B Coefficient: Measuring Direct Predictor Effects

In quantitative empirical inquiry, establishing the exact rate at which an outcome variable shifts in response to an isolated unit change in an explanatory variable is foundational to scientific inference. The unstandardized regression coefficient, universally represented in statistical literature and computational output as the B coefficient, serves as the primary metric for quantifying this direct, real-world relationship within linear and generalized linear modeling frameworks. By preserving the original observational scale of both predictor and criterion measures, the B coefficient bridges theoretical econometric and psychometric paradigms with actionable, tangible interpretation across diverse scientific domains.

B Coefficient

1. Concise Definition

The B coefficient (often written as an unstandardized regression coefficient, partial slope, or raw metric weight) represents the expected change, in raw measurement units, of a dependent or criterion variable associated with a one-unit increase in an independent predictor variable, holding all other covariates in the model constant. When estimated within an ordinary least squares or maximum likelihood framework, it quantifies the isolated linear rate of change attributable to a specific explanatory feature. Consequently, it reflects the directional magnitude and natural baseline impact of that variable within its native physical, psychological, or economic units of measurement.

Unlike its standardized counterpart (frequently denoted by the lowercase Greek letter beta, β), the unstandardized B coefficient is deliberately metric-dependent. It retains the operational scale of the collected data, allowing researchers to evaluate ecological or clinical shifts directly without transforming variables into abstract standard deviation metrics. Within multivariable specifications, the B coefficient functions as a partial derivative of the estimated conditional expectation function, isolating the independent predictive capacity of a singular predictor while analytically neutralizing the confounding influences of mutually correlated indicators.

Beyond basic ordinary least squares linear regression, the mathematical construct of the B coefficient generalizes across the broader taxonomy of generalized linear models (GLMs). In logistic, Poisson, and proportional hazards regressions, the estimated B coefficient characterizes the rate of change in the transformed link function (such as log-odds or log-rates) per unit change in the explanatory index. This conceptual continuity makes the B coefficient one of the most foundational and ubiquitous parameters in empirical research, serving as the core engine for quantitative hypothesis testing, structural equation modeling, and causal inference.

2. Etymology & Linguistic Origin

The nomenclature of the B coefficient derives from early modern statistical notation developed in Western Europe during the late nineteenth and early twentieth centuries. The Latin letter B was formally introduced as an algebraic shorthand for the secondary parameter of a linear relationship, succeeding A (which traditionally designated the intercept or constant origin) in the classical linear equation Y = A + B(X). This sequential typographic convention mirrors foundational European algebraic traditions tracing back to René Descartes’s seventeenth-century designation of early alphabetic letters for known constants and terminal letters for variables.

The term coefficient itself is derived from the Neo-Latin formation coefficiens, combining the prefix co- (meaning “together with” or “jointly,” from the classical Latin cum) and the present participle of efficere (signifying “to bring about,” “to accomplish,” or “to produce effect”). The linguistic term was first popularized in a mathematical context by the French mathematician François Viète in the late sixteenth century to denote a numerical or symbolic factor multiplying an algebraic quantity. When British polymath Francis Galton and mathematical statistician Karl Pearson formalized the mathematical paradigms of regression and correlation, Latinate designations were preserved to distinguish empirical sample estimators (denoted by Roman capital or lowercase letters such as B or b) from population parameters (traditionally denoted by Greek characters such as β).

3. Pronunciation & Grammatical Form

The term is pronounced phonetically in standard academic English as /biː koʊ.ɪˈfɪʃ.ənt/ (bee koh-ih-FISH-uhnt). Morphologically and syntactically, “B coefficient” operates as a compound noun phrase, where “B” acts as an uninflected attributive nominal modifier qualifying the head noun “coefficient.” In contemporary statistical manuscripts, authors frequently pluralize the head noun as “B coefficients” when referring to a vector of regression slopes within multiple regression matrices.

In standard typesetting conventions, the designation alternates between an italicized uppercase Roman letter (B) and an italicized lowercase Roman letter (b), depending on the governing stylistic guidelines. For instance, the American Psychological Association (APA) style guide stipulates the use of italicized uppercase B to designate unstandardized regression slopes, reserving the Greek lowercase β strictly for standardized regression weights. Grammatically, the term typically occurs within prepositional constructs such as “the unstandardized B coefficient for socioeconomic status” or as a predicate nominative describing model output parameters.

4. Detailed Conceptual Explanation

To grasp the foundational mechanics of the B coefficient, one must examine the mathematical formalization of ordinary least squares (OLS) multiple regression. In a canonical model involving a dependent variable Y and a vector of k continuous or categorical predictors X₁, X₂, …, X_k, the theoretical equation expresses the conditional expectation of Y given X as a linear combination of parameters plus a stochastic error term (ε). Each unstandardized partial regression slope B_j reflects the partial derivative of the expectation function with respect to X_j, formally expressed as ∂E(Y|X)/∂X_j = B_j.

This partial derivative property indicates that B isolates the unique marginal association between that specific predictor and the outcome. In simple bivariate regression, the unstandardized coefficient is computed directly as the ratio of the sample covariance between X and Y to the sample variance of X (Cov(X, Y) / Var(X)). However, in multivariable scenarios, the estimation must purge the predictor of all shared variance with other explanatory variables. Through matrix algebra, the vector of unstandardized coefficients B is mathematically derived via the ordinary least squares closed-form solution B = (X^TX)^-1X^TY, where X represents the design matrix of predictors and Y constitutes the observed response vector.

Because the design matrix X and the response vector Y are expressed in their original, unstandardized units of observation, the resulting values of B are intrinsically tied to those scales. If an outcome is measured in annual income (dollars) and the predictor is measured in completed education (years), a calculated B value of 4,500 indicates that each additional calendar year of education completed corresponds to an expected increase of $4,500 in annual income, assuming all other covariates in the regression model remain invariant. Altering the unit of measurement—such as converting years of schooling to months, or annual income to thousands of dollars—proportionately scales the numeric magnitude of B, while leaving its statistical significance, test statistics, and underlying explanatory power completely unchanged.

Furthermore, the conceptual boundaries of the B coefficient must be distinguished across diverse modeling paradigms. In linear models, the B coefficient indicates an additive, constant marginal effect across the entire continuum of the predictor variable. Conversely, in non-linear or generalized linear models utilizing non-identity link functions (such as the logit link in logistic regression or the log link in Poisson models), B characterizes an additive shift on the unobservable latent or transformed scale. Converting this parameter into actionable real-world probability changes requires calculating nonlinear transformations, such as computing an odds ratio via exponentiation (e^B) or estimating marginal effects at representative values of the data distribution.

5. Historical Development

The intellectual lineage of the B coefficient is deeply intertwined with the development of the method of least squares and early mathematical biometrics during the nineteenth century. Carl Friedrich Gauss and Adrien-Marie Legendre independently synthesized the mathematical principles of minimizing squared residual errors between observed coordinates and hypothetical mathematical functions in the early 1800s, primarily to solve complex astronomical and geodetic orbital trajectories. However, these mathematical tools were initially conceptualized as deterministic error-correction procedures rather than probabilistic models for variable biological or social phenomena.

In the late nineteenth century, British polymath Francis Galton adapted least-squares methodology to explore human heredity, observing that the stature of adult offspring exhibited a statistical tendency to “revert” or “regress” toward the general population mean relative to the stature of their parents. In his seminal 1886 treatise on regression toward mediocrity, Galton introduced the concept of the regression line. Building upon Galton’s conceptual discoveries, mathematical statistician Karl Pearson established the formal algebraic apparatus of bivariate and multiple regression throughout the 1890s. Pearson derived the analytical foundations of partial correlation and unstandardized multiple regression coefficients, establishing the mathematical relationship between the raw covariances of empirical variables and the slope parameters of multidimensional least-squares planes.

Throughout the early decades of the twentieth century, the conceptual framework of the B coefficient was further refined by the legendary statistician Ronald A. Fisher, who integrated regression analysis into a rigorous system of null-hypothesis significance testing and the analysis of variance (ANOVA). Fisher developed exact distribution theory for the standard error of the B coefficient, introducing the Student’s t-test distribution to assess whether an empirical sample regression slope differed significantly from an expected null value of zero. Simultaneously, in agricultural and genetic research, Sewall Wright developed path analysis in the 1920s, formalizing the distinct analytical properties of unstandardized path coefficients (B) versus standardized path coefficients (β) when evaluating direct and indirect causal hypotheses within directional path diagrams.

By the mid-twentieth century, the rapid growth of econometrics spearheaded by the Cowles Commission, Jan Tinbergen, and Ragnar Frisch expanded the application of B coefficients into complex systems of simultaneous equations and dynamic macroeconomic forecasting. The advent of modern mainframe and desktop computing during the 1970s and 1980s further democratized the estimation of B coefficients across behavioral, psychological, and medical sciences through popular statistical packages such as SPSS, SAS, and later R and Stata. Today, the B coefficient remains an indispensable metric across computational data science, structural equation modeling, and contemporary causal inference paradigms.

6. Theoretical Foundations

The mathematical validity and statistical properties of the B coefficient rest upon the classical Gauss-Markov theorem within ordinary least squares regression theory. According to this fundamental theorem, under a specific set of rigorous structural assumptions, the unstandardized OLS estimator B is guaranteed to be the Best Linear Unbiased Estimator (abbreviated as BLUE). This means that among the theoretical universe of all linear, unbiased estimation algorithms, the least-squares B coefficient demonstrates the minimal possible sampling variance, thereby offering maximal statistical efficiency and precision.

The foundational Gauss-Markov assumptions required for the B coefficient to achieve BLUE status include:

• Linearity in Parameters: The relationship between the criterion variable and the explanatory predictors must be structurally linear in terms of its parameters, such that Y = B₀ + B₁X₁ + … + B_kX_k + ε.
• Strict Exogeneity: The conditional expectation of the stochastic residual error term ε given the explanatory matrix X must equal zero (E(ε|X) = 0). This crucial assumption implies that no unobserved confounding variables omitted from the model correlate systematically with the included predictors.
• Absence of Perfect Multicollinearity: The design matrix X must have full column rank, meaning that no predictor variable can be written as an exact linear combination of any other explanatory factors.
• Spherical Errors (Homoscedasticity and No Autocorrelation): The stochastic error terms must possess constant variance across all levels of the predictors (Var(ε_i|X) = σ²) and must exhibit zero statistical covariance or autocorrelation across distinct observations (Cov(ε_i, ε_j|X) = 0 for all i ≠ j).

When researchers additionally assume that the error terms are normally distributed (ε ~ N(0, σ²I)), the sampling distribution of each estimated B_j follows an exact Student’s t-distribution centered at the true population parameter. This distributional assumption enables the construction of exact confidence intervals and two-tailed hypothesis tests. If any of these core theoretical assumptions are violated—such as the presence of heteroscedasticity, endogeneity, or substantial measurement error—the resulting B coefficient may become inefficient, biased, or statistically inconsistent, necessitating remedial estimation techniques such as generalized least squares, instrumental variables estimation, or robust standard error adjustments.

7. Key Components, Types & Dimensions

To fully contextualize the B coefficient in modern computational and psychometric workflows, it is vital to categorize its primary variations, operational components, and dimensional characteristics:

• Unstandardized B Coefficient (Slope): The primary regression parameter that reflects the absolute change in the raw, unstandardized units of the criterion variable for every one-unit adjustment in the predictor variable.
• Standard Error of the B Coefficient (SE_B): The estimated standard deviation of the sampling distribution of B across repeated hypothetical sampling from the target population. It serves as the denominator in calculating the Student’s t-statistic (t = B / SE_B) and determines the width of parameter confidence intervals.
• Intercept or Constant (B₀): The specific baseline unstandardized coefficient that indicates the expected numerical value of the criterion variable when all included explanatory variables in the regression equation are set to zero.
• Standardized Regression Coefficient (β): The transformation of the B coefficient obtained when all predictor and criterion variables are rescaled to have a mean of zero and a standard deviation of one prior to model estimation. It quantifies change in terms of standard deviation units rather than raw ecological metrics.
• Partial Regression B Coefficient: An unstandardized slope estimated within a multivariable linear framework that represents the unique effect of a predictor after statistically partialling out and controlling for all other covariates in the system.
• Logistic B Coefficient: The unstandardized slope parameter estimated within a binary or multinomial logistic regression, reflecting the linear change in the natural logarithm of the odds (log-odds) of an event occurring per single-unit change in the predictor.

8. Examples & Illustrative Cases

Consider an educational psychologist investigating the determinants of academic performance among secondary school students. The researcher specifies a multivariable linear regression model where the dependent variable is an standardized achievement examination score (scaled from 0 to 100 points). The two primary independent variables are the number of weekly study hours (measured in raw hours) and family income (measured in thousands of dollars annually). Following ordinary least squares estimation on an empirical sample of 800 students, the computational software outputs the following equation: Exam Score = 45.20 + 1.85(Study Hours) + 0.12(Family Income).

In this empirical case, the unstandardized B coefficient for weekly study hours is B = 1.85 (with an estimated standard error of SE_B = 0.25, t = 7.40, p < .001). This parameter demonstrates that for every additional hour of weekly study time a student engages in, their predicted examination score increases by exactly 1.85 percentage points, holding family income strictly constant. Because the B coefficient preserves the authentic scale of the measurement, educational policymakers can readily evaluate whether an intervention that raises study time by five hours per week (yielding an anticipated improvement of 5 × 1.85 = 9.25 exam points) produces a meaningful difference in real-world educational achievement.

Conversely, consider a clinical pharmacologist examining systemic systolic blood pressure (measured in millimeters of mercury, mmHg) as a function of daily dosage of an antihypertensive compound (measured in milligrams, mg) alongside baseline patient age (measured in completed years). The estimated regression output generates a B coefficient for drug dosage of B = -0.75 (SE_B = 0.15, p < .001). This negative unstandardized slope reveals that for each incremental milligram of the pharmaceutical agent administered daily, patient systolic blood pressure drops on average by 0.75 mmHg, after statistically adjusting for the physiological effects of aging. The retention of concrete physical metrics (mg and mmHg) makes the B coefficient vital for clinical safety, therapeutic dosage calibration, and medical decision-making.

9. Measurement & Assessment

Evaluating the statistical integrity and substantive reliability of an estimated B coefficient requires a comprehensive, multi-step assessment protocol. The initial stage of evaluation focuses on null-hypothesis significance testing. The analyst formulates the conventional null hypothesis stating that the true population slope is zero (H₀: B_j = 0), indicating no linear association between the predictor and the outcome. The empirical test statistic is computed by dividing the parameter estimate by its estimated standard error: t = B_j / SE(B_j), which is evaluated against a Student’s t-distribution with N – k – 1 degrees of freedom (where N is sample size and k is the number of predictors).

Because relying solely on point estimates and p-values can obscure sampling uncertainty, reporting standards in contemporary empirical science emphasize the construction of confidence intervals. A conventional 95% confidence interval for an unstandardized B coefficient is calculated using the formula: CI_95% = B_j ± [t_crit × SE(B_j)], where t_crit is the two-tailed critical threshold corresponding to the chosen alpha level. If this computed confidence interval excludes zero, the analyst rejects the null hypothesis at that significance threshold, while the width of the interval directly communicates the precision of the estimator.

In addition to sampling precision, the assessment of a B coefficient requires rigorous diagnostic screening for multicollinearity and model specification errors. When two or more explanatory predictors correlate strongly with one another, the Variance Inflation Factor (VIF) for the predictor spikes. A high VIF (frequently identified by thresholds exceeding 5.0 or 10.0) inflates the standard error of the B coefficient, destabilizing the numerical estimate, widening confidence intervals, and increasing the risk of Type II statistical errors. Diagnostic workflows must evaluate residual normality, studentized residuals, leverage values, and Cook’s distance to ensure that individual extreme outliers do not disproportionately bias the estimated B coefficient.

10. Applications & Practical Significance

The unstandardized B coefficient plays an indispensable operational role across clinical, educational, industrial, and macroeconomic sectors. In clinical epidemiology and public health policy, unstandardized parameters are preferred over standardized indices because health interventions occur on established physical and biological scales. When epidemiologists study the relationship between daily particulate air pollution exposure (measured in μg/m³) and pediatric asthma hospital admissions, the resulting B coefficient provides a direct calculation of expected hospital admissions per increment of ambient pollution. This raw metric information is necessary for environmental regulatory agencies to formulate statutory air quality limits and calculate public health economic costs.

Within organizational psychology and human resource analytics, the B coefficient informs personnel selection and compensatory frameworks. When industrial psychologists evaluate the predictive validity of a cognitive ability assessment on supervisory job performance ratings, the unstandardized slope reflects the concrete gain in job output units per unit increase on the pre-employment assessment battery. This quantification enables accurate utility analysis, allowing corporate leaders to calculate the projected financial return on investment from implementing psychometric testing systems.

Furthermore, in econometric and macro-financial modeling, unstandardized B coefficients are fundamental for fiscal and monetary policy simulations. Central banking institutions estimate B coefficients to project how incremental adjustments in national benchmark interest rates (measured in basis points) systematically alter annualized headline inflation rates, private capital investment, and consumer borrowing behaviors. Because financial policy decisions involve tangible capital and distinct currency denominations, unstandardized metrics provide an interpretable, realistic foundation for quantitative decision modeling.

11. Research & Empirical Evidence

Decades of methodological literature have debated the relative empirical utility of unstandardized B coefficients versus standardized β weights. In a foundational paper on causal modeling, statistical methodologist Kenneth Bollen demonstrated that standardized coefficients are inherently unstable when comparing structural models across different independent populations or observational time periods. Bollen illustrated that because standardized coefficients conflate the underlying structural association with the sample-specific variance of the variables, an apparent difference in β between two distinct demographic groups often reflects nothing more than variance heterogeneity across groups rather than a true discrepancy in the biological, psychological, or social mechanism under study.

Complementing these foundational insights, psychometric researchers John Hunter and Frank Schmidt highlighted in their meta-analytic work that unstandardized B coefficients remain invariant to artificial sample selection artifacts and restriction of range phenomena that frequently distort standardized metrics. When empirical samples suffer from range restriction on a predictor—such as selecting only high-achieving applicants within an educational testing setting—the observed standard deviation of the predictor drops sharply. This compression artificially attenuates both the Pearson correlation coefficient (r) and the standardized β weight. In contrast, under classical linear assumptions, the expected value of the unstandardized B coefficient remains unbiased and structurally stable across both unrestricted and restricted populations.

More recent empirical work within contemporary causal inference frameworks, such as the structural causal models developed by Judea Pearl, reinforces the central role of unstandardized B coefficients in identifying structural path parameters. In directed acyclic graphs (DAGs) and non-parametric structural equation systems, the partial unstandardized slope reflects the direct causal mechanism operating between variables. Modern observational studies in medicine, sociology, and behavioral economics rely heavily on unstandardized regression slopes to calculate average treatment effects (ATE) and marginal structural parameters, confirming the B coefficient’s foundational status in contemporary statistical and computational modeling.

12. Cultural & Cross-Cultural Considerations

When implementing regression analyses within cross-cultural, cross-national, or global comparative research frameworks, the substantive interpretation of the unstandardized B coefficient requires careful contextual adjustments. Because unstandardized coefficients are defined entirely by the original measurement scales of the variables, their direct comparability across distinct cultural settings depends on the psychometric equivalence and semantic congruence of those measures. If a psychological construct—such as subjective emotional distress or job satisfaction—is measured using translated Likert instruments that demonstrate differential item functioning (DIF) across cultural cohorts, the resulting unstandardized B coefficients will reflect cultural response biases rather than genuine cross-cultural structural differences.

Moreover, cross-national economic studies must navigate complex currency conversion challenges when utilizing B coefficients. If an international comparative study examines the impact of household income on childhood educational attainment across multiple sovereign nations, reporting an unstandardized B coefficient in raw local currency denominations (e.g., Japanese Yen versus United States Dollars versus Swiss Francs) prevents meaningful international comparisons. Cross-cultural quantitative researchers must adjust such monetary variables using purchasing power parity (PPP) indices or standardized international monetary units to preserve comparability.

In cross-cultural educational testing, such as the Program for International Student Assessment (PISA) and the Trends in International Mathematics and Science Study (TIMSS), researchers frequently use B coefficients to analyze the relationship between socioeconomic status and academic literacy across participating nations. Methodologists emphasize that interpreting these unstandardized slopes requires evaluating whether the underlying educational structures, institutional tracking practices, and social mobility mechanisms share structural equivalence across the societies under investigation. A one-unit change in an index of economic, social, and cultural status may represent vastly different lived experiences, educational opportunities, and material access in a developing nation compared to an industrialized post-welfare state.

13. Criticisms, Debates & Limitations

Despite its mathematical elegance and broad real-world utility, the unstandardized B coefficient faces substantive criticisms and analytical limitations. The most prominent practical challenge is that unstandardized B coefficients cannot be compared directly with one another within the same regression equation to determine the relative importance of different predictors. Because each B coefficient is expressed in its own raw physical, temporal, or scale metric, a variable measured in small increments (such as body weight in grams) may yield a tiny B coefficient, whereas a variable measured in broad increments (such as educational attainment in decades) may yield a large B coefficient, even if the former predictor has a much stronger statistical association with the outcome.

This relative incomparability has sparked persistent debates between substantive domain researchers and methodological purists. Applied practitioners frequently rely on standardized beta weights (β) or variance-decomposition techniques (such as dominance analysis and relative weight analysis) to establish predictor rank-order hierarchies within complex observational models. Furthermore, critics emphasize that unstandardized B coefficients are vulnerable to misinterpretation when the underlying explanatory predictor is an arbitrary, non-standardized psychological composite score. In psychometric inventories lacking natural physical units—such as an unanchored 17-item self-report scale of organizational commitment—the real-world meaning of a single-unit change on the questionnaire is ambiguous, diminishing the intuitive interpretive advantages of the raw B coefficient.

An additional substantive limitation stems from the vulnerability of the OLS B coefficient to measurement error in the independent variables. Classical measurement error in explanatory predictors induces regression dilution (or attenuation bias), systematically driving the estimated B coefficient toward zero and producing inconsistent parameter estimates. Finally, in observational cross-sectional research designs, unstandardized B coefficients are frequently misinterpreted as direct causal effects. In the absence of randomized experimental manipulation, rigorous natural experiment conditions, or exhaustive conditioning on all confounding variables via directed acyclic graphs, an unstandardized B coefficient merely captures conditional correlational associations rather than genuine causal mechanics.

14. Related Terms & Distinctions

To avoid conceptual confusion in empirical data modeling, the B coefficient must be distinguished from several related statistical indices:

• B Coefficient vs. Standardized Beta (β): The B coefficient quantifies the marginal change in the outcome in raw, unstandardized measurement units per one-unit increase in the predictor. In contrast, the standardized beta weight (β) quantifies that relationship in terms of standard deviation units, allowing researchers to evaluate relative predictive magnitude across diverse variables within a single sample.
• B Coefficient vs. Pearson Correlation Coefficient (r): The Pearson correlation coefficient measures the degree of linear association between two variables on a standardized scale bounded strictly between -1.0 and +1.0, treating both variables symmetrically. The B coefficient represents an asymmetric directional regression slope that depends directly on the observational metric of the variables and can take any finite real value from -∞ to +∞.
• B Coefficient vs. Odds Ratio (OR): In binary logistic regression, the B coefficient represents the linear change in the natural logarithm of the odds of an outcome per unit change in the explanatory factor. The odds ratio corresponds to the exponentiated B coefficient (OR = e^B), transforming the additive logarithmic parameter into a multiplicative factor reflecting the relative odds of the event.
• B Coefficient vs. Intercept (B₀): The intercept is the baseline constant parameter indicating the expected value of the dependent variable when all predictors equal zero, whereas the slope B coefficients (B₁, …, B_k) govern the rate of directional change across the multidimensional predictor space.

15. Summary & Key Takeaways

The unstandardized B coefficient represents one of the most foundational and practically meaningful parameters in modern quantitative methodology. By quantifying the expected change in a dependent criterion variable in its native observational units per one-unit change in an explanatory predictor—while holding other covariates constant—the B coefficient translates abstract statistical models into tangible, real-world relationships. From its historical origins in the nineteenth-century biometric breakthroughs of Galton and Pearson to its rigorous mathematical formulation under the Gauss-Markov theorem, the B coefficient provides an intuitive and mathematically robust tool for empirical discovery.

While standardized metrics are helpful for comparing the relative statistical strength of disparate predictors within a single sample, unstandardized B coefficients are essential for testing theoretical models across independent populations, conducting range-invariant meta-analyses, and informing evidence-based policy interventions. When properly supported by diagnostic tests for collinearity, measurement validity, and model specification, the B coefficient serves as an indispensable bridge connecting abstract quantitative data to concrete scientific, clinical, and social phenomena.

References

• Bollen, K. A. (1989). Structural equations with latent variables. John Wiley & Sons. https://doi.org/10.1002/9781118619179
• Cohen, J., Cohen, P., West, S. G., & Aiken, L. S. (2003). Applied multiple regression/correlation analysis for the behavioral sciences (3rd ed.). Lawrence Erlbaum Associates. https://doi.org/10.4324/9780203774441
• Galton, F. (1886). Regression towards mediocrity in hereditary stature. The Journal of the Anthropological Institute of Great Britain and Ireland, 15, 246–263. https://doi.org/10.2307/2841583
• Hunter, J. E., & Schmidt, F. L. (2004). Methods of meta-analysis: Correcting error and bias in research findings (2nd ed.). SAGE Publications. https://doi.org/10.4135/9781412985031
• Pearl, J. (2009). Causality: Models, reasoning, and inference (2nd ed.). Cambridge University Press. https://doi.org/10.1017/CBO9780511803161