In short: A category average describes the category and none of your products, because price elasticity by pack and channel varies with pack size, channel, brand position and competitive set. The spread matters because the sign of the decision often flips inside it, so a price move that adds margin in one channel destroys it in another. Regressing log quantity on log price is biased, since prices are chosen in response to expected demand rather than set at random. Cross-price effects frequently dominate own-price effects for a business with a portfolio, which is where most of the value in the analysis sits.
A category study says demand in your category has an elasticity of around minus one point four. You apply it, model a three percent price increase, expect a four percent volume decline, and take the margin.
What happens is a two percent decline in one channel, an eight percent decline in another, and almost nothing on the large pack while the small pack falls off a cliff. The average was right and it described none of your actual products.
Elasticity varies enormously within a category, and the variation is systematic rather than random. Using one number across a portfolio throws away exactly the information that makes pricing a decision rather than a guess.
The published meta-analyses are useful here, and mostly for what they say about spread rather than about the central figure. Tellis pooled elasticities from econometric sales models in 1988 and reported a mean of about minus one point seven six. Bijmolt, van Heerde and Pieters repeated the exercise in 2005 across 1,851 estimates and reported a mean nearer minus two point six, along with the finding that elasticities have grown more negative over time. Two careful meta-analyses seventeen years apart differ by nearly a full point on the central estimate, which is itself informative: if the literature cannot pin a category average to within a point, a figure lifted from a category study will not describe your 750ml pack in the discounter.
Where the variation comes from
Four sources, and each is worth estimating separately because each points at a different action.
Pack size. Small packs are frequently bought on impulse or for immediate consumption, and the price is a small absolute amount, so response is weak. Large packs are stock-up purchases where the shopper is more price aware and more likely to compare unit prices. The same brand can have quite different elasticity across its pack range.
Channel. A shopper in a discounter is a different shopper, or the same shopper in a different mindset, from one in a convenience store. Price comparison is easier in some channels than others, and competitive assortment differs.
Brand position. Strong brands with loyal buyers hold volume better under a price increase. That is most of what brand equity means commercially, and elasticity is one of the few ways to measure it directly.
Price point. Response is not linear. Crossing a psychological threshold or moving past a competitor's shelf price produces a step change that a constant elasticity model cannot represent. A move from 2.49 to 2.59 and one from 2.95 to 3.05 are not the same event even though the percentage is similar.
What the difference is worth
The reason the spread matters is that the decision frequently flips sign inside it, and the arithmetic takes a minute.
Take an item at 10.00 with a variable cost of 6.00, so a contribution margin of 4.00, selling 100,000 units and producing 400,000 of margin. Put the price up three percent to 10.30 and the unit margin becomes 4.30. At an elasticity of minus one point four, volume falls four point two percent to 95,800 units, and margin becomes 411,940, a gain of about twelve thousand. At an elasticity of minus two point six, volume falls seven point eight percent to 92,200 units, and margin becomes 396,460, a loss of about three and a half thousand. Same item, same move, opposite answer, and both of those elasticities are figures the published meta-analyses report as category averages.
The breakeven is worth carrying around, because it is a single division. A price rise of p percentage points on a contribution margin of m percent can absorb a volume loss of p divided by m plus p. Here that is three over forty-three, or just under seven percent, and dividing by the three percent price move gives a breakeven elasticity of about minus two point three. Anything less elastic than that and the move makes money. Anything more elastic and it does not.
That threshold does the real work in a pricing meeting, because it converts an argument about a coefficient nobody can pin down into a question about which side of minus two point three an item sits on. People who cannot agree on an elasticity to one decimal place can usually agree on which side of a threshold an item sits, and where they cannot, you have identified the item worth testing.
Notice also that the threshold moves with margin. On a twenty percent margin the same three percent rise absorbs three over twenty-three, about thirteen percent of volume, and the breakeven elasticity is around minus four point three. Low margin items tolerate far more elasticity than high margin ones, which is the opposite of the intuition most commercial teams bring into the room.
The estimation problem nobody mentions
The standard approach regresses log quantity on log price, and the coefficient is the elasticity. It is simple and it has a well-known flaw that will bias your results in a specific direction.
Prices get set by people responding to demand conditions. Promotions are scheduled when demand is expected to be strong, prices are cut when volume is weak, and increases are timed to periods when the business feels confident.
That correlation between price and unobserved demand conditions biases the estimate, usually toward finding less price sensitivity than exists, because the price cuts that were scheduled into weak periods appear to have produced less lift than they did. Villas-Boas and Winer set the problem out formally for brand choice models in Management Science in 1999 and showed that ignoring it moves the coefficients materially.
There is a related error that produces a much larger distortion and gets noticed far less often. In most transaction histories, the great majority of observed price variation is promotional, and shoppers respond to a temporary price cut far more strongly than to a change in the everyday price, because a promotion carries a signal about scarcity of the offer as well as a lower price. Fit one coefficient across all the weeks and you get something dominated by the promoted ones. The symptom is an elasticity in the region of minus three or minus four on an item whose everyday price has moved twice in two years. Use that figure to model a list price increase and you will forecast a volume collapse that does not happen, talk yourself out of a move that would have made money, and never find out.
The check takes an afternoon. Refit with promoted weeks excluded, and compare the two coefficients. A gap of a factor of two or more means the number in your pricing model is a promotion elasticity wearing the wrong label, and the everyday-price decision needs the restricted estimate. Report both, and label which one each pricing decision should use.
Three practical mitigations.
Control for what drives the pricing decision. Seasonality, competitor prices, distribution, and the promotional calendar. The more of the pricing rationale you can include as controls, the less of it is left in the error term to bias the coefficient.
Use variation you did not choose. Cost pass-through events, tax or duty changes, and competitor-initiated moves you had to follow are closer to exogenous than your own planned pricing, and they give cleaner identification.
Test with deliberate variation. A price test across matched store groups is the only method that fully resolves this, and it is available more often than it is used. Two or three tests a year on the products where the stakes are highest will teach you more than any amount of modelling on observational data.
Cross-price is where the money is
Own-price elasticity tells you what happens to an item when you move its price. Cross-price elasticity tells you what happens to everything else, and for a business with a portfolio that second effect frequently dominates.
Increase the price of the 500ml and some of its buyers move to the 750ml, which may carry a better margin. The net portfolio effect can be positive even when the item's own volume falls. Without cross-price terms, that entire dynamic is invisible and the decision looks worse than it is.
Put figures on it. The 500ml sells at 2.00 against a cost of 1.20, so a margin of 0.80 on 50,000 units, which is 40,000. The 750ml sells at 2.60 against a cost of 1.55, a margin of 1.05 on 30,000 units, which is 31,500. Portfolio margin is 71,500. Now raise the 500ml five percent to 2.10, taking its unit margin to 0.90, with an own elasticity of minus one point eight. Volume falls nine percent to 45,500, and that line contributes 40,950.
Stop there and the move looks worth about 950, which is thin enough that most teams would leave it. Carry on and account for where the lost units went. If four in ten of the 4,500 units that left the 500ml reappear on the 750ml, that pack sells 31,800 and contributes 33,390. Portfolio margin is 74,340, a gain of 2,840, and roughly two thirds of the value of the decision sits in the substitution rather than in the item that was repriced.
The switching rate is the parameter that matters, and it is rarely measured. Four in ten is an assumption there, and a defensible range might be two to six in ten, which puts the gain between 1,900 and 3,800. That is still a decision you can make, and a very different conversation from one where the answer is 950 and substitution was never mentioned.
Estimating a full cross-price matrix across a large range is not feasible, because the number of parameters explodes and there is not enough independent variation in the data to identify them. The workable approach constrains the structure: allow substitution within a category and need state, weight it by attribute proximity, and estimate a small number of parameters that govern the pattern rather than one per pair.
That is less flexible than a full matrix and it is estimable from real data, which the full matrix is not.
Sample size, and when to stop
The most common failure in elasticity work is estimating at a grain the data cannot support, then acting on a number that is noise.
An item in a channel with eleven observed price points over two years, most of them clustered, does not have an identified elasticity. Fitting one produces a coefficient with an enormous standard error, and reporting it without the error invites someone to act on it.
The practical approach is hierarchical. Estimate at the level where you have data, then borrow. A category-channel estimate from pooled data, shrunk toward it for items with thin evidence, and item-specific estimates only where the item has enough price variation of its own to identify one. Report which items have their own estimate and which are inheriting.
That is more honest than a table where every row has a number and nothing indicates which numbers are real.
What to do with it
Two uses, and the second is more valuable than the first.
The obvious use is pricing decisions: what does this move cost in volume and gain in margin, across the portfolio including the substitution.
The less obvious use is pack architecture. Elasticity by pack tells you where the price ladder between sizes is wrong. A ladder where the step from small to medium is priced below what shoppers will bear leaves margin on the table, and one priced above it pushes volume down to a pack with worse economics. Testing the ladder against actual response is a larger opportunity than any individual price move, and it is rarely examined because pack pricing tends to be set once and inherited.
The limit
Elasticity estimated from historical variation describes response within the range of prices actually observed. Extrapolating to a move larger than anything in your history is extrapolation, and consumer response to a large increase is frequently non-linear in ways a small-move coefficient cannot anticipate.
Estimates also decay. Consumer price sensitivity shifts with economic conditions, and an elasticity fitted through a period of low inflation will not describe behaviour in a high one. Refit regularly and treat a two year old estimate with suspicion.
Finally, elasticity says nothing about the competitive response, which for a material price move is often the dominant factor. A model that predicts your volume decline while assuming competitors hold price is answering half the question, and the other half is a judgement about their position rather than a number from your data. Making that judgement explicit alongside the estimate is better than embedding it silently in a scenario.
Start by estimating separately for your three largest packs in your two largest channels. If those six numbers are close together, the category average was fine. They usually are not.