A point biserial correlation measures the strength of the relationship between two variables when one is continuous (like a test score or blood pressure reading) and the other is binary, meaning it has only two possible values (like yes/no, male/female, or treated/untreated). It produces a value between -1 and +1, just like a standard Pearson correlation, where values closer to the extremes indicate a stronger relationship and values near zero indicate a weak or nonexistent one.
How It Differs From a Standard Correlation
A standard Pearson correlation measures the linear relationship between two continuous variables, like height and weight. The point biserial correlation is actually a special case of the Pearson correlation, adapted for situations where one variable is dichotomous. If you coded your binary variable as 0 and 1 and ran a regular Pearson correlation, you’d get the exact same result as a point biserial calculation. The dedicated formula simply makes the logic behind the math more transparent when one variable is binary.
This distinction matters because researchers encounter binary variables constantly. Did the patient receive the treatment or not? Did the student pass or fail? Is the participant in the experimental group or the control group? Whenever you want to know whether that binary grouping is associated with differences in a measured outcome, the point biserial correlation is the appropriate tool.
What the Formula Does
The calculation compares the average score on the continuous variable for each of the two groups defined by the binary variable. Specifically, it takes the difference between those two group means, multiplies it by a factor based on how evenly the groups are split, and divides by the overall spread (standard deviation) of the continuous variable.
In plainer terms: the correlation will be strong when the two groups have very different average scores and the scores within each group are relatively tightly clustered. It will be weak when the group averages are similar or when there’s so much variation within each group that the group difference gets lost in the noise. The proportion of cases in each group also plays a role. Extremely unbalanced groups (say, 95% in one category and 5% in the other) will push the maximum achievable correlation below 1.0, even if the relationship is strong.
How to Interpret the Result
The output is a correlation coefficient, typically written as rpb, ranging from -1 to +1. The sign tells you the direction: a positive value means the group coded as 1 has higher scores on the continuous variable, while a negative value means the group coded as 0 has higher scores. The magnitude tells you the strength of the association.
Cohen’s widely used guidelines suggest that r = ±0.10 represents a small effect, r = ±0.30 a medium effect, and r = ±0.50 a large effect. These are rough benchmarks, not rigid cutoffs. Context matters enormously. A correlation of 0.10 between a surgical procedure and a serious complication could be highly meaningful in clinical terms, even though the number looks small by conventional standards. Always consider what a given effect size means in practical terms for the specific situation you’re examining.
Testing Whether the Result Is Statistically Significant
A point biserial correlation of, say, 0.25 in your data doesn’t automatically mean a real relationship exists in the broader population. It could be a fluke of your particular sample. To check, you can convert the correlation to a t-statistic using a straightforward formula and then look up the associated p-value with degrees of freedom equal to your total sample size minus two.
This is mathematically equivalent to running an independent-samples t-test comparing the means of your two groups. If the t-test is significant, the point biserial correlation is significant, and vice versa. They’re two ways of expressing the same underlying comparison. The t-test tells you whether the group means differ; the point biserial correlation tells you how strongly group membership is associated with the continuous outcome. The correlation is essentially the effect size version of the t-test.
You can also construct a confidence interval around the correlation. If the interval doesn’t include zero, you can conclude the relationship is statistically significant at your chosen threshold.
Real-World Examples
In education, the point biserial correlation is one of the most common tools for evaluating test items. If you want to know whether a particular exam question distinguishes between students who know the material and those who don’t, you correlate whether each student got that question right (binary: correct or incorrect) with their total exam score (continuous). A high positive correlation means the question is doing its job. A near-zero or negative correlation suggests the question is confusing, poorly worded, or testing something unrelated to the rest of the exam.
In medical research, you might correlate a binary variable like whether a patient has a specific condition (present or absent) with a continuous health outcome score. One study found a statistically significant positive correlation between sex (coded as a binary variable for women) and a global health maintenance score, helping researchers quantify how strongly that demographic factor was associated with health behaviors.
Other common applications include correlating treatment group assignment with recovery time, employment status with a psychological well-being score, or the presence of a genetic marker with a biomarker level.
Assumptions and Limitations
The point biserial correlation assumes the continuous variable is roughly normally distributed within each of the two groups. It also assumes the binary variable represents a true dichotomy, meaning the two categories are naturally distinct rather than an artificial split of an underlying continuous trait. If you’ve taken a continuous variable like age and split it into “young” and “old” at some cutoff, a different technique (the biserial correlation) is more appropriate, because chopping a continuous variable into two bins throws away information and distorts the result.
Like all correlations, the point biserial measures linear association and doesn’t establish causation. A strong correlation between group membership and an outcome tells you the two are related but not why. Outliers in the continuous variable can also inflate or deflate the result, so checking your data for extreme values before interpreting the coefficient is worth the effort.
How to Calculate It in Software
Most statistical software packages make this calculation simple. In Python, the SciPy library has a dedicated function called pointbiserialr in its stats module. You pass in your binary array and your continuous array, and it returns both the correlation coefficient and the p-value.
In R, you can use the cor.test function with your binary variable coded as 0 and 1, since the Pearson correlation on those inputs produces the same result. The ltm package also has a dedicated biserial.cor function. In SPSS, you can run a standard Pearson correlation between the two variables, with the binary variable coded numerically, and the output will be the point biserial correlation. No special procedure is required because, mathematically, it’s the same computation.