Correlation or Regression Coefficient Transformation¶
Coefficient harmonization method — instantiates Effect Size Standardization
Converts association estimates — correlations and regression slopes — into comparable effect-size units, and inter-converts between the correlation and mean-difference families.
Not every effect arrives as a group contrast. Much of it comes as association: a correlation coefficient, an unstandardized regression slope in raw units, a standardized beta. Correlation or Regression Coefficient Transformation is the mechanism that makes these commensurable — with each other and, critically, with effects from the mean-difference world. Its defining move is the bridge: it standardizes a raw slope into unit-free form and inter-converts across effect families (a correlation r into a standardized mean difference d, and back), so a study reporting an association can sit in the same table as one reporting a group difference. Where its siblings each live inside one metric family, this one's whole reason to exist is crossing between families while keeping the conversion auditable.
Example¶
An economist assembles evidence on how income relates to self-reported wellbeing. One study reports a correlation of r = 0.20 between log-income and a life-satisfaction scale. Another reports an unstandardized regression slope: 0.8 satisfaction points per $10,000, which is uninterpretable next to a correlation until it is standardized — multiply by the SD of income and divide by the SD of satisfaction to get a standardized beta of about 0.18. A third study only compared "high-income" vs. "low-income" groups and reported d = 0.45. To place all three on one axis, the analyst converts the group-difference d into a correlation-family point-biserial value (here about 0.22, given the split), and now has ~0.20, ~0.18, ~0.22 — genuinely comparable. Each conversion is logged with its formula and the inputs it used (the split proportion for the d↔r step, the two SDs for the slope), so a reviewer can retrace every hop and see the sign was preserved throughout.
How it works¶
The distinguishing work is choosing and recording the right bridge:
- Standardize raw slopes. Multiply an unstandardized coefficient by the predictor's SD over the outcome's SD to get a standardized beta comparable across models.
- Inter-convert families. Apply the standard algebraic maps — r ↔ d, r ↔ standardized beta, d ↔ point-biserial — appropriate to the estimand, using the group-split proportion where a mean-difference conversion needs it.
- Stabilize correlations for combination. Move to Fisher's z before averaging or comparing correlations, since raw r is bounded and skewed near ±1.
- Preserve sign and log the hop. Fix which direction is positive and record every conversion — source metric, formula, inputs, result — in a traceable chain.
It harmonizes metrics; it does not itself attach uncertainty intervals or judge importance.
Tuning parameters¶
- Standardizing SDs — which variances define "standardized" for a slope; using sample vs. population SDs, or a restricted-range SD, rescales the coefficient.
- Conversion formula choice — the specific r↔d map, which embeds assumptions (e.g. an even group split, bivariate normality) that may not hold.
- Fisher-z vs. raw-r combination — whether correlations are averaged on the stabilized scale (recommended) or naively, which biases pooled associations toward zero.
- Partial vs. zero-order — whether a coefficient is adjusted for covariates; mixing partial and simple correlations in one table compares unlike quantities.
When it helps, and when it misleads¶
Its strength is reach: it lets association-heavy literatures — psychology, economics, education — enter a synthesis that would otherwise be blocked because half the studies "don't report a d." The Fisher transformation is what makes correlations safe to combine at all.[n1]
Its hazard is that every cross-family conversion smuggles in assumptions, and the deceptively simple ones bite hardest: the d↔r map depends on the group-split proportion, so a conversion done at an assumed 50/50 split silently distorts effects from unbalanced designs, and standardizing a slope by a range-restricted SD inflates it just as it would an SMD. The classic misuse is pooling standardized betas from models with different covariate sets as if they estimated the same thing. The guarding discipline is to record every conversion's formula and inputs so the assumptions are visible and reversible, and to refuse conversions across estimands that are not actually the same construct.
How it implements the components¶
scale_and_unit_inventory— catalogs each source coefficient's type, scale, and units so the correct standardizing or conversion map is chosen.standardization_rule— declares the slope-standardization and family-conversion formulas that map raw associations to comparable effect sizes.conversion_traceability_table— logs every hop (source metric, formula, inputs, result) so a multi-step conversion chain stays auditable and reversible.directionality_and_sign_convention— fixes positive/negative meaning so signs survive the family conversions consistently.
It does not preserve the original raw group means or produce a base d from scratch — that raw_effect_estimate_record and estimand_definition work is Standardized Mean Difference Calculation; this sibling starts from an already-estimated coefficient and bridges between families rather than computing the first standardized effect. Attaching intervals is Confidence Interval Propagation via uncertainty_attachment.
Related¶
- Instantiates: Effect Size Standardization — supplies the cross-family bridge that brings association estimates onto the common axis.
- Sibling mechanisms: Standardized Mean Difference Calculation · Hedges Correction Application · Confidence Interval Propagation · Risk Ratio or Odds Ratio Standardization · Absolute Risk Difference Translation · Meta-Analytic Effect Harmonization · Minimal Important Difference Anchoring · Forest Plot or Effect Table Display
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Correlation or Regression Coefficient Transformation operates as a computation, comparison, model, or analytic representation used to infer, estimate, or choose because it converts association estimates — correlations and regression slopes — into comparable effect-size units, and inter-converts between the correlation and mean-difference families.
Independent corroboration: The frozen evidence defines Correlation or Regression Coefficient Transformation as 'Converts association estimates — correlations and regression slopes — into comparable effect-size units, and inter-converts between the correlation and mean-difference families', so its operative form is Analysis, Modeling & Optimization.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Single lineage
Present-day reach: Specialized
Rationale: Meta-analysis and applied statistics cohered formulas that standardize slopes and convert association coefficients into comparable effect-size families.
Related originating lineages:
- Mathematics — Algebraic relationships among standardized covariance, correlation, and mean-difference measures supply the conversion identities.
Review resolution: Transformations of correlation and regression coefficients cohered in statistical inference; mathematical derivation is foundational but not a separate applied lineage.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
[n1] Fisher's z transformation applies the inverse hyperbolic tangent to a correlation, converting the bounded, skewed r scale into one that is approximately normal with a variance depending only on sample size. Averaging or testing correlations on the z scale and back-transforming avoids the bias that plagues naive averaging of raw r values near ±1. ↩