Observational Price Variation in Scanner Data Does Not Reproduce Experimental Price Elasticities

The paper's primary experimental elasticity estimate is approximately −0.34 (p. 9). This estimate remains inelastic even for price changes of +/-20% and over time horizons of several months. On its face, this implies that the grocery retailer is consistently pricing in the inelastic portion of the demand curve – a striking result that deserves more scrutiny than the paper provides.

Consider what an elasticity of −0.34 means in practice. Suppose the retailer raises prices by 20%. At this elasticity, quantity sold falls by roughly 7%. Selling 7% fewer units at 20% higher price yields approximately 12% higher revenue.

Now consider the grocery industry's actual financial structure. Kroger, a large publicly listed Midwestern grocery retailer that matches the paper's description, reports gross margins of 20–25% and net profit margins of 2–3%. A revenue increase of 12%, with modest volume loss, would flow almost entirely to the bottom line. Gross margins would jump to roughly 30–35%, and net profit margins could rise to 13–15%: an order of magnitude above actual industry profitability.

The fact that grocery margins remain thin suggests either that the experimental elasticity is biased toward zero or that the experiment is measuring something other than the demand curve the retailer faces in equilibrium.

In applied pricing contexts, experimental prices are frequently not implemented as intended. Sources of error range from technical failures to human overrides of prices that staff perceive as "particularly outrageous."

The authors report that 93% of price changes during the observational period stem from temporary price reductions. If the experimental price variation was implemented without promotional signaling, it may not have been salient enough for shoppers to notice.

The paper does not discuss stockout events, which could especially bias experimental elasticity estimates. Price reductions increase the likelihood of stockouts, which censor the observed demand response.

The process of averaging elasticity estimates across products may introduce bias. The products included in the experiment were a non-random sample of relatively high-selling items.

Because the experimental price perturbations were mean-zero across products, a typical customer's total basket price may not have changed meaningfully.

The specifications estimated for Figure 10 would in principle allow quantification of substitution effects, though the paper does not report these estimates. It is important to distinguish between the category-level price elasticity of demand and the more elastic demand curve that a single retailer or brand faces in a competitive market.

• Clarification needed: what is the main estimator? I'm confused about what is being reported in the main results. Are the reported elasticities coming from OLS log–log regressions, or from 2SLS? Page 9 mentions that four models are estimated, but then I can't see which one is used in, e.g., Figure 2. The interpretation depends heavily on this, and the paper would benefit from making the baseline estimator and its identifying variation completely explicit up front (and keeping the OLS/IV results clearly separated).

• Framing concern: DID on elasticities is not automatically "observational bias." Independently of OLS vs. 2SLS, I find the DID framing confusing. The paper effectively treats the estimated elasticities as the outcome in a difference-in-differences design and concludes that the experiment "causes" an average reduction in the magnitude of own-price elasticities of roughly 2. I don't think that object can be interpreted as "observational bias" without further caveats. The key issue (which the authors hint at) is that the experiment affects the distribution of prices. Elasticities estimated during the experimental period are therefore elasticities at different price levels, and thus at different points on the demand curve, than those prevailing in the pre-period. There is no reason, even under perfect identification, that those elasticities should coincide with pre-period elasticities if elasticities are not constant in prices. Put differently, I understand the authors as suggesting that a DID effect of zero on the elasticity would be the null under "no bias," but that is not generally true once the treatment itself changes the price support over which the elasticity is evaluated.

• What I take away instead: functional-form restrictiveness. The interpretation that seems most natural is that the log-linear demand specification is too restrictive in this context. A constant-elasticity log–log model implicitly assumes the elasticity is invariant to the price level, which is exactly what the experiment appears to contradict. The central empirical fact is then less about "bias" per se and more about model misspecification / lack of flexibility: observational estimation under a restrictive functional form can deliver elasticities that fail to transport across price regimes. This, in turn, is a clean motivation for the large literature on flexible / semiparametric / nonparametric demand estimation.
• That said, the magnitude is large and worth digging into. The estimated change is substantial, so it may not be fully explained by "different price levels imply different elasticities." I think it would be interesting to explore how much of the change in elasticities can be explained by this channel (maybe using standard approaches to flexible demand estimation). This would give a sharper quantification of the residual change in elasticities that is truly driven by the gap between the experimental and the observational variation.

This paper speaks to a broader question in empirical IO / quantitative marketing: whether the demand-estimation methods we routinely apply to scanner data can be used reliably for policy analysis, including policies motivated by animal welfare (and related concerns such as emissions from meat production). A central policy-relevant object in this setting is substitution between meat and plant-based products. If plant-based options draw substantial demand away from meat when their prices fall—i.e., if the relevant cross-price elasticities (or diversion) are large—then subsidizing plant-based products could plausibly be an effective instrument to improve animal welfare (and reduce emissions). If instead meat and plant-based products are effectively separate markets—in the extreme case, because plant-based demand primarily comes from consumers who were already vegetarian—then such a subsidy would not achieve its intended goal.

At the moment, the paper speaks primarily to own-price elasticities. In future research, it would be very interesting to expand the analysis to cross-price elasticities between meat and plant-based products. It's very likely that a restrictive log-log model that assumes constant cross-price elasticities would also tend to not do well at matching the patterns in the experimental data. This would confirm that specifying a flexible functional form matters for the key questions mentioned above that rely on correctly capturing diversions across products.