Discount buyer statistics and the repeat purchase denominator

Updated 14 source citationsSources checked 13 November 2026

On this page
  1. Quick answer
  2. Key Takeaways
  3. Do deeper discounts bring in customers who buy less later?
  4. Why does the denominator change the apparent direction?
  5. What do the other two experiments add?
  6. Can customer history alone explain the difference?
  7. What would make a deeper offer worth keeping?
  8. How large could the missing nonresponder outcome be?
  9. Methodology and Sources
Download study text

Quick answer

Discount buyer statistics can point in opposite directions when a report changes its denominator. In one randomized catalog study of established customers, buyers responding to the deeper promotion later ordered 6.89 units each, versus 7.67 in the control group, approximately 10.17% fewer. But more people responded to the promotion. Those initial responders generated approximately 219.87 later units per 1,000 people mailed, versus 154.59 for control. The second calculation still excludes later purchases by people who skipped the initial catalog. (Said Maybe calculations, Anderson and Simester, Nov 2026).

This analysis recalculates published tables from three field experiments initiated in 1999, published in February 2004, and followed for different periods. It asks how a merchant should evaluate a deeper offer when that offer changes who buys. The outcome is product units, not profit or complete customer lifetime value. The distinction matters before concluding either that discount buyers are worse customers or that a promotion creates more total future demand.

Key Takeaways

Do deeper discounts bring in customers who buy less later?

There is no single answer in the published study tables. The conditional comparison differs between the established-customer study and the two prospective-customer studies. Those studies also used different catalogs and follow-up periods.

Study and original populationFollow-upLater units per initial buyer, controlLater units per initial buyer, deeper promotionCalculated relative difference
A, previously purchasing customers28 months7.676.8910.17% fewer
B, prospective customers24 months1.591.8214.47% more
C, prospective customers22 months0.991.3334.34% more

Relative differences divide the deeper-promotion mean by the control mean and subtract one. The inputs are published to two decimal places, so the recalculations are approximate. They describe customers who purchased from the initial test catalog, not everyone assigned a catalog. (Said Maybe calculations, Anderson and Simester, Nov 2026).

The source marks the Study A and C future-unit differences as significant at the 1% level. Study B's table has no significance star for that row, and the accompanying text describes the result as marginal, with p below 0.10. The three directional comparisons should not be presented as equally certain. (Anderson and Simester, Feb 2004).

For a merchant, the conditional question is useful: what did customers entering under each offer subsequently buy? It is not sufficient to answer the budget question: what did making that offer do for the full eligible audience?

Why does the denominator change the apparent direction?

A promotion can increase the number of initial buyers while reducing the later average among those buyers. Dividing by the buyers and dividing by the original audience therefore answer different questions.

Discount buyer statistics sensitivity frontier: for an assumed control nonresponder mean of 0.20 later units, the promotion nonresponder mean must be about 0.135 for full-audience later units to break even, despite a positive responder-only contribution.
Figure 1: The missing nonresponders can change the full-audience answer. Source: Anderson and Simester, Table 3 inputs, Feb 2004. Original aggregate analysis of the cited source. Study populations and limits appear in the text.

Study A mailed control to 37,758 customers and the deeper promotion to 18,708. There were 761 initial buyers under control and 597 under promotion. Comparing those raw buyer counts would favor the much larger control audience. Dividing by the people mailed gives the opposite initial-response comparison, 2.015% versus 3.191%. (Said Maybe calculations, Anderson and Simester, Nov 2026).

To retain the initial audience as a denominator while measuring only the responder cohort, calculate:

Initial buyers × later units per initial buyer ÷ people mailed × 1,000

For control, the calculation is 761 × 7.67 ÷ 37,758 × 1,000, approximately 154.59 units. For promotion, it is 597 × 6.89 ÷ 18,708 × 1,000, approximately 219.87 units. The responder cohort's later units per original audience are approximately 42.23% higher even though later units per initial buyer are approximately 10.17% lower. (Said Maybe calculations, Anderson and Simester, Nov 2026).

Does that establish a forty-two percent increase in all future purchases?

No. It is an increase in the calculated contribution from initial responders, using the original audience size as the denominator. Someone who skipped the initial catalog contributes zero to that particular two-stage measure even if they bought from a later catalog. Their later purchases have not become observed zeros in reality.

The paper's tables of responder outcomes do not supply the missing future purchases for the full mailed audience. The authors discuss selection and forward buying in their wider analysis. Our recalculation does not overturn those analyses or recover the missing population outcomes. It makes the change in denominator visible.

That is why a dashboard needs precise outcome names. “Future units among initial buyers per 1,000 assigned” is longer than “customer value,” but it tells a reviewer what was counted. A familiar label cannot repair an incomplete outcome.

What do the other two experiments add?

They show that the same reporting distinction exists even where the conditional and audience-based responder measures move in the same direction. They do not supply a common long-run multiplier for discounting.

StudyInitial responders' later units per 1,000 mailed, controlInitial responders' later units per 1,000 mailed, promotion
AApproximately 154.59Approximately 219.87
BApproximately 3.23Approximately 6.85
CApproximately 2.28Approximately 5.33

Each row uses its own mailed audience, initial buyer count, rounded future-unit mean and follow-up period. None includes later units purchased by initial nonresponders. The small prospective-customer values mainly reflect the much lower initial response in those separate studies; they should not be ranked against Study A as though the audiences had equal prior relationships or exposure. (Said Maybe calculations, Anderson and Simester, Nov 2026).

The three initial response differences are approximately 1.176 percentage points in A, 0.174 in B and 0.171 in C. Approximate independent-binomial 95% intervals are 0.887 to 1.465 points, 0.135 to 0.212 points and 0.127 to 0.215 points, respectively. These intervals apply to initial response only. The tables do not provide the variance needed to put comparable intervals around our later-unit calculations. (Said Maybe calculations, Anderson and Simester, Nov 2026).

Can customer history alone explain the difference?

The raw tables cannot isolate customer history from every other difference between the studies. Established and prospective customers were not randomly assigned to have different histories.

Study A changed prices on 36 products and left another 50 unchanged. Across the changed products, the mean sale price was $133.81 in control and $77.17 in the promotion version. These are catalog-product averages, not the average paid price of every responding buyer. (Anderson and Simester, Feb 2004).

The experiments used different catalogs, changed prices on different sets of products and followed purchasing for different lengths of time. A was an established-customer test; B and C used prospective households. Comparing those studies is useful for identifying a hypothesis about acquisition and retention, but it is not equivalent to a single test that varies customer history while holding everything else fixed.

The setting matters too. These were direct catalog experiments involving durable goods, initiated between January and April 1999. They were not contemporary checkout text-message experiments. Their value for a DTC team is the measurement problem and the observed contrast in a documented setting, rather than a current benchmark for how large a recovery discount should be. (Anderson and Simester, Feb 2004).

A deeper offer can change both behavior and the composition of the buyer group. If a merchant observes different later purchasing among offer users, it needs evidence to distinguish those possibilities. The observed difference alone does not reveal which customers were persuaded, which were already willing to buy, or which shifted a future purchase into the present.

What would make a deeper offer worth keeping?

Evaluate the offer against an unchanged comparison for the eligible audience, using the outcome the business actually needs. Customer averages and first-order conversion can explain part of the result without deciding the whole question.

DecisionOutcome to retainWhat a narrower report leaves out
Did the offer create more initial orders?Completed orders per eligible customer assignedCustomers who ignored the offer
Did it change later demand?Later orders or units for the full assigned audience over a fixed windowLater purchases by initial nonbuyers
Did the offer improve the business result?Contribution after discounts, fulfillment, refunds and relevant costsThe cost of acquiring the extra order
Did it change the buyer mix?Initial response and subsequent outcomes shown separatelySelection into the buyer cohort

This table is a proposed evaluation structure. The published unit means do not provide all of these outcomes. Multiplying units by a headline product price would not recover margin, and a higher response rate would not establish a profitable offer.

For an abandoned-checkout program, keep the actual offer rules with the eligibility record. Resolve the shopper's pricing question using the current approved offer, while the experiment compares the consequences of offering it. A price objection is not permission to invent a discount, and an accepted discount is not proof that the order would otherwise have been lost.

The distinction between promotion eligibility and later behavior is examined in shopping cart coupon statistics. The difference between an observed paid price and a verified discount is covered in ecommerce price history statistics.

How large could the missing nonresponder outcome be?

For Study A, the published responder tables imply approximately 65.284 more later units per 1,000 originally mailed from the initial-buyer cohorts. The full-audience difference also needs later purchases by people who did not buy from the first catalog. Let uC and uP be average later units among those initial nonbuyers under control and promotion. The full-audience difference per person is approximately 0.065284 + 0.968089×uP − 0.979845×uC. The coefficients are the original arms' nonresponse shares. (Anderson and Simester, original Table 3 reconstruction, source checked Nov 2026).

That equation identifies exactly what the responder-only calculation leaves out. If the control nonresponders averaged 0.20 later units, promotion nonresponders would need about 0.135 later units each for the full-audience difference to break even. These are worked values of unreported quantities, not observed purchase rates or recovered records.

Why can the missing outcomes change the sign?

With a worked 0.20-unit average for control nonresponders and zero for promotion nonresponders, the implied full-audience difference is approximately -130.69 later units per 1,000 mailed. If both nonresponder groups instead averaged 0.20, the difference is approximately 62.93 units per 1,000. The same published initial-responder results are consistent with both examples because those results do not supply the missing group means.

Nonnegative future units alone do not establish the direction: either arm's unreported nonresponder mean could be sufficiently larger. An upper bound or an explicit equality assumption would narrow the possibilities, but neither can be silently supplied. Assuming both missing groups have zero future purchases is one strong completion assumption, not a neutral consequence of their initial nonresponse.

Does equal later purchasing among nonresponders solve the problem?

Under the separate worked assumption uC=uP=u, the positive responder-cohort gap is reduced by the larger promotion response share. The break-even common mean is about 5.553 later units per initial nonresponder over the 28-month window. This is an algebraic threshold, not evidence that the nonresponders' means were equal or close to that amount. The groups are selected by a response the promotion can change.

The scope remains the published responder tables. The paper's wider long-run analysis is not overturned by this sensitivity, and its other studies have different audiences and follow-up windows. The missing-mean equation simply prevents the 219.87 versus 154.59 responder-chain values from being mislabeled a complete future-demand effect.

For a current offer test, retain later orders for every originally assigned eligible customer, including those who ignore the first offer. Keep the observation window, identity join and treatment of missing records explicit. The nonresponder sensitivity frontier and calculation make the remaining inputs visible without inventing a profit or lifetime-value result.

Methodology and Sources

We transcribed the mailed counts, initial buyer counts and future-unit means from Tables 3, 5 and 7 of Anderson and Simester's 2004 paper, then checked the rendered source pages. The source experiments began between January and April 1999. Study A followed initial purchasers for twenty-eight months, B for twenty-four and C for twenty-two. We calculated initial response directly from counts, the relative difference between published future-unit means, and initial responders' later units per 1,000 mailed. The future means are rounded, so derived totals and rates are approximate. We did not pool the studies, estimate missing purchases or invent confidence intervals for means whose variance is unavailable.

Study B has a small source inconsistency. The prose on journal page nine assigns 148,703 people to promotion and 148,702 to control, while Table 5 assigns those numbers the other way around. We use the visually checked table and preserve the discrepancy rather than silently repair it. Publication was February 2004. Citations naming Said Maybe calculations refer to the original calculations made in November 2026, not a claim that the experiment is recent.

Glossary

  • Initial responder: A customer who purchased from the test catalog.
  • Conditional mean: An average calculated only among a selected group, here the initial buyers.
  • Original audience: Everyone mailed the assigned catalog version.
  • Responder contribution: Later units from initial buyers, divided by the original audience size.
  • Nonresponder: Someone who did not buy from the initial test catalog, regardless of whether they bought later.
  • Selection: A change in who enters the measured buyer group.
  • Forward buying: Purchasing sooner or buying more now instead of during a later period.
  • Contribution: Revenue after the costs relevant to the decision, which the unit tables do not reconstruct.

Sources

Cite this page

Use this citation when referring to the analysis. Retain the original source population and period with any quoted finding.

Said Maybe. "Discount buyer statistics and the repeat purchase denominator." Said Maybe, November 2026. https://saidmaybe.com/guides/discount-buyer-repeat-purchase-statistics/
<a href="https://saidmaybe.com/guides/discount-buyer-repeat-purchase-statistics/">Discount buyer statistics and the repeat purchase denominator</a> (Said Maybe, November 2026)