Nine dollar price ending statistics across three retail experiments
https://saidmaybe.com/guides/nine-dollar-price-ending-retail-statistics/
## Quick answer
Nine dollar price ending statistics support a more specific conclusion than the familiar claim that a price ending in 9 always sells more. Anderson and Simester reported three catalog experiments with positive selected 9 ending estimates, but the estimates varied in precision and the third depended on whether a sale cue was present. These were whole-dollar endings, including prices such as USD 39, not a blanket test of every .99 price. (Anderson and Simester (https://www.kellogg.northwestern.edu/faculty/anderson_e/htm/PersonalPage_Files/Papers/Effects_of_9_Price_Endings_on_Retail_Sales.pdf), Mar 2003)
A price ending is worth testing when the resulting price still makes sense for the product and its margin. This analysis preserves the experimental context, compares the published coefficients with their uncertainty and separates the ending from the amount charged. Changing USD 44 to USD 39 changes both.
## Key Takeaways
All three selected specifications estimated a positive whole-dollar 9 ending effect. (Anderson and Simester (https://www.kellogg.northwestern.edu/faculty/anderson_e/htm/PersonalPage_Files/Papers/Effects_of_9_Price_Endings_on_Retail_Sales.pdf), Mar 2003).
The third plotted estimate applies to products without a sale cue. (Anderson and Simester (https://www.kellogg.northwestern.edu/faculty/anderson_e/htm/PersonalPage_Files/Papers/Effects_of_9_Price_Endings_on_Retail_Sales.pdf), Mar 2003).
Historical dollar endings do not directly establish an effect for every .99 ecommerce price. (Anderson and Simester (https://www.kellogg.northwestern.edu/faculty/anderson_e/htm/PersonalPage_Files/Papers/Effects_of_9_Price_Endings_on_Retail_Sales.pdf), Mar 2003).
## What was actually randomized?
The retailer mailed catalog versions with different prices to customer samples. Study 1 sent three versions to 20,000 customers each. Study 2 sent two versions to 31,250 each, and Study 3 sent three versions to 90,000 each. The paper describes assignment through geographical customer groups rather than independently repricing every shopper's individual transaction. (Anderson and Simester (https://www.kellogg.northwestern.edu/faculty/anderson_e/htm/PersonalPage_Files/Papers/Effects_of_9_Price_Endings_on_Retail_Sales.pdf), Mar 2003)
Figure description: Nine dollar price ending statistics: selected log demand coefficients were 0.351, 0.152 and 0.071 across three catalog studies, with approximate 95 percent intervals; the third estimate applies to items without sale cues.
Figure 1: Positive price ending estimates had different precision and conditions, rather than one transferable conversion lift. Source: Anderson and Simester (https://www.kellogg.northwestern.edu/faculty/anderson_e/htm/PersonalPage_Files/Papers/Effects_of_9_Price_Endings_on_Retail_Sales.pdf), Mar 2003. Original comparison from the cited study. Study period and calculation limits appear in the text.
The demand models used item by version observations. The plotted specifications have 219, 422 and 924 observations, respectively. Those counts are different from the number of households mailed. A large recipient base does not make every individual product effect precisely estimated. (Anderson and Simester (https://www.kellogg.northwestern.edu/faculty/anderson_e/htm/PersonalPage_Files/Papers/Effects_of_9_Price_Endings_on_Retail_Sales.pdf), Mar 2003)
For an ecommerce test, hold the price display consistent from product page through cart, checkout and the eventual message. A shopper who sees two different prices across surfaces has received a different intervention from a shopper who consistently sees one price point. The test must also allow for any change in margin from the price itself.
## How uncertain were the selected estimates?
The selected coefficient and standard error pairs are .351 and .172 for Study 1, .152 and .041 for Study 2, and .071 and .030 for the no sale cue specification in Study 3. Our approximate intervals use the coefficient plus or minus 1.96 standard errors. (Anderson and Simester (https://www.kellogg.northwestern.edu/faculty/anderson_e/htm/PersonalPage_Files/Papers/Effects_of_9_Price_Endings_on_Retail_Sales.pdf), Mar 2003)
This gives intervals of about .014 to .688, .072 to .232 and .012 to .130 in coefficient units. These are calculations from the reported models. They are not intervals for a shopper conversion percentage, and the specifications should not be averaged into a single store forecast. (Anderson and Simester (https://www.kellogg.northwestern.edu/faculty/anderson_e/htm/PersonalPage_Files/Papers/Effects_of_9_Price_Endings_on_Retail_Sales.pdf), Mar 2003)
The first result is much less precise than its headline alone suggests. The third is narrower but conditional on a different product presentation. Interval overlap does not test whether the effects changed across studies. The catalogs, item mix and model terms differ, so a claim that the technique became weaker over time would require more than arranging the point estimates in order.
## What does the sale cue change?
The third experiment explicitly distinguishes the presence of sale cues. The plotted .071 coefficient is for a 9 ending without a sale cue. The paper's broader interpretation is that price endings can serve as information about price levels, with a different role when a sale label already supplies such a cue. (Anderson and Simester (https://www.kellogg.northwestern.edu/faculty/anderson_e/htm/PersonalPage_Files/Papers/Effects_of_9_Price_Endings_on_Retail_Sales.pdf), Mar 2003)
That is a reason to keep promotional presentation in a price experiment's specification. A new regular price, a discounted price and a sale label can send different signals even when the displayed amount is the same. A merchant testing an ending while simultaneously adding a countdown, comparison price and promotional email cannot isolate the contribution of the ending.
The source also examines new versus established items. Treat that as a condition to consider, rather than evidence that every newly launched product will respond. A shopper's knowledge of the item and the store's prior pricing may differ from the catalog audience's knowledge.
## How should a merchant judge a price point?
Judge the price policy through contribution after product cost, fulfillment, discount expense and returns. More units do not automatically make a lower price profitable. A price ending can increase demand while the reduction in collected revenue still outweighs the gain in volume.
Use a design that separates price level from the cue as far as the store's commercial constraints allow. Define eligible products, protect price consistency and retain the assigned comparison even if the shopper returns later. Decide whether the outcome is units, purchasers, order value or contribution per eligible visitor before reading the results.
The historical evidence gives a reason to run a focused price test. It does not authorize an unsupported reference price or a fictitious sale claim. The useful transfer is the experimental question: does this complete price policy improve the store's outcome for the relevant products and shoppers?
## Is a log-demand coefficient the same as a percentage sales lift?
The selected model reports coefficients on the logarithm of unit demand. Holding the model's other terms fixed, the exact demand multiplier associated with an ending indicator is the exponential of its coefficient. The common shortcut of multiplying the coefficient by 100 is an approximation, particularly for the largest estimate. (Anderson and Simester, model specification and Tables 2, 4 and 7 (https://www.kellogg.northwestern.edu/faculty/anderson_e/htm/PersonalPage_Files/Papers/Effects_of_9_Price_Endings_on_Retail_Sales.pdf), Mar 2003)
Our transformations of the selected .351, .152 and .071 coefficients give approximately 42.05%, 16.42% and 7.36% modeled demand differences, respectively. These are conditional interpretations of the published log model, not observed conversion-rate changes or forecasts for a new store. The first transformation is notably larger than the 35.1% linear approximation. Retain the underlying coefficient, selected specification and uncertainty when citing the transformed number.
Changing a displayed price also changes its level, not just its final digit. The source model includes price controls. Its ending coefficient therefore cannot simply be pasted onto a merchant's proposed change from 40 to 39 and treated as the predicted total effect. That proposal changes the price level and the ending together.
## How much extra unit demand would preserve contribution?
Consider a worked price change from 40 to 39 currency units, with variable cost unchanged. At zero variable cost, unit sales must increase by approximately 2.56% to preserve contribution. At a variable cost of 25, contribution per unit falls from 15 to 14 and the required unit increase is 7.14%. At a variable cost of 35, contribution falls from 5 to 4 and the hurdle becomes 25%. These are accounting calculations using declared example inputs, not effects measured in the price-ending experiments.
Figure description: For a worked price change from 40 to 39, contribution break-even requires 2.56 percent more unit sales at zero variable cost, 7.14 percent more at cost 25 and 25 percent more at cost 35. The required increase rises as the remaining unit contribution narrows.
Figure 2: The contribution hurdle depends on the margin left after the price change. A positive demand effect alone does not establish a profitable pricing decision. Source: Original contribution arithmetic; Anderson and Simester, price-ending experiment context (https://www.kellogg.northwestern.edu/faculty/anderson_e/htm/PersonalPage_Files/Papers/Effects_of_9_Price_Endings_on_Retail_Sales.pdf), Mar 2003. Original comparison from the cited study. Study period and calculation limits appear in the text.
Calculate the hurdle by dividing the old contribution per unit by the new contribution per unit, then subtracting one. If the new unit contribution is zero or negative, selling more units cannot recover the old positive contribution under this simple model. Shipping subsidies, payment fees, returns and fulfillment costs belong in variable cost when the price decision changes them.
## What should a price-ending test keep separate?
Decide whether the question concerns the ending itself or the whole proposed price policy. A test of a real price change can be commercially useful even when it cannot isolate the psychological contribution of the final digit. Describe it as the combined change rather than attributing all movement to the ending.
Keep the sale cue, product presentation, shipping conditions and eligibility consistent if they are not part of the decision. The historical third experiment shows why the sale cue deserves attention. An ending effect selected from a no-sale-cue specification does not supply the same prediction when a promotional claim is added.
Judge the actual policy on contribution per assigned shopper over a suitable window. Preserve unit demand, revenue, cancellations and product mix as explanatory outcomes. An increase in demand concentrated in a low-margin item can produce a different result from the same unit increase spread across the assortment. The source supplies credible reasons to test the price presentation, while the margin hurdle makes clear what that test needs to earn.
## Which calculations can a reader reproduce?
The download retains the published inputs, their units and their comparison populations. The following arithmetic uses those inputs. It does not add a new sample or a new causal comparison.
Calculation | Expression | Result
Study 1 approximate coefficient lower 95 percent limit | 0.351-1.96*0.172 | 0.014
Study 1 approximate coefficient upper 95 percent limit | 0.351+1.96*0.172 | 0.688
Study 2 approximate coefficient lower 95 percent limit | 0.152-1.96*0.041 | 0.072
Study 2 approximate coefficient upper 95 percent limit | 0.152+1.96*0.041 | 0.232
Study 3 approximate coefficient lower 95 percent limit | 0.071-1.96*0.030 | 0.012
Study 3 approximate coefficient upper 95 percent limit | 0.071+1.96*0.030 | 0.130
## Methodology and Sources
We read the cited primary source and the decisive methods, results and exceptions on 13 November 2026, retaining its PDF and text under the source ledger. The download distinguishes reported aggregates from our arithmetic, with units and analysis populations preserved. The formulas shown here can be rerun from analyze.py. We do not reconstruct individual records from rounded aggregates or treat an implementation suggestion as a measured result. (Anderson and Simester (https://www.kellogg.northwestern.edu/faculty/anderson_e/htm/PersonalPage_Files/Papers/Effects_of_9_Price_Endings_on_Retail_Sales.pdf), Mar 2003).
Whole-dollar 9 endings such as USD 39, not a direct test of every .99 price. Catalog households and item-level demand, not current ecommerce checkout conversion. The third plotted specification applies to items without a sale cue. Approximate intervals use reported asymptotic standard errors and rounded coefficients. Their overlap is not a direct test of whether the studies differ.
## Glossary
Price ending: The final whole dollar digit tested in these catalog experiments.
Log demand coefficient: A model parameter on a logarithmic demand scale. Exponentiating it gives a demand ratio within the stated model.
Standard error: A measure of the statistical uncertainty around an estimated coefficient.
Confidence interval: A model-based interval whose interpretation depends on the statistical and sampling assumptions used to construct it.
Price control: An adjustment for the amount charged, helping distinguish that amount from its ending.
Sale cue: Wording that signals a promotion in the source experiment.
## Sources
Anderson and Simester: Effects of $9 Price Endings on Retail Sales: Evidence from Field Experiments (https://www.kellogg.northwestern.edu/faculty/anderson_e/htm/PersonalPage_Files/Papers/Effects_of_9_Price_Endings_on_Retail_Sales.pdf)
Reproduce the calculations
The aggregate source inputs retain the study period, unit and denominator.
Inspect the additional analysis
Reported inputs and original calculations retain their units and conditional scope.
Replay the additional calculations
Run this file beside its CSV to verify the arithmetic and explicit scope.
Quick answer
Nine dollar price ending statistics support a more specific conclusion than the familiar claim that a price ending in 9 always sells more. Anderson and Simester reported three catalog experiments with positive selected 9 ending estimates, but the estimates varied in precision and the third depended on whether a sale cue was present. These were whole-dollar endings, including prices such as USD 39, not a blanket test of every .99 price. (Anderson and Simester, Mar 2003)
A price ending is worth testing when the resulting price still makes sense for the product and its margin. This analysis preserves the experimental context, compares the published coefficients with their uncertainty and separates the ending from the amount charged. Changing USD 44 to USD 39 changes both.
The retailer mailed catalog versions with different prices to customer samples. Study 1 sent three versions to 20,000 customers each. Study 2 sent two versions to 31,250 each, and Study 3 sent three versions to 90,000 each. The paper describes assignment through geographical customer groups rather than independently repricing every shopper's individual transaction. (Anderson and Simester, Mar 2003)
Figure 1: Positive price ending estimates had different precision and conditions, rather than one transferable conversion lift. Source: Anderson and Simester, Mar 2003. Original comparison from the cited study. Study period and calculation limits appear in the text.
The demand models used item by version observations. The plotted specifications have 219, 422 and 924 observations, respectively. Those counts are different from the number of households mailed. A large recipient base does not make every individual product effect precisely estimated. (Anderson and Simester, Mar 2003)
For an ecommerce test, hold the price display consistent from product page through cart, checkout and the eventual message. A shopper who sees two different prices across surfaces has received a different intervention from a shopper who consistently sees one price point. The test must also allow for any change in margin from the price itself.
The selected coefficient and standard error pairs are .351 and .172 for Study 1, .152 and .041 for Study 2, and .071 and .030 for the no sale cue specification in Study 3. Our approximate intervals use the coefficient plus or minus 1.96 standard errors. (Anderson and Simester, Mar 2003)
This gives intervals of about .014 to .688, .072 to .232 and .012 to .130 in coefficient units. These are calculations from the reported models. They are not intervals for a shopper conversion percentage, and the specifications should not be averaged into a single store forecast. (Anderson and Simester, Mar 2003)
The first result is much less precise than its headline alone suggests. The third is narrower but conditional on a different product presentation. Interval overlap does not test whether the effects changed across studies. The catalogs, item mix and model terms differ, so a claim that the technique became weaker over time would require more than arranging the point estimates in order.
The third experiment explicitly distinguishes the presence of sale cues. The plotted .071 coefficient is for a 9 ending without a sale cue. The paper's broader interpretation is that price endings can serve as information about price levels, with a different role when a sale label already supplies such a cue. (Anderson and Simester, Mar 2003)
That is a reason to keep promotional presentation in a price experiment's specification. A new regular price, a discounted price and a sale label can send different signals even when the displayed amount is the same. A merchant testing an ending while simultaneously adding a countdown, comparison price and promotional email cannot isolate the contribution of the ending.
The source also examines new versus established items. Treat that as a condition to consider, rather than evidence that every newly launched product will respond. A shopper's knowledge of the item and the store's prior pricing may differ from the catalog audience's knowledge.
Judge the price policy through contribution after product cost, fulfillment, discount expense and returns. More units do not automatically make a lower price profitable. A price ending can increase demand while the reduction in collected revenue still outweighs the gain in volume.
Use a design that separates price level from the cue as far as the store's commercial constraints allow. Define eligible products, protect price consistency and retain the assigned comparison even if the shopper returns later. Decide whether the outcome is units, purchasers, order value or contribution per eligible visitor before reading the results.
The historical evidence gives a reason to run a focused price test. It does not authorize an unsupported reference price or a fictitious sale claim. The useful transfer is the experimental question: does this complete price policy improve the store's outcome for the relevant products and shoppers?
The selected model reports coefficients on the logarithm of unit demand. Holding the model's other terms fixed, the exact demand multiplier associated with an ending indicator is the exponential of its coefficient. The common shortcut of multiplying the coefficient by 100 is an approximation, particularly for the largest estimate. (Anderson and Simester, model specification and Tables 2, 4 and 7, Mar 2003)
Our transformations of the selected .351, .152 and .071 coefficients give approximately 42.05%, 16.42% and 7.36% modeled demand differences, respectively. These are conditional interpretations of the published log model, not observed conversion-rate changes or forecasts for a new store. The first transformation is notably larger than the 35.1% linear approximation. Retain the underlying coefficient, selected specification and uncertainty when citing the transformed number.
Changing a displayed price also changes its level, not just its final digit. The source model includes price controls. Its ending coefficient therefore cannot simply be pasted onto a merchant's proposed change from 40 to 39 and treated as the predicted total effect. That proposal changes the price level and the ending together.
Consider a worked price change from 40 to 39 currency units, with variable cost unchanged. At zero variable cost, unit sales must increase by approximately 2.56% to preserve contribution. At a variable cost of 25, contribution per unit falls from 15 to 14 and the required unit increase is 7.14%. At a variable cost of 35, contribution falls from 5 to 4 and the hurdle becomes 25%. These are accounting calculations using declared example inputs, not effects measured in the price-ending experiments.
Figure 2: The contribution hurdle depends on the margin left after the price change. A positive demand effect alone does not establish a profitable pricing decision. Source: Original contribution arithmetic; Anderson and Simester, price-ending experiment context, Mar 2003. Original comparison from the cited study. Study period and calculation limits appear in the text.
Calculate the hurdle by dividing the old contribution per unit by the new contribution per unit, then subtracting one. If the new unit contribution is zero or negative, selling more units cannot recover the old positive contribution under this simple model. Shipping subsidies, payment fees, returns and fulfillment costs belong in variable cost when the price decision changes them.
Decide whether the question concerns the ending itself or the whole proposed price policy. A test of a real price change can be commercially useful even when it cannot isolate the psychological contribution of the final digit. Describe it as the combined change rather than attributing all movement to the ending.
Keep the sale cue, product presentation, shipping conditions and eligibility consistent if they are not part of the decision. The historical third experiment shows why the sale cue deserves attention. An ending effect selected from a no-sale-cue specification does not supply the same prediction when a promotional claim is added.
Judge the actual policy on contribution per assigned shopper over a suitable window. Preserve unit demand, revenue, cancellations and product mix as explanatory outcomes. An increase in demand concentrated in a low-margin item can produce a different result from the same unit increase spread across the assortment. The source supplies credible reasons to test the price presentation, while the margin hurdle makes clear what that test needs to earn.
The download retains the published inputs, their units and their comparison populations. The following arithmetic uses those inputs. It does not add a new sample or a new causal comparison.
Calculation
Expression
Result
Study 1 approximate coefficient lower 95 percent limit
0.351-1.96*0.172
0.014
Study 1 approximate coefficient upper 95 percent limit
0.351+1.96*0.172
0.688
Study 2 approximate coefficient lower 95 percent limit
0.152-1.96*0.041
0.072
Study 2 approximate coefficient upper 95 percent limit
0.152+1.96*0.041
0.232
Study 3 approximate coefficient lower 95 percent limit
0.071-1.96*0.030
0.012
Study 3 approximate coefficient upper 95 percent limit
We read the cited primary source and the decisive methods, results and exceptions on 13 November 2026, retaining its PDF and text under the source ledger. The download distinguishes reported aggregates from our arithmetic, with units and analysis populations preserved. The formulas shown here can be rerun from analyze.py. We do not reconstruct individual records from rounded aggregates or treat an implementation suggestion as a measured result. (Anderson and Simester, Mar 2003).
Whole-dollar 9 endings such as USD 39, not a direct test of every .99 price. Catalog households and item-level demand, not current ecommerce checkout conversion. The third plotted specification applies to items without a sale cue. Approximate intervals use reported asymptotic standard errors and rounded coefficients. Their overlap is not a direct test of whether the studies differ.
Use this citation when referring to the analysis. Retain the original study scope with any figures.
Said Maybe. "Nine dollar price ending statistics across three retail experiments." Said Maybe, November 2026. https://saidmaybe.com/guides/nine-dollar-price-ending-retail-statistics/
<a href="https://saidmaybe.com/guides/nine-dollar-price-ending-retail-statistics/">Nine dollar price ending statistics across three retail experiments</a> (Said Maybe, November 2026)