AI Glossary / Statistics and probability / Advanced

Biserial correlation

Definition

Biserial correlation estimates the association between a continuous variable and an underlying continuous variable that has been observed only as two categories. The usual model treats the two-category observation as a thresholded normally distributed variable. It differs from point-biserial correlation, which is simply the Pearson correlation between the measured continuous variable and the observed binary indicator.

What it means

Suppose a system retains only whether a continuous score exceeded a cutoff. Biserial correlation addresses the relationship with the score behind that indicator, under a specified latent-variable model. Point-biserial correlation addresses the relationship with the indicator itself.
The normality assumption concerns an underlying variable, not a claim that a zero-one column has a normal distribution. A naturally binary category does not automatically justify imagining a normally distributed quantity behind it.

How it works

Let D equal one above a threshold on a latent standard-normal variable Z. Let p be the proportion with D = 1, q = 1-p, and t the normal threshold satisfying P(Z > t) = p. Write phi(t) for the standard-normal density at that threshold.
The usual moment correction is r_b = r_pb sqrt(pq)/phi(t), where r_pb is the observed point-biserial correlation. Under a joint normal model for the continuous measurement and Z, the population version recovers their latent correlation. The correction is model-based; it does not reconstruct individual missing scores.
An equivalent sample calculation uses the difference between the two group means divided by the overall standard deviation, multiplied by pq/phi(t), when that standard deviation uses divisor n. Mixing formulas with different standard-deviation conventions can introduce a finite-sample discrepancy.

Three practical examples

• Score analysis: an analyst studies a numeric measurement alongside an indicator created by thresholding another score and distinguishes observed from latent association.
• Assessment research: researchers examine whether a continuous total measure relates to an assumed continuous tendency behind a binary response, stating the model explicitly.
• Data-retention review: a team compares the information available from original numeric values with what remains after retaining only above-cutoff indicators.

Think of it this way

Imagine seeing a lamp that turns on when a hidden dial crosses a line. Point-biserial correlation concerns the lamp's on-off state. Biserial correlation attempts to infer association with the hidden dial, using assumptions about how its values are distributed.

Worked example

Consider an original population model. Let Z and E be independent standard-normal variables, and define a continuous measurement X = 0.6Z + 0.8E. Because 0.6 squared plus 0.8 squared equals one, X has variance one. Its correlation with Z is 0.6. Suppose only D = 1 when Z > 0 is retained, so p = q = 0.5.
The density at the threshold is phi(0) = 1/sqrt(2 pi), approximately 0.3989423. Under this model, the observed point-biserial correlation is 0.6 times phi(0)/sqrt(0.5 times 0.5), approximately 0.4787307. Thresholding has changed the population correlation from 0.6 to about 0.479.
Applying the biserial correction gives 0.4787307 times 0.5/0.3989423, approximately 0.6. The correction works here because the data-generating model satisfies the assumptions used to derive it.
A separate tiny sample illustrates the arithmetic using observed values. Suppose the D = 0 group has X values 0 and 2, and the D = 1 group has values 2 and 4. Each group contains two observations, so p = q = 0.5. Their means are 1 and 3, giving a difference of 2. The overall mean is 2.
The four squared deviations from that overall mean are 4, 0, 0 and 4. With divisor n = 4, the overall variance is 2 and the standard deviation is sqrt(2). The point-biserial correlation is therefore [2/sqrt(2)] times sqrt(0.25) = 1/sqrt(2), approximately 0.7071068.
The biserial estimate is [2/sqrt(2)] times 0.25/0.3989423, approximately 0.8862269. Equivalently, multiply 0.7071068 by the correction factor 0.5/0.3989423, approximately 1.2533141. Both paths give the same estimate. This four-observation calculation demonstrates the formula; such a small dataset provides little basis for assessing the latent model or precision.
If the binary coding is reversed, the group mean difference changes from positive two to negative two. The point-biserial and biserial estimates become approximately -0.7071068 and -0.8862269. The strength has not changed, but the category designated as one has.
Finally, suppose another sample with an equal split has observed point-biserial correlation 0.9. Applying the same moment correction gives approximately 1.1279827, outside the correlation range. Such a result is not a correlation stronger than perfect association. It signals that this unconstrained moment estimate is inadmissible, potentially because of sampling variation or unsuitable assumptions. It should prompt investigation rather than automatic interpretation as a valid latent correlation.

Illustrative case study

Fictional company Willow Metrics retains a continuous performance measure and a thresholded version of a separate score. Analysts first establish how the threshold was defined and which category is coded one. They calculate the observed association and separately assess whether a latent normal model is defensible.
Their report labels the two coefficients clearly and explains that the larger corrected value answers a different question. If original scores are available, they inspect them directly instead of relying solely on the inferred relationship.

Common misconceptions and mistakes

Biserial, point-biserial and rank-biserial correlations are different measures. Similar names do not make their assumptions or interpretations interchangeable.
A larger corrected magnitude is not automatically more trustworthy. The model can be inappropriate, the estimate unstable or an unconstrained calculation outside the allowable range. When one category is absent, the required calculation is undefined. A further mistake is treating association with a latent variable as evidence that changing a threshold or an underlying trait will cause an outcome to improve.

Business use

The measure can clarify analyses involving thresholded scores, particularly when teams need to distinguish associations with recorded labels from associations with assumed underlying quantities. Reports should identify the coding, threshold model, estimation method and uncertainty. Retaining original continuous information, when available and appropriate, permits a more direct analysis.

People also ask

Is this ordinary Pearson correlation on zero-one data?

That is point-biserial correlation.

Does the correction recover missing individual values?

No; it estimates an association under assumptions.

Is a value above one valid?

No; an unconstrained estimate outside the range needs investigation.

Check your understanding

Why did the population correction return 0.6?

Answer: The example satisfied the joint normal threshold model.

Why did reversing the binary code reverse the sign?

Answer: It swapped the group means in the difference.

Why was the small sample not model validation?

Answer: Correct arithmetic does not establish the distributional assumptions.

Related terms

• Point biserial correlation ( r pb )
• Tetrachoric correlation
• Latent variable
• Normal distribution
• Point-biserial correlation
• Polyserial correlation
• Dichotomization
• Pearson correlation

Learn next

Compare observed and corrected correlations at different thresholds under the same simulated joint normal model, tracking category balance and estimation variability.

More in Statistics and probability

• Binomial test - A binomial test assesses a specified probability of success using a count of successes in a fixed number of independent binary trials with a common success...
• Birnbaum's theorem - Birnbaum's theorem is a foundational statistical result linking the sufficiency principle, the conditionality principle and the likelihood principle.
• Bisecting K-Means - Bisecting K-Means is a divisive clustering method that repeatedly splits one existing cluster into two using K-Means until the requested number of clusters...
• Bispectrum - The bispectrum is a third-order frequency-domain description of a signal or random process.
• Bivariate analysis - Bivariate analysis examines two variables together to describe or investigate their relationship.
• Bivariate normal model - A bivariate normal model describes two random variables jointly using a two-dimensional normal distribution.

Sources

David Howell, Statistical Methods for Psychology glossary: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html . William Revelle, psych package documentation, biserial and polyserial correlations: https://search.r-project.org/CRAN/refmans/psych/html/tetrachor.html . The population model, sample calculations and company scenario are original illustrations.

Previous: Bisecting K-Means Next: Bispectrum