail once you have chosen it.

For a wider view of how the technique fits a modern modeling workflow, see our companion article on a contemporary pooled regression approach.

More data points make every estimate more robust

The more observations a model sees, the more reliably it can measure an effect. Pooling regions multiplies the observation count, and that has two consequences. The model estimates every coefficient on more data, so the measures are steadier. And effects that only exist at the regional level become measurable at all, which a national model can never do.

Robustness is not only a matter of more data. It also depends on how you check the model, which we cover in the three layers of MMM validation.

Regional modeling stacks the data into one equation

Take a product sold through three main UK retailers: Boots, Asda, and Superdrug. Suppose each has weekly data running from 1 January 2020 to 31 December 2022. That is three years, or roughly 156 weekly observations per account. Pool the three together and the dataset becomes 156 x 3 = 468 rows.

The number of observations roughly tripled. Ordinary least squares stacks the data from every account and fits a single regression line through all of it. That expansion buys three things: more robust measures, the ability to measure impact at the regional level, and the ability to capture more media granularity than a national model would support.

Pooling regions with different averages breaks OLS

Pooled OLS behaves well only when the stacked regions sit at comparable levels. In practice they often do not. An analyst frequently pools high-selling regions with low-selling ones. Skip the transformation and run OLS on the raw data, and the model splits the difference: it overestimates sales in the regions that sell below the pooled average and underestimates them in the regions that sell above it.

The clearest example is price against sales. Look at each region on its own and the relationship is the expected one: when price rises, sales fall, and when price drops, sales rise. Pool several regions without transforming the data, fit one line through all the points, and the slope can flip. Price and sales appear to move together, which is the opposite of what economics and every individual region tell you.

Normalization makes pooled regression valid

If the point of pooling is to read regional detail, you usually need to normalize the data first. There are several ways to do it: subtract the variable’s mean, divide it by its own mean, or divide it by the mean of the dependent variable.

Subtracting each region’s mean is worth naming precisely: it is the within transformation, which is exactly a fixed-effects specification. That is why analysts often call the pooled model, once normalized this way, pooled regression with fixed effects. Two cases need no separate step. Log-linear modeling normalizes through the log transform itself, so the data is already on a comparable scale. National models do not need normalization at all, because there is only one region. Normalized this way, the working assumption is that a variable’s effect is the same across regions in percentage terms, which is a fair default. We cover the log-linear route, and when to prefer it, in log-linear modeling for Marketing Mix Modeling.

Splitting variables captures different regional sensitivities

In MassTer, the Normalization processor handles this transformation, and estimating the pooled data then reports the average impact of each activity across all fitted regions. For a national TV campaign, that is the right default: you assume the percentage impact is equal across regions, and once you multiply it by the size of each region, the absolute impact is larger in bigger regions, which is what you would expect.

Sometimes that assumption is too strong. Some regions really are more sensitive to an activity than others. When an analyst models north against south and believes southern customers are more price-sensitive, the Splitter transformation lets them split the price variable across the two regions and estimate a separate coefficient, and therefore a separate elasticity, for each. The rule of thumb is simple: assume the same percentage contribution across regions by default, and split a variable only when there is reason to believe a region behaves differently.

We cover the mechanics of preparing and dividing those regional inputs in data splitting in Marketing Mix Modeling.

Partial pooling extends the same logic for 2026

The choice between one shared coefficient and a separate coefficient per region is not binary. It is a spectrum, and pooled OLS and the Splitter sit at its two ends.

• Complete pooling (pooled OLS): one coefficient for every region. Efficient and stable, but it assumes regions behave identically in percentage terms once normalized.
• No pooling (per region): a separate coefficient for each region, which is the Splitter taken to its limit. Flexible, but small or sparse regions have too little data and their estimates get noisy.
• Partial pooling (hierarchical models): the model draws each region’s coefficient from a shared distribution, so every region’s estimate borrows strength from the others: it pulls small regions toward the group average while large regions keep their own signal.

Partial pooling is the idea behind geo-level Bayesian hierarchical MMM, the approach used in Google’s open-source Meridian and set out in Google’s research on geo-level Bayesian hierarchical media mix modeling. Meridian supports models with 50 or more geographies on two to three years of weekly data, and it produces tighter credible intervals for smaller markets precisely because it shares information across regions. Meta’s Robyn reaches for a different tool, regularized ridge regression, rather than a Bayesian hierarchy, but the goal is the same: stable, granular estimates when regions differ.

None of this replaces the pooled regression described above. It is the same instinct expressed with more statistical machinery. Complete pooling and no pooling are the extremes; partial pooling is the considered middle. Regression remains the engine, a point we make at length in why modern MMM still runs on its oldest engine.

Frequently asked questions

Conclusion

Pooled regression lets Marketing Mix Modeling measure below the national line, by store, retailer, product, or market. Pooling enlarges the dataset, which makes every measure more robust and makes regional effects visible in the first place. You estimate pooled OLS with ordinary least squares, and its validity depends on the stacked regions sitting on a comparable scale. When they are not, normalization keeps the results valid, unless you are already working in log-linear form, where the transform does the job for you.

For the fundamentals this builds on, from data preparation to model estimation, see our Comprehensive MMM Guide.

Those regional coefficients do not stay in the model either. They feed budget decisions downstream through response curves and channel synergy.

The deeper decision is how much to let regions differ. Assume one shared effect, allow every region its own, or let them share strength through partial pooling. Before your next regional model goes to a planning meeting, it is worth asking one question: are you assuming your regions behave the same because they do, or because you never checked?