An ecommerce analysis reports that a redesigned checkout doubles the odds of completion. A slide then announces a 100% conversion lift. Those statements are not equivalent. The gap can be substantial when many shoppers already complete checkout, even if the statistical model itself was fitted correctly.
EcomToolkit’s editorial rule is to translate analytical output into the scale of the decision. A trading team needs the expected number of additional completed purchases, the population involved and the uncertainty around the estimate. This guide uses hypothetical examples to distinguish odds, probabilities and relative changes without presenting an association as proven causation.
Table of Contents
- Start with a count of outcomes
- Compare three ways to describe a change
- Translate an odds ratio using a baseline
- Do not turn adjusted odds into a universal forecast
- Preserve uncertainty and commercial meaning
- Build a reporting handoff that survives editing
- The EcomToolkit view
Start with a count of outcomes
Suppose 100 eligible checkout starters include 20 completed purchases and 80 non-completions. The completion probability is 20 divided by 100, or 20%. The odds are 20 divided by 80, or 0.25. Probability compares successes with all eligible outcomes; odds compare successes with failures.
That difference in denominator is the entire source of the common reporting error. A model or statistical package may use odds because they have useful mathematical properties, but customers and trading teams generally discuss probabilities. Keep both definitions available in the analysis rather than trying to infer them from the word lift.
The UCLA explanation of logistic-regression odds ratios provides the formal link between probability, odds and model coefficients. For a probability p, odds equal p divided by one minus p. Conversely, probability equals odds divided by one plus odds.
Define the denominator before any calculation. Checkout starters, sessions and customers create different conversion measures. An odds ratio for completion among checkout starters cannot be applied directly to all website sessions. Document the eligibility window, repeated visits and treatment of purchases that complete after the reporting cutoff.
Compare three ways to describe a change
Imagine two hypothetical groups with 100 eligible checkout starters each. Group A has 20 completions; group B has 40. The probabilities are 20% and 40%, giving a probability ratio of 2 and an absolute difference of 20 percentage points. The odds are 0.25 and approximately 0.667, giving an odds ratio of approximately 2.67.
| Measure | Group A | Group B | Comparison |
|---|---|---|---|
| Completed purchases | 20 of 100 | 40 of 100 | 20 additional completions |
| Completion probability | 20% | 40% | 20 percentage points higher |
| Probability ratio | Baseline | 0.40 / 0.20 | 2.00 |
| Odds | 20 / 80 = 0.25 | 40 / 60 = 0.667 | About 2.67 times the odds |
These are different summaries of the same fixture, not conflicting results. The error appears when the 2.67 odds ratio is described as 167% higher conversion probability. That wording exaggerates the relative probability change, which is 100% in this example.
Use percentage points when describing the absolute change and percent when describing the relative change. A chart annotation reading simply plus 20% is ambiguous here. Prefer a full sentence: completion increased from 20% to 40%, an absolute increase of 20 percentage points in this illustrative dataset.
Translate an odds ratio using a baseline
An odds ratio alone does not identify the new probability. If p0 is the baseline probability and OR is an odds multiplier applicable to that comparison, the transformed probability is p1 = OR × p0 / (1 − p0 + OR × p0). The baseline is therefore a required input, not an optional footnote.
For OR = 2 and p0 = 0.20, the result is 0.40 / 1.20, or about 33.3%. The absolute change is about 13.3 percentage points and the relative probability increase is about 66.7%. It is not a move from 20% to 40%.
The same odds multiplier gives different absolute and relative probability changes at other baselines. A low baseline can make odds and probability ratios numerically close, but closeness should be checked rather than assumed. In checkout-stage reporting, completion can be common enough for the difference to affect a business case materially.
| Baseline probability | Probability after doubling odds | Absolute change | Relative probability increase |
|---|---|---|---|
| 1% | About 1.98% | About 0.98 points | About 98.0% |
| 20% | About 33.33% | About 13.33 points | About 66.7% |
| 50% | About 66.67% | About 16.67 points | About 33.3% |
| 80% | About 88.89% | About 8.89 points | About 11.1% |
All table entries are arithmetic illustrations. They do not forecast what a checkout change will deliver. A measured association, a hypothetical multiplier and a causal treatment effect are different forms of evidence, even when the formulas used to display them look similar.
Do not turn adjusted odds into a universal forecast
A logistic-regression coefficient often describes a conditional odds ratio while other model variables are held fixed. Applying that coefficient to an overall store conversion rate can mix incompatible populations. Device, acquisition source and customer history may differ between the baseline summary and the model comparison.
Ask the analyst for predicted probabilities under clearly defined scenarios. If an overall effect is required, calculate predictions for the intended target population and average them using an agreed method. Keep the model, covariate distribution and weighting policy with the result so another reviewer can reproduce it.
Do not interpret a change between adjusted and unadjusted odds ratios as proof of one specific bias. Odds ratios have statistical properties that complicate that comparison. The research discussion in A Most Odd Ratio explains why careful interpretation is necessary. Specialist review is appropriate when the decision depends on an adjusted model.
A regression on observational storefront data does not establish that switching platforms or displaying a badge caused the outcome. Customers who encounter a feature may already differ from those who do not. Keep causal language tied to the study design, and explain unresolved selection effects in ordinary business terms.
Preserve uncertainty and commercial meaning
A transformed point estimate still needs uncertainty. Use an interval appropriate to the original estimation procedure and the target probability contrast. If the baseline is also estimated, treating it as fixed can understate uncertainty. Ask for the full calculation rather than converting only the most attractive number in a model output.
For a randomised checkout experiment, report group sizes, completions, absolute probabilities and a suitable interval for the difference or ratio. Our guide to conversion-rate confidence intervals covers the broader reporting discipline. Match the analysis unit to assignment and account for repeated customer observations where relevant.
Translate a defensible absolute change into an illustrative order count only after fixing the eligible population. A two-percentage-point improvement among 10,000 comparable checkout starters corresponds arithmetically to 200 additional completions. It does not mean 200 additional orders across a different traffic mix or a future month with unknown demand.
Then evaluate margin, refunds, payment failures and implementation costs separately. More checkout completions need not produce proportionally more contribution. Avoid multiplying a conversion claim by headline average order value and presenting the resulting revenue as secured financial value.
Build a reporting handoff that survives editing
Write the metric name beside the number in the chart itself. Put absolute group probabilities close to any odds ratio, and distinguish observed counts from model predictions. This reduces the risk of a correct analysis becoming an incorrect executive summary after several rounds of editing.
Have a reviewer reconstruct the headline from the displayed counts. For the 20-versus-40 example, they should recover both the probability ratio of 2 and the odds ratio of approximately 2.67. A disagreement is a useful signal to stop the publication and inspect denominators or wording.
Keep a short prohibited-translation note in the reporting template: odds must not be relabelled probability without a baseline and a valid transformation. That is more useful than banning odds ratios altogether, because the statistic may be entirely appropriate for the model that produced it.
When comparing segments, retain the same completion definition and observation period. A high-intent returning-customer cohort can show a very different absolute effect from first-time mobile visitors. Present that difference explicitly rather than implying that one headline multiplier describes every shopper equally well.
The EcomToolkit view
A conversion claim should remain understandable after the statistical terminology is removed. State what happened to the probability, how many eligible shoppers were observed, and what the evidence supports. Keep odds ratios available for technical review without asking a commercial audience to guess their meaning.
Bring the model output and proposed headline to a focused EcomToolkit measurement review. The priority is to align the business claim with the actual estimand before it becomes a forecast, budget request or platform recommendation.
