A Comprehensive Guide to Selecting, Calculating, and Interpreting Correlation and Association Coefficients in Medical Research

2.1. Pearson Correlation Coefficient

A common descriptive approach for examining the relationship between two quantitative variables is a scatter plot, in which each variable is represented on one axis. The closer the data points cluster around a straight line, the stronger the linear association—either positive or negative.

The most widely used measure of such a relationship is the Pearson correlation coefficient, which quantifies the degree of linear correlation between two variables.

It is defined as the covariance of the variables divided by the product of their standard deviations, effectively providing a normalized measure of covariance with values ranging from -1 to 1. A positive value indicates that X and Y tend to increase or decrease together, whereas a negative value implies that higher values of one variable are associated with lower values of the other.

The maximum value of 1 is obtained if the points in the scatter plot lie exactly on a straight line. Conversely, the correlation coefficient is negative if high values of Y tend to go with low values of X, and vice versa [7]. For two quantitative variables, X and Y, the Pearson correlation coefficient

where x̄ and ȳ are the sample means of the X and Y variables, respectively. The formula uses sample means and sample standard deviations, making it an estimate of the population correlation coefficient.

2.2. Spearman Correlation

The Spearman correlation coefficient, similar to the Pearson correlation coefficient, assesses the strength of a monotonic relationship between two variables. Unlike Pearson's coefficient, it is based on the ranks of the data rather than their raw values. To compute this coefficient, the data for each variable are first ranked in ascending or descending order. Next, the difference (D) between the ranks for each pair of observations is determined. Finally, given n pairs of observations, the Spearman correlation coefficient is estimated as:

The Spearman rank correlation is less sensitive to outliers and relies on fewer assumptions compared to the Pearson correlation coefficient [1].

When ties occur, average ranks are assigned to tied observations. Most statistical software automatically handles ties using this approach. For small sample sizes in R, we recommend using cor.test(..., method = 'spearman', exact = TRUE) to obtain exact p-values; for larger samples, the asymptotic approximation is appropriate.

2.3. Phi Coefficient

The Phi coefficient is used to assess the association between two dichotomous (two-level) variables, typically represented in a 2×2 contingency table. It is calculated as the square root of the chi-square statistic divided by the sample size [8]:

Note that the Phi coefficient is strictly applicable only to 2×2 contingency tables. For tables with more than two categories, alternative measures such as Cramér's V (see Section 2-5) or the Contingency Coefficient (Section 2-4) should be used. For instance, in diagnostic test accuracy studies, Phi can be used to quantify the association between a dichotomous test result (positive/negative) and disease status (present/absent).

2.4. Contingency Coefficient

If either variable has more than two levels, the Phi coefficient can exceed 1. To address this issue, Pearson introduced the Contingency Coefficient, defined as [6]:

In the above formula, χ² denotes the chi-square statistic and n represents the sample size.

2.5. Cramér's V

Since the contingency coefficient seldom reaches its theoretical maximum of 1, Cramér proposed a scaled measure, known as Cramér's V, defined as [9]:

Here, n denotes the sample size, and min(r,c) represents the smaller of the number of rows or columns. It should be noted that for 2×2 tables, the values of Cramér's V and the Phi coefficient are identical.

2.6. Tetrachoric Correlation Coefficient

The tetrachoric correlation coefficient estimates the association between two dichotomous variables under the assumption of an underlying bivariate normal distribution. In other words, the two variables are theoretically continuous and normally distributed but have been artificially dichotomized [1]. This coefficient is widely applied in various fields, including factor analysis of binary items and the assessment of inter-rater agreement. For a two-by-two contingency table with cell counts a, b, c, and d, the tetrachoric correlation coefficient (r_t) is estimated as:

The formula presented is a convenient approximation suitable for most practical applications. For greater accuracy, particularly with small samples or extreme marginal proportions, we recommend using maximum-likelihood estimation as implemented in R's psych::tetrachoric() function, which yields more precise estimates.

2.7. Tschuprow's T

Tschuprow's T is a measure of association between two nominal variables, taking values in the range from 0 to 1 [9]. If the number of rows and columns is the same and equal to 2, the values of Phi, Cramér's V, and Tschuprow's T coincide. Tschuprow's T statistic is estimated as:

where χ² denotes the chi-square statistic, n represents the sample size, and r and c correspond to the number of rows and columns, respectively.

While Cramér's V is more commonly used in practice, Tschuprow's T is often preferred when the number of rows and columns are unequal, as it is less biased toward larger tables. For rectangular tables (r ≠ c), Tschuprow's T provides a more conservative estimate of association compared to Cramér's V. Using the simulated dataset (nominal1 × nominal2, a 3×3 table), Tschuprow's T was calculated as 0.167, indicating weak association.

2.8. Lambda Correlation Coefficient

The lambda coefficient quantifies the strength of association between two nominal variables. It is derived from the Proportional Reduction in Error (PRE) in predicting the dependent variable when the independent variable is used, relative to the prediction error without it. Denoting the prediction error with the independent variable as E1 and the error without it as E2, the lambda coefficient is calculated as:

According to the above formula, the lambda coefficient ranges from 0 to 1. It should be noted that this coefficient is not symmetric [3].

For example, in a study examining the association between comorbidity status (categorical, e.g., none, mild, severe) and mortality (yes/no), Lambda can be calculated with mortality as the dependent variable (Y), measuring the proportional reduction in error in predicting mortality when comorbidity status is known. If Lambda = 0.045, this indicates that comorbidity status provides only a 4.5% improvement in predicting mortality compared to guessing without that information.

In medical research, the choice of dependent variable should be guided by the research question and clinical context. Typically, the dependent variable is the clinical outcome of interest (e.g., disease status, mortality, treatment response), while the independent variable is the predictor or risk factor (e.g., comorbidity status, risk factor category). Researchers should clearly specify which variable is treated as dependent when reporting Lambda coefficients.

2.9. Gamma Coefficient

The gamma coefficient (Goodman and Kruskal's gamma) measures the degree of association between two sets of ranked data and is applicable to ordinal variables. In datasets with a large number of tied ranks, the gamma coefficient is often preferred over alternative measures, such as Somers' D or Kendall's Tau, due to its robustness to outliers. The value of Goodman and Kruskal's gamma is calculated as:

where Nc denotes the total number of concordant pairs, and Nd represents the total number of discordant pairs [10]. For two dichotomous variables, Yule's Q is typically used instead.

2.10. Yule's Q Coefficient

Yule's Q is a measure of association between two binary variables [11]. It is a special case of Goodman and Kruskal's gamma for a two-by-two table. Its value is obtained from the following equation:

2.11. Uncertainty Coefficient

The uncertainty coefficient, also known as the entropy coefficient, is a measure of association for nominal variables. It quantifies the proportion of entropy (uncertainty) in the variable Y that can be explained by the variable X, denoted by U(X|Y), and takes values between 0 and 1 [12]. Suppose we have samples from two discrete random variables, X and Y. By constructing the joint distribution, P_X,Y (x, y), and the conditional distribution, P_X|Y (x|y), the uncertainty coefficient can be computed as follows:

where

A value of 1 indicates complete certainty and perfect predictability, whereas a value of 0 corresponds to maximum uncertainty.

For example, if we consider symptom severity (nominal variable X) and laboratory test results (nominal variable Y), an Uncertainty Coefficient U(X|Y) = 0.25 would indicate that knowledge of the laboratory result reduces the uncertainty in predicting symptom severity by 25% compared to having no information about the laboratory result.

2.12. Kendall's Tau Coefficient

Kendall's Tau is a nonparametric statistic that measures the ordinal association between two variables and was introduced by Maurice Kendall in 1938 [2]. It assesses the degree of similarity between the rankings of the data across variables. The coefficient ranges from −1 to 1:

• A value of 1 indicates perfect agreement: all pairs are concordant.
• A value of -1 indicates perfect disagreement: all pairs are discordant.
• A value of 0 indicates the absence of association.

Let (x₁, y₁), (x₂, y₂),…,(x_n, y_n) be a set of joint observations.

• A pair of observations (x_i, y_i) and (x_j, y_j), where i < j, is concordant if the ranks for both elements agree:

sign(x_j − x_i) = sign(y_j − y_i)

• The pair is discordant:

sign(x_j − x_i) = −sign(y_j − y_i)

• The pair is tied: if x_i = x_j or/and y_i = y_j.

Three standard variants of Kendall's Tau have been developed to accommodate different types of data structures:

2.12.1. Tau-a

This version is applied when there are no tied ranks. It represents the simplest form of Kendall's Tau and is calculated by the difference between the proportions of concordant and discordant pairs as:

2.12.2. Tau-b

The most common version, appropriate for square contingency tables where X and Y have the same number of categories. It adjusts for ties on both variables:

2.12.3. Tau-c

Designed for rectangular contingency tables (different numbers of categories for X and Y). It accounts for table dimensions and is robust across tables with different shapes:

In the above relations, m is the smaller of the number of rows (r) and columns (c), and:

• N_c = Number of concordant pairs
• N_d = Number of discordant pairs
• n = Total number of observations
• T_x = Number of pairs tied only on X
• T_y = Number of pairs tied only on Y

2.13. Somers' D

Somers' Delta (Somers' D) is an asymmetric measure of association between two ordinal variables, originally proposed by Robert H. Somers in 1962 [13]. It is conceptually analogous to a regression coefficient for ordinal data, quantifying the extent to which knowledge of one variable enhances the prediction of the other’s order. Somers' D ranges between −1 and 1.

• 1: Perfect agreement; knowledge of X allows perfect prediction of Y's rank.
• -1: Perfect disagreement.
• 0: No predictive power.

A key distinction between Somers' D and Kendall's Tau is that Somers' D is asymmetric, meaning that the measure depends on which variable is designated as the predictor (independent) variable and which as the response (dependent) variable. Formally, Somers' D can be expressed in terms of Kendall's Tau-b as follows:

Somers' D is denoted as D(Y|X), where Y is the dependent (outcome) variable and X is the independent (predictor) variable, emphasizing that the measure depends on which variable is designated as the outcome.

However, a more direct and commonly used computational form, which explicitly reflects its asymmetric nature, is given by:

Where:

• N_c = Number of concordant pairs
• N_d = Number of discordant pairs
• T_y = the number of pairs that are tied on the dependent variable (Y) but not on the independent variable (X)

Interpretation of the Formula:

The denominator N_c + N_d + T_y represents the total number of observation pairs that are not tied on the independent variable (X). Accordingly, Somers' D measures the difference between the proportion of concordant and discordant pairs among those for which X differentiates between cases. Pairs tied on X are excluded, since when X does not distinguish between two observations, it provides no information for predicting their relative ordering on Y.

For instance, in a study examining the relationship between treatment intensity (ordinal X, e.g., low, medium, high) and pain score improvement (ordinal Y, e.g., no improvement, mild improvement, significant improvement), D(Y|X) would quantify how well treatment intensity predicts pain improvement.

2.14. Cohen's Kappa Coefficient

Cohen's kappa coefficient (κ) is a statistical measure that quantifies inter-rater (and intra-rater) reliability for categorical data [14]. The coefficient is computed using the following equation:

where P₀ denotes the proportion of observed agreement, while P_E represents the expected agreement by chance. Table 1 illustrates the contingency for inter-rater agreement analysis.

Where:

• a = Number of cases where both raters agreed on a “Yes” response
• b = Number of cases in which Rater 1 responded “Yes” but Rater 2 responded “No”
  
  
  
• c = Number of cases where Rater 1 said “No” but Rater 2 said “Yes”
• d = Number of cases where both Raters said “No”
• n = Total number of cases, calculated as (a + b + c + d)

The observed agreement and the expected agreement by chance for computing Cohen's Kappa in this contingency table are defined as follows:

and

Interpretation Guidelines:

• κ ≤ 0: No agreement
• 0.01-0.20: Slight agreement
• 0.21-0.40: Fair agreement
• 0.41-0.60: Moderate agreement
• 0.61-0.80: Substantial agreement
• 0.81-1.00: Almost perfect agreement

2.15. Eta Squared Coefficient

Eta squared (η²) is a measure of effect size, defined as the proportion of the total variance in the outcome variable (Y) that is explained by an independent variable (X). In other words, it quantifies the share of variability in Y attributable to X. The eta squared coefficient can be estimated using the following formula:

Here SS_effect denotes the sum of squares for the effect of interest, while SS_total represents the total sum of squares, including all effects, interactions, and error variance in the ANOVA. Another related measure, partial eta squared, quantifies the proportion of variance attributable to a specific effect relative to the variance associated with that effect plus its corresponding error variance. The partial eta squared coefficient is calculated as follows:

Both η² and partial η² take values between 0 and 1.

2.16. Omega Squared

Omega squared (ω²), similar to Eta squared, repre- sents the proportion of variance in the dependent variable that is explained by an independent variable. Omega squared is generally considered a less biased estimator than Eta squared, particularly for small sample sizes [15]. It can be calculated using the following formula:

Where df_effect and MS_error are the degrees of freedom associated with the effect (calculated as k - 1, where k is the number of groups or levels of the independent variable in an ANOVA model) and the Mean Square Error, respectively.

2.17. Intraclass Correlation Coefficient (ICC)

The Intraclass Correlation Coefficient (ICC) is a widely used measure of reliability in test-retest, intrarater, and interrater assessments. Originally, ICC was introduced as a modification of the Pearson correlation coefficient [4]. In contemporary practice, however, ICC is computed based on mean squares derived from Analysis of Variance (ANOVA) [16]. Several forms of ICC estimators have been proposed, most of which can be defined within the framework of a random-effects model:

Where:

• Y_ij denotes the j-th observation (e.g., a specific rater's score) in the i-th group (e.g., a subject/patient).
• μ represents the overall grand mean across all observations.
• α_i is the random effect of the i-th group, reflecting the deviation of that group's mean from μ. This effect is typically assumed to follow a normal distribution, α_i ~ N(0,σ_α²).
• ϵ_ij denotes the random error associated with each measurement, assumed to be distributed as ϵ_ij ~ N(0,σ_ϵ²).

Accordingly, the total variance associated with any observation is σ_α² + σ_ϵ².

The population ICC in this model is defined as the proportion of the between-group variance relative to the total variance:

The ICC ranges from 0 to 1 and represents the proportion of total variance attributable to differences between groups. McGraw and Wong introduced ten distinct forms of ICC, classified according to the model (one-way random, two-way random, or two-way mixed), the type (single measurement or the mean of k measurements/raters), and the definition (consistency or absolute agreement) [17].

Commonly accepted thresholds for ICC interpretation are: <0.50 = poor reliability, 0.50-0.75 = moderate reliability, 0.75-0.90 = good reliability, and >0.90 = excellent reliability

The McGraw-Wong classification distinguishes 10 ICC forms based on model (one-way random, two-way random, two-way mixed), type (single measurement or average of k measurements), and definition (consistency or absolute agreement) [17].

For example, in assessing inter-rater agreement in radiology where two radiologists independently evaluate the presence of a lesion on MRI scans, ICC(2,1) (two-way random model, single measurement, absolute agreement) would be appropriate to assess reliability.

2.18. Point-Biserial Correlation

The point-biserial correlation coefficient (r_pb) quantifies the association between a continuous variable and a naturally dichotomous variable [18]. For instance, it can be used to assess the correlation between drug administration (yes/no) and blood pressure. The point-biserial correlation is computed as:

Here, M₁ and M₀ denote the means of the continuous variable for the groups with binary outcomes 1 and 0, respectively; s is the standard deviation of the continuous variable. The quantities p and q represent the proportions of observations for which the binary variable takes the values 1 and 0, respectively.

2.19. Biserial Correlation

The biserial correlation coefficient r_b quantifies the association between a continuous variable and a variable that has been artificially dichotomized (that is, a continuous variable split into two categories using a cutoff point—for example, students classified as passing or failing an exam). The biserial correlation coefficient is estimated as:

where y is the value of the standard normal density at z, with P(Z ≤ z) = p and q = 1 - p. The remaining terms M₀, M₁, s, p and q are defined in the same way as for the point-biserial correlation.

The biserial correlation assumes that the underlying continuous variable follows a normal distribution. When this assumption is violated (common with skewed clinical variables such as laboratory values or biomarker concentrations), the point-biserial correlation (which makes no such assumption) or Spearman's correlation may be more appropriate alternatives. Practitioners should assess normality using histograms or Shapiro-Wilk tests before applying biserial correlation.
 

3.1. Key Recommendations for Correlation Analysis

(1) Check Assumptions: Always ensure that the assumptions of the selected correlation coefficient (such as linearity, normality, and appropriate measurement level) are satisfied. Applying an unsuitable coefficient may produce misleading results.

(2) Visualize Data: Before calculating a correlation, plot the data (using a scatter plot for continuous variables or a clustered bar chart for categorical variables) to examine the relationship's structure.

(3) Interpret with Caution: Remember that correlation does not imply causation. A significant correlation may result from an unmeasured third variable (confounder).

(4) Consider the Context: Select the correlation coeffi- cient based on the research question and the characte- ristics of the data, rather than solely for convenience.

(5) Report Accurately: When presenting correlations, always provide the coefficient value, sample size (n), and the p-value or confidence interval. For measures like Somers' D, clearly indicate which variable is considered the dependent variable.

(6) Be Aware of Limitations: Certain coefficients, such as Pearson's r, are sensitive to outliers. Others, such as the Contingency Coefficient, have an upper bound below 1, making comparisons across different table sizes difficult.

4.1. Pearson Correlation (r)

Based on Fisher's z-transformation for linear correlation between two continuous variables [7, 19]:

where

• n = required sample size0
• Z_α/2 = critical value for type I error (α), typically 1.96 for α = 0.05 (two-tailed)
• Z_β = critical value for type II error (β), typically 0.84 for β = 0.20 (80% power)
• r = anticipated Pearson correlation coefficient

4.2. Spearman Correlation (ρ)

Same formula as Pearson, using Spearman's rho as the effect size [1, 19]. Valid for monotonic relationships:

where

• ρ = anticipated Spearman correlation coefficient

4.3. Point-Biserial Correlation (r_pb)

Same formula as Pearson, using point-biserial correlation as the effect size [18]:

where

• r_pb = anticipated point-biserial correlation coefficient

4.4. Kendall's Tau-b (τ_b)

Modified formula for Kendall's rank correlation coefficient [2, 19]:

• τ_b = anticipated Kendall's Tau-b coefficient

4.5. Intraclass Correlation Coefficient (ICC)

Based on the simplified guide for determining sample size requirements for estimating ICC [16, 20]:

where

• ρ_ICC = anticipated ICC value

4.6. Cohen's Kappa (κ)

Alternative formula for Cohen's Kappa (κ) [14, 21] is:

where

• κ = anticipated Kappa coefficient
• κ₀ = hypothesized Kappa under null hypothesis (usually 0)
• P = chance agreement probability

4.7. Biserial Correlation (r_b)

where

• r_b = anticipated biserial correlation coefficient ([18, 22]).

4.8. Cramér's V (V)

Formula for chi-square based association [9]:

where

• df = degrees of freedom based on the smaller of (rows - 1, columns - 1)
• V = anticipated Cramér's V coefficient
• r = number of rows in the contingency table
• c = number of columns in the contingency table

An alternative formula is:

where

• w = Cohen's w effect size derived from V, w=V×dfw=V×df (Cohen's w effect size)

4.9. Phi Coefficient (ϕ)

The required sample size for 9 Phi Coefficient (ϕ) [8, 23] can be calculated as follows:

Alternative formula (for 2×2 tables):

• ϕ = anticipated Phi coefficient

4.10. Gamma Coefficient (G)

Using Fisher's transformation approximation for Gamma coefficient [10, 24]:

where

• G = anticipated Goodman and Kruskal's Gamma coefficient

As shown in Fig. (1), the process begins with identifying the measurement scales of the variables (continuous, ordinal, nominal, or dichotomous), followed by selecting the appropriate coefficient from 19 commonly used measures, including Pearson's r, Spearman's ρ, Cramér's V, Cohen's κ, and ICC. The selected coefficient can then be implemented in major statistical software packages (R, SPSS, Stata, SAS, Python, etc.) and interpreted based on its magnitude and the research context. This framework provides a comprehensive guide that bridges statistical theory and practical application in medical research.

5.1. Calculation of Correlation Coefficients

5.2. Sample Output and Interpretation

Based on the values obtained, only three coefficients (Pearson, Spearman, and ANOVA effect sizes) demonstrated strong and meaningful relationships. Most coefficients for nominal and ordinal variables (e.g., Phi, Cramér's V, Lambda, Gamma) were relatively small, indicating weak associations in this simulated dataset. This pattern underscores the importance of selecting correlation coefficients that align with the variable types and the true strength of relationships in the data (Table 3).

5.3. Real Medical Data Example

To demonstrate practical application, we supplement the simulated data with an example from a published biomarker study [25] that examines the correlation between serum biomarkers and clinical outcomes. The data show a Pearson correlation of 0.56 between biomarker level (continuous) and disease severity score (continuous), and Cramér's V of 0.41 between biomarker category (low/mid/high) and disease stage (I/II/III). Complete analysis code is provided in the supplementary materials.