Can the binary codes be 1 and 2 instead of 0 and 1?

Yes. An increasing linear recoding preserves the correlation, while reversing the order changes its sign.

Is normality required to calculate it?

No. Distributional assumptions concern particular inference procedures, not the arithmetic definition.

Can it be calculated if everyone is in one group?

No. The binary variable then has zero variance, so correlation is undefined.

AI Glossary / Statistics and probability / Intermediate

Point biserial correlation ( r pb )

Definition

Point biserial correlation, commonly written r_pb, measures the association between a binary variable and a numerical variable. It is the ordinary Pearson correlation calculated when the binary variable is coded with two distinct values, conventionally 0 and 1. The sign indicates which coded group has the higher numerical mean, and the magnitude depends on the mean separation, within-group variation and group proportions. It is a descriptive association measure; it does not establish that group membership causes the difference.

What it means

The statistic connects a comparison of two group means with the familiar correlation scale from -1 to +1. If observations coded 1 tend to have larger numerical values than observations coded 0, the correlation is positive. Reversing those codes reverses the sign without changing the strength of association. The same mean difference can yield different correlations when the overall spread or group balance changes. Consequently, comparisons across datasets require attention to how groups were sampled and how the continuous measure was defined, not just the reported coefficient.

How it works

Pair each binary indicator with its numerical measurement and calculate Pearson correlation, or use the equivalent group-mean formula. With population-style empirical standard deviation s_N and group proportions p and q, the formula is (mean_1 - mean_0) times sqrt(pq), divided by s_N. If the standard deviation instead uses denominator N minus one, a corresponding adjustment is needed. Both groups must occur, and the numerical variable must vary. For confidence intervals or hypothesis tests, use methods appropriate to independence, sample size and distributional conditions. Computing the descriptive coefficient itself does not require the observations to be normally distributed.

Three practical examples

Education analysis: summarize the association between a yes/no course-completion indicator and a numerical assessment score.
Service operations: compare measured resolution times between cases with and without a specified process feature.
AI diagnostics: examine the association between a binary error indicator and a continuous case characteristic, while investigating confounding and subgroup composition before interpreting the relationship.

Think of it this way

Think of placing two labeled sets of measurements on the same number line. Point biserial correlation summarizes how well the labels align with higher or lower positions relative to the overall spread. Overlapping groups can still have a nonzero correlation, and a clear group separation does not reveal why the groups differ.

Worked example

Take an original artificial dataset with group 0 values 2, 4 and 6, and group 1 values 6, 8 and 10. The group means are 4 and 8; the overall mean is 6. Squared deviations from the overall mean sum to 40, so the empirical variance with denominator 6 is 40/6. Both group proportions equal one half. Therefore r_pb = (8-4) times sqrt(0.5 times 0.5), divided by sqrt(40/6), which equals approximately 0.7746. Direct Pearson correlation between the code sequence 0,0,0,1,1,1 and the six values gives the same result. Reversing the codes gives -0.7746. Under the usual pooled two-group test assumptions, the associated t statistic is r times sqrt[(N-2)/(1-r^2)] = sqrt(6), approximately 2.4495, with four degrees of freedom. That test relationship does not remove the assumptions needed for inferential interpretation.

Illustrative case study

In a fictional case, Cedar Learning compares assessment scores for participants who completed a voluntary workshop and those who did not. The point biserial correlation is positive, but participants chose whether to attend. Analysts explain that prior motivation and experience may differ between groups. They report group sizes, means and distributions alongside the correlation and avoid calling it the causal effect of the workshop. A later randomized study would address a different and stronger causal question.

Common misconceptions and mistakes

Point biserial correlation is not a special coefficient that generally disagrees with Pearson correlation on the same binary coding and observations. It differs from biserial correlation, which involves assumptions about a latent continuous variable underlying a dichotomy. Arbitrarily splitting a continuous variable into high and low categories can discard information and change the relationship being measured. A large sample can make a small association statistically significant without making it practically important. Likewise, the coefficient's sign depends on coding, and a zero correlation means equal group means under the formula, not that the full group distributions are identical.

Business use

For business reporting, show the original-unit mean difference and group sizes as well as r_pb. This makes it easier to judge whether the association is operationally meaningful. Check whether group proportions reflect the target population or deliberate sampling; a case-control-style sample can change the coefficient's interpretation. For predictive use, validate on the deployment population and avoid selecting features solely because their individual correlations look strong. For explanations, investigate selection and confounding before treating the binary attribute as an intervention that would change the outcome.

Check your understanding

Reversing the 0 and 1 labels changes r_pb from 0.4 to what value?

Answer: -0.4. The group comparison is the same, but its direction relative to the coding reverses.

Related terms

Biserial correlation
Binary variable
Confounding
Pearson correlation
mean difference
two-sample t-test

Learn next

Learn how correlation relates to a two-group comparison, then examine how unequal group sizes, sampling design and artificial dichotomization affect interpretation.

Sampling design

Sources

SciPy, pointbiserialr documentation, formula and relationship to Pearson correlation: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.pointbiserialr.html

Previous: Pod slurping Next: Point estimate