ORDEEN · Choice Modelling · 3 MIN READ

What is the Haldane Correction? Fixing Zero-Cell Probabilities

The Haldane correction adds 0.5 to each contingency table cell to stabilise odds-ratio estimates when some cells contain zero observations.

Book a 20-min demo What is the Haldane Correction? Fixing Zero-Cell Probabilities

The Haldane correction (sometimes called the Haldane-Anscombe correction) adds a small constant — classically $0.5$ — to each cell of a contingency table before computing a ratio, log, or odds-based statistic. It is the standard fix for the problem that any cell containing zero observations makes an odds ratio either zero or undefined, and makes its log statistic $\pm \infty$.

Ordeen applies this correction inside its segment-level and interaction tests so empty cells do not produce spurious infinite-ratio effects, and reports every tested interaction with a 95% credible interval and proper multiple-comparison control. See how it fits in the full platform in our methodology.

The problem

In interaction testing between two categorical attributes, each cell of the contingency table records how often a particular pair of levels appeared together in observed choices. For attributes with many levels, or small samples, some cells inevitably contain zero observations.

The naive odds ratio for a $2 \times 2$ sub-table is:

$$\mathrm{OR} = \frac{n_{11} \cdot n_{22}}{n_{12} \cdot n_{21}}$$

If $n_{12}$ or $n_{21}$ is zero, the ratio diverges to infinity. If $n_{11}$ or $n_{22}$ is zero, it collapses to zero. Either way the test statistic becomes nonsensical and would, uncorrected, manufacture a highly “significant” interaction finding from what is actually a sparse-data artefact.

How the correction works

Haldane’s fix adds $0.5$ to every cell of the table before computing the statistic:

$$\mathrm{OR}_{\text{Haldane}} = \frac{(n_{11} + 0.5)(n_{22} + 0.5)}{(n_{12} + 0.5)(n_{21} + 0.5)}$$

This stabilises the estimator and its variance. The $0.5$ is a Bayesian-flavoured choice — equivalent to a Jeffreys prior of $\mathrm{Beta}(0.5, 0.5)$ on each cell probability — but any small positive constant achieves the same stabilising effect; $0.5$ is simply the convention.

Why it matters in conjoint

Interaction grids in CBC and ACBC studies routinely contain sparse cells, especially when prohibited combinations limit the design space or when a level is used sparingly. Running the interaction tests without a small-cell correction, then applying Benjamini-Hochberg on top, will cheerfully “discover” interactions driven entirely by individual zero cells. The correction has to happen before the multiplicity adjustment, not after.

How Ordeen uses it

Ordeen uses the Haldane correction on the interaction results it reports, not to decide which interactions enter the model (that is decided by a data-coverage check). When a level pair has fewer than 20 observations, 0.5 is added to the counts before its test statistic is computed, and only then is Benjamini-Hochberg applied across all the cell tests. Interaction terms estimated inside the Hierarchical Bayes model are judged by whether their 95% credible interval excludes zero. The result is interaction findings that reflect genuine signal rather than sparse-cell arithmetic.

Related terms

Benjamini-Hochberg · Conjoint analysis · Full methodology

The Haldane correction (sometimes called the Haldane-Anscombe correction) adds a small constant — classically $0.5$ — to each cell of a contingency table before computing a ratio, log, or odds-based statistic. It is the standard fix for the problem that any cell containing zero observations makes an odds ratio either zero or undefined, and makes its log statistic $\pm \infty$.

Ordeen applies this correction inside its segment-level and interaction tests so empty cells do not produce spurious infinite-ratio effects, and reports every tested interaction with a 95% credible interval and proper multiple-comparison control. See how it fits in the full platform in our methodology.

The problem

In interaction testing between two categorical attributes, each cell of the contingency table records how often a particular pair of levels appeared together in observed choices. For attributes with many levels, or small samples, some cells inevitably contain zero observations.

The naive odds ratio for a $2 \times 2$ sub-table is:

$$\mathrm{OR} = \frac{n_{11} \cdot n_{22}}{n_{12} \cdot n_{21}}$$

If $n_{12}$ or $n_{21}$ is zero, the ratio diverges to infinity. If $n_{11}$ or $n_{22}$ is zero, it collapses to zero. Either way the test statistic becomes nonsensical and would, uncorrected, manufacture a highly “significant” interaction finding from what is actually a sparse-data artefact.

How the correction works

Haldane’s fix adds $0.5$ to every cell of the table before computing the statistic:

$$\mathrm{OR}_{\text{Haldane}} = \frac{(n_{11} + 0.5)(n_{22} + 0.5)}{(n_{12} + 0.5)(n_{21} + 0.5)}$$

This stabilises the estimator and its variance. The $0.5$ is a Bayesian-flavoured choice — equivalent to a Jeffreys prior of $\mathrm{Beta}(0.5, 0.5)$ on each cell probability — but any small positive constant achieves the same stabilising effect; $0.5$ is simply the convention.

Why it matters in conjoint

Interaction grids in CBC and ACBC studies routinely contain sparse cells, especially when prohibited combinations limit the design space or when a level is used sparingly. Running the interaction tests without a small-cell correction, then applying Benjamini-Hochberg on top, will cheerfully “discover” interactions driven entirely by individual zero cells. The correction has to happen before the multiplicity adjustment, not after.

How Ordeen uses it

Ordeen uses the Haldane correction on the interaction results it reports, not to decide which interactions enter the model (that is decided by a data-coverage check). When a level pair has fewer than 20 observations, 0.5 is added to the counts before its test statistic is computed, and only then is Benjamini-Hochberg applied across all the cell tests. Interaction terms estimated inside the Hierarchical Bayes model are judged by whether their 95% credible interval excludes zero. The result is interaction findings that reflect genuine signal rather than sparse-cell arithmetic.

Related terms

Benjamini-Hochberg · Conjoint analysis · Full methodology

Related guides

Contract Analysis

Jurisdiction-Grounded AI Contract Review & Risk

Why generic chatbots hallucinate legal contract risk, and how Ordeen grounds contract review, portfolio risk scoring and renewal drafting in governing law.

Read the guide
Forecasting

Intermittent Demand Forecasting: Croston, TSB & MinT

How to forecast lumpy, sparse SKU demand and reconcile hierarchies: Croston's method, SBA, TSB, MinT reconciliation and out-of-sample AI signal validation.

Read the guide
Price Monitoring

Retail Checkout Price Monitoring vs Sticker Scrapers

Why basic retail price scrapers miss real checkout undercutting: how delivery thresholds, variant SKU matching and screenshot evidence protect brand RRP.

Read the guide