Statistics Glossary → Haldane-Anscombe (zero-cell) correction
Haldane-Anscombe (zero-cell) correctionconcept
Adding a small constant (usually 0.5) to every cell of a contingency table that contains a zero count, so that the odds ratio and its logarithm can be computed.
In a 2 x 2 table the odds ratio is (a·d) / (b·c). If one cell is zero, the ratio becomes 0 or infinite and its logarithm and standard error are undefined. The Haldane-Anscombe correction adds 0.5 to each of the four cells, giving a finite odds ratio, log odds ratio and confidence interval. The same idea is used in log-linear models, where a small delta is added to every cell to stop parameters drifting to infinity because of zero cells.
The correction is a practical fix rather than an exact method: the added value pulls the estimate somewhat towards 1 (0 on the log scale), and the result can be sensitive to the constant chosen. For sparse tables, conditional maximum likelihood or exact methods may be more reliable, and in meta-analysis methods such as Mantel-Haenszel that cope with zero-cell studies. Some implementations add 0.5 only when a zero cell is present; the rule used should be reported.
An everyday example
In a small clinical trial, none of 15 patients on the new drug has a side effect, compared with 4 of 15 on placebo. Because of the zero cell, the odds ratio is 0 and no confidence interval can be formed. Adding 0.5 to each cell gives 0.5 / 15.5 and 4.5 / 11.5, an odds ratio of about 0.08, and a confidence interval can now be computed as well.
In MerQur
In the Log-Linear Model, the 'Cell delta' option adds a constant (delta) to the cell counts to avoid boundary estimates in zero cells; it is applied only in the saturated model. In MR x MR Crosstab, 'Additional tests: Cell association measures' computes the odds ratio and Woolf interval for a zero cell by adding 0.5 to the cells (Haldane-Anscombe), and Fisher's Exact Test also gives a Haldane CI among its confidence-level outputs.
Analyses where this concept appears
Fisher's Exact TestLog-Linear ModelMR × MR Crosstab