This ABA point-biserial correlation calculator summarizes linear association between one prespecified binary grouping variable and one continuous response. It exposes the group coding, counts, means, response spread, group-balance factor and coefficient so another qualified reviewer can reproduce the arithmetic. The result describes the displayed units and cannot establish a treatment effect, experimental control, clinical importance or causation.

Clinicians & ABA Professionals / Data, Outcomes and Clinical Decision-Making.

Decision question and coding key

Write the descriptive question before entering values: "Among these prespecified independent units, how is membership in these two defined groups associated with this continuous response?" Name the unit, the two eligible groups, the response, its units and its direction. State why a binary grouping variable is appropriate and why every row is eligible.

Define the reported difference as Group 1 mean minus Group 0 mean. A positive coefficient then means that the group coded 1 has the higher response mean; a negative coefficient means the reverse. This direction belongs in the record before anyone sees the data.

Current SciPy pointbiserialr documentation describes point-biserial correlation for a binary variable and a continuous variable. It also notes its equivalence to ordinary Pearson correlation when the binary variable is coded numerically. Current SciPy pearsonr documentation supplies that linear-correlation cross-check. NIST scatterplot guidance supports examining the displayed values rather than relying on a coefficient alone. These statistical sources do not make the method an ABA standard or validate a clinical interpretation.

The current BACB Ethics Codes and BCBA Test Content Outline provide professional context for competence, measurement, documentation and data-based decision-making. Neither prescribes point-biserial correlation. The Standards for Educational and Psychological Testing support tying an interpretation to its intended use and available evidence.

Scope record

Complete this specification without looking at the coefficient.

FieldPrespecified entryDecision or descriptive questionIndependent unitGroup 0 definitionGroup 1 definitionWhy exactly two groups are eligibleCoding rationale and planned sign interpretationContinuous response definitionResponse units and directionMeasurement windowInclusion ruleExclusion ruleMissing-data ruleData version and extraction timeIntended descriptive useRequired clinical and methods reviewers

Do not collapse three or more categories into two after seeing results. Do not turn an ordered category into a binary label merely to obtain this coefficient. If the grouping variable reflects a phase, condition or treatment assignment, the coefficient still does not recover the logic of the design.

Row-level input table

Preserve the original group label and the fixed numeric code. The calculator uses 0 for Group 0 and 1 for Group 1. The coding key determines the sign.

Row keyOriginal group labelFixed code, 0 or 1Continuous responseResponse unitsIncluded under rule?Missingness or context note12345Add rows

Use one row per independent unit. Repeated observations from the same client, clinician, classroom or site are not made independent by placing them on separate rows. Stop for a clustered or repeated design unless a qualified methods reviewer specifies an analysis that represents that dependence.

Formula and audit trail

Let n0 and n1 be the two group counts, N = n0 + n1, ybar0 and ybar1 the group means, ybar the overall response mean, and sy the sample standard deviation across all included responses.

SS_y = sum((Y_i - y_bar)^2)
s_y = sqrt(SS_y / (N - 1))
mean difference = y_bar_1 - y_bar_0
balance factor = sqrt((n0 * n1) / (N * (N - 1)))
r_pb = ((y_bar_1 - y_bar_0) / s_y) * balance factor

Calculate from unrounded values. Round only reported displays. The displayed formula uses the sample standard deviation with denominator N - 1; pairing it with the stated balance factor is essential. As a computational audit, calculate ordinary Pearson correlation between the fixed 0/1 code and the same response values. The two coefficients should agree within the stated numerical tolerance. Disagreement means the rows, coding, missingness or standard-deviation convention need review.

The coefficient is undefined when either group is empty or the response has zero variance. A group with one row can produce arithmetic, but it provides fragile descriptive evidence and should trigger review rather than confident interpretation.

Blank calculation record

Use this ABA point-biserial correlation calculator as a reproducible worksheet, not as a score-only output.

QuantityFormula or provenanceFull-precision valueDisplay valuen0Count of fixed code 0n1Count of fixed code 1Nn0 + n1ybar0Mean response in Group 0ybar1Mean response in Group 1y_barMean response across all rowsSS_ysum((Yi - ybar)^2)s_ysqrt(SS_y / (N - 1))Mean differenceybar1 - ybar0Balance factorsqrt((n0n1)/(N(N-1)))r_pbMean-difference formulaPearson cross-checkCorrelation of code and responseAbsolute cross-check differenceabs(rpb - rPearson)

Also retain side-by-side dot plots, strip plots or another direct display of the response within each group. Record range, unusual observations, group imbalance and any provenance question next to the coefficient.

Fictional worked example

The values below are fictional and exist only to show arithmetic. They are not treatment recommendations, performance expectations or norms.

Fictional rowGroup labelCodeResponseAComparison02BComparison04CComparison05DTarget16ETarget19

The group counts are n0 = 3, n1 = 2 and N = 5. The means are ybar0 = 3.6666666667, ybar1 = 7.5 and ybar = 5.2. Across all five responses, SSy = 26.8 and s_y = sqrt(26.8 / 4) = 2.5884358211.

The mean difference is 7.5 - 3.6666666667 = 3.8333333333. The balance factor is sqrt((3 2) / (5 4)) = 0.5477225575. Therefore:

r_pb = (3.8333333333 / 2.5884358211) * 0.5477225575
= 0.8111474582

Pearson correlation between codes [0, 0, 0, 1, 1] and responses [2, 4, 5, 6, 9] is also 0.8111474582. The positive sign means the group coded 1 has the higher mean in this record. The label itself has no intrinsically positive direction.

A side-by-side display would show the five source responses and the small Group 1 count. The coefficient should be reported beside that display and the unstandardized fictional mean difference of 3.8333333333, not by itself.

Removing fictional row E changes the coefficient to 0.6831300511. That comparison reveals sensitivity to one row and the smaller Group 1 count. The sensitivity finding neither identifies a data error nor authorizes deletion.

Coding and transformation checks

CheckExpected resultReview meaningSwap 0 and 1 labels without changing rowsSign reverses; magnitude stays the sameInterpretation must name the coding keyAdd the same constant to every responseCoefficient stays the sameResponse origin does not determine associationMultiply every response by a positive constantCoefficient stays the samePositive unit conversion does not alter rMultiply every response by a negative constantSign reversesResponse direction must be documentedResponse is constantr_pb is undefinedNo response variation is availableEither group is emptyr_pb is undefinedA binary comparison requires both groupsOne group has only one rowArithmetic may be definedInterpretation is fragile and needs explicit reviewRemove each row onceRecord every leave-one-out valueSensitivity is a review prompt, not a deletion rule

These are calculation checks, not permissions for post hoc recoding. A clinician who changes the group labels, response direction, inclusion rule or row set after seeing the preferred result has changed the question.

Interpretation boundaries

Point-biserial correlation ranges from -1 to 1, subject to numerical precision. The sign follows the group code and response direction. The magnitude summarizes standardized linear separation in the displayed response between the two coded groups. No universal ABA threshold labels it weak, moderate, strong, clinically important or actionable.

Do not interpret r_pb as the probability that one group has a higher value, diagnostic accuracy, agreement, reliability or treatment effect. Squaring the coefficient gives the usual squared linear-correlation quantity for the coded variables; it is not the proportion of response variation caused by group membership. Group membership may be related to selection, phase, setting, measurement opportunity, baseline level or another variable. Correlation does not imply causation.

Severe group imbalance, restricted response range and influential observations can make the coefficient unstable or unrepresentative. Inspect each group's distribution and retain the raw values. If the scientific question concerns a difference in proportions, paired change, ordinal groups, repeated measurements, phase effects or covariate adjustment, use a design-appropriate method selected before looking at a preferred result.

Failure modes that stop the worksheet

  • A row's group or response cannot be traced to its source record.
  • The supposed binary variable contains more than two meaningful categories.
  • The coding key or response direction was selected after outcomes were reviewed.
  • Units are repeated or nested, but the worksheet treats them as independent.
  • One group is empty, the response is constant or missingness leaves an unusable record.
  • Group imbalance is so severe that one or two rows dominate the comparison.
  • Exclusions, transformations, weights or imputations were introduced after inspecting the coefficient.
  • The intended claim extends beyond the displayed units or requires causal, inferential or diagnostic-validity evidence.

Document the stop, preserve the source rows and obtain qualified clinical and methods review. Do not keep recalculating until the coefficient looks reassuring.

Privacy, security and accessible review

Use the minimum necessary data and replace direct identifiers with approved row keys. The HHS Privacy Rule summary and HHS Security Rule summary provide federal context; organizational policy and applicable law still govern storage, access, transmission, retention and disclosure.

Provide the coding key and results in text and tables rather than color alone. Label plots, state response units and direction, preserve a machine-readable table, and use accessible headings. A coefficient without the definitions and group displays is not an adequate review artifact.

Reviewer questions

  • Were the grouping variable, response, unit and inclusion rules fixed before calculation?
  • Are the two groups genuinely binary and mutually exclusive for this question?
  • Does every row represent an independent unit, or is dependence being hidden?
  • Do r_pb and the Pearson 0/1 cross-check agree?
  • Is the unstandardized group-mean difference reported beside the coefficient?
  • Do the group plots show imbalance, range restriction or influential rows?
  • Would swapping the group code reverse only the sign and not the substantive interpretation?
  • Are clinical magnitude and client or caregiver priorities discussed apart from the coefficient?
  • Are every named external clinical, statistical, privacy, security, accessibility and legal review still pending where applicable?

Related resources

Sources