This ABA Pearson correlation calculator measures linear association between two complete variables recorded on the same prespecified units. It keeps every pair, exposes the centered cross-products and denominator, and pairs the coefficient with a scatterplot and leave-one-out sensitivity check. The result describes the displayed data; it does not establish agreement, treatment effect, experimental control, or causation.

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

Start with a linear association question

Official NIST correlation documentation defines the product-moment coefficient from centered sums for two equal-length variables. Current SciPy pearsonr documentation likewise describes Pearson r as a measure of linear relationship and identifies constant and nearly constant inputs as calculation boundaries. NIST's scatterplot guidance shows why the graph must be examined for nonlinearity, changing spread, and unusual observations. These statistical references do not make Pearson r an ABA standard or validate a clinical interpretation.

Use the worksheet only when the planned question is linear: across the specified units, do higher values of X tend to accompany proportionally higher or lower values of Y? The variables need not share units, but their definitions, direction, pairing, and collection conditions must be defensible. A coefficient close to zero means little linear association in this set of pairs. It does not mean that no structured relationship exists.

Sequential measures from one person, learners nested within a clinician, repeated ratings, or observations linked by time are not made independent merely because they can be placed in two columns. Preserve those structures and obtain methods review instead of presenting a simple paired coefficient as though each row were independent.

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

Freeze the design before seeing r

Record the analysis plan before calculating a preferred answer.

Design fieldPrespecified entryObservable questionIndependent unit of analysisVariable X definition, units, and directionVariable Y definition, units, and directionPairing keyWhy a linear summary is relevantInclusion and exclusion rulesMissing-pair ruleStop; complete pairs requiredTransformation ruleNone unless justified before reviewData version and extraction timeIntended descriptive useScatterplot review planQualified reviewers

Direction matters. Higher scores can reflect improvement for accuracy and independence, while lower scores can reflect improvement for latency or errors. Write the direction of each variable in words. Reversing an axis after seeing the sign is a new analysis, not a correction.

Preserve the paired observations

Retain original precision and an auditable pairing key.

Pairing keyOriginal XOriginal YCollection contextSource/version noteUnit 1Unit 2Unit 3...

Do not silently delete an unusual point, substitute a group mean, or break a pair. A suspected entry error should be resolved against its source and documented. A valid extreme observation may reveal an important context, limited range, or a model that does not fit; its influence is information to inspect.

Plot Y against X before calculating r. Mark meaningful subgroups, phases, settings, and collection changes. A curve, separate clusters, a funnel-shaped spread, or one distant X value can make one linear coefficient incomplete or misleading.

Calculate the centered components

For n complete pairs, calculate the two means:

xbar = sum(xi)/n

ybar = sum(yi)/n

For every row preserve dxi=xi-xbar, dyi=yi-ybar, dxi*dyi, dxi^2, and dyi^2.

UnitXYdxdydx*dydx^2dy^2123...

Then sum the centered columns:

Sxx = sum(dx_i^2)

Syy = sum(dy_i^2)

Sxy = sum(dxi*dyi)

r = Sxy / sqrt(Sxx*Syy)

Use unrounded quantities through the final division. The centered X and Y deviations should each sum to zero apart from ordinary floating-point tolerance. If Sxx=0 or Syy=0, one variable is constant and r is undefined. A nearly constant input also deserves a stop because subtracting almost equal numbers can make the result numerically unstable.

Fictional five-pair calculation

Consider fabricated values X=[1,2,3,4,8] and Y=[2,3,5,7,6]. They are arithmetic fixtures, not client, caregiver, clinician, assessment, or practice data. The means are xbar=3.6 and ybar=4.6.

UnitXYdxdydx*dydx^2dy^2112-2.6-2.66.766.766.76223-1.6-1.62.562.562.56335-0.60.4-0.240.360.164470.42.40.960.165.765864.41.46.1619.361.96

The audit totals are Sxy=16.2, Sxx=29.2, and Syy=17.2. Therefore:

r = 16.2 / sqrt(29.2*17.2) = 0.7228686085

The coefficient is positive because larger X values generally accompany larger Y values, but the fifth pair has a distant X value and does not continue the pattern seen in the first four rows. The scatterplot makes that feature visible.

As a declared influence check, recalculate after withholding each row one at a time while preserving the full result as primary. With only the first four fabricated pairs, r=0.9897782666. The difference from 0.7228686085 shows sensitivity to the fifth pair. It does not authorize deleting that pair. Review provenance, context, range, and whether a single line is a suitable description.

Transformation, symmetry, and sign checks

Adding any constant to all X values or all Y values leaves r unchanged. Multiplying an entire variable by a positive constant also leaves r unchanged. Multiplying exactly one variable by a negative constant reverses the sign and preserves the magnitude. Swapping X and Y leaves r unchanged.

These are arithmetic checks, not licenses for post hoc recoding. A nonlinear monotonic transformation can change Pearson r because spacing matters. Spearman rho is based on ranks and answers a different question about monotonic ordering.

An exact positive straight line produces r=1; an exact negative straight line produces r=-1. Neither value proves agreement. For example, Y can equal 100+10X and still have perfect positive correlation while differing substantially in level and scale. Lin concordance correlation and Bland-Altman analysis address different agreement questions.

Nonlinearity and restricted range

A strong U-shaped or inverted-U pattern can yield r near zero even though Y changes systematically with X. Inspect the raw scatterplot and do not translate “little linear association” into “no relationship.” Similarly, a coefficient calculated in a narrow X range may differ greatly from one calculated across a broader range. Range restriction is a property of the sampled pairs, not a reason to extrapolate.

Pooling distinct phases, settings, or populations can create, mask, or reverse a coefficient. If the groups matter to the intended interpretation, retain the labels and obtain a design-aware analysis. Do not search many subgroup splits and report only the most favorable r.

Influence review without automatic deletion

For each leave-one-out calculation, record the omitted row, recalculated r, and difference from the full-data coefficient. Large changes identify a pair that strongly affects this summary. They do not identify a data error or an observation that should be removed.

Sensitivity row omittedRecalculated rDifference from full rProvenance/context reviewUnit 1Unit 2...

Any exclusion requires a rule that is defensible independently of the desired result. Report the complete-data coefficient, the prespecified sensitivity analysis, and the reason for any separately approved exclusion.

Interpret only the displayed linear association

Use the ABA Pearson correlation calculator as an inspectable record. Report variable definitions and directions, independent unit, pairing key, complete-pair count, means, Sxy, Sxx, Syy, denominator, r, scatterplot pattern, range, subgroup or phase labels, influence checks, and intended descriptive use.

Pearson r has no universal ABA threshold for weak, moderate, strong, clinically important, or actionable association. Its magnitude does not measure treatment effect, accuracy, reliability, agreement, interchangeability, or experimental control. Correlation also does not imply causation. A third variable, selection process, shared time trend, or common measurement artifact can produce association.

This page intentionally does not calculate a p-value or confidence interval. Inference requires a sampling framework, assumptions, and qualified review that are not supplied by two copied columns.

Stop conditions and escalation

Pause for incomplete or unmatched pairs, repeated or clustered units, time-series dependence, a near-constant variable, uncertain pairing, post hoc transformations or exclusions, influential observations with unresolved provenance, strong nonlinearity, range restriction that defeats the question, changing variance, weights, covariate adjustment, multiple-comparison searches, or planned generalization beyond the displayed pairs.

Obtain statistical or research-methods review for confidence intervals, hypothesis tests, permutation or bootstrap procedures, hierarchical or time-series models, partial correlation, power, sample-size planning, or any causal interpretation. Preserve the raw data and graphs while a more suitable design is selected.

Privacy and provenance

Use the minimum row-level information needed for review. Store identifiable material only in approved systems, limit access, and preserve the source version and calculation record. The HHS Privacy Rule summary and HHS Security Rule summary describe federal requirements for regulated entities. State and organizational rules may add duties. This page is not legal advice.

Copyable result record

Result fieldValueData versionX and Y definitions, units, and directionsIndependent unit and pairing keyComplete pairs nX mean / Y meanSxySxx / SyyDenominatorPearson rScatterplot shape and rangePhase or subgroup reviewLeave-one-out sensitivity rangeMissing or excluded entriesIntended interpretation and limitsReviewer and review date

Related resources

Sources