The kvr2 package provides functions to calculate nine types of coefficients of determination ($R^2$) as classified by Kvalseth (1985).

The coefficient of determination, $R^2$, is one of the most common metrics for assessing model fit. However, its mathematical definition is not unique. While various formulas yield identical results in standard linear regression with an intercept, they can diverge significantly—sometimes producing negative values or values exceeding 1—when applied to:

• Models without an intercept (No-intercept models)
• Power regression models
• Other fits via transformations (e.g., log-log models)

This package is specifically designed for models that can be represented as lm objects in R. This includes:

• Standard Linear Models (with an intercept)
• No-intercept Models (e.g., lm(y ~ x - 1))
• Power Regression Models (fitted via log-transformation, such as lm(log(y) ~ log(x)))

Note: This package does not support general non-linear least squares (nls) or other complex non-linear modeling frameworks. It focuses on the mathematical sensitivity of $R^2$ within the context of linear estimation and its common transformations.

The primary goal of kvr2 is not to provide a definitive “best” for every scenario, but to serve as an educational and diagnostic resource. Many users rely on the single value provided by standard software, but as this package demonstrates, that value is sensitive to the underlying mathematical definition and the software’s internal defaults.

Through this package, users can:

• Understand Mathematical Sensitivity: Observe firsthand how different algebraic formulas (eight + one definitions) can lead to dramatically different interpretations of the same model fit, especially in non-intercept models.
• **Diagnose Negative **: It is imperative to acknowledge that a negative (typically in) should not be interpreted as a “bug”; rather, it functions as a critical diagnostic signal. This signal indicates that the model predicts outcomes that fall below the mean of a simple horizontal line.
• Evaluate Robustness and Transformations: Explore Kvalseth’s recommendations for using for consistency and for robustness against outliers, and see how behaves when models are fitted in transformed spaces (e.g., log power regression models).

The package calculates nine indices based on Kvalseth (1985):

• $R^2_1$ to $R^2_8$: A classification of existing and historical formulas used in statistical literature and software.
• $R^2_9$: A robust version of the coefficient of determination based on median absolute deviations, as proposed in the original paper.

You can install the released version of kvr2 from CRAN with:

install.packages("kvr2")

You can install the development version of kvr2 like so:

remotes::install_github("indenkun/kvr2")

kvr2 provides a simple way to observe how different $R^2$ definitions behave across various model specifications.

In standard linear models with an intercept, most $R^2$ definitions yield identical results. However, they can diverge significantly in models without an intercept or in power regression models.

library(kvr2)

# Dataset from Kvalseth (1985)
df1 <- data.frame(x = 1:6, y = c(15, 37, 52, 59, 83, 92))

# Case A: Linear regression with intercept (Values are consistent)
model_int <- lm(y ~ x, data = df1)
r2(model_int)
#> R2_1 :  0.9808 
#> R2_2 :  0.9808 
#> R2_3 :  0.9808 
#> R2_4 :  0.9808 
#> R2_5 :  0.9808 
#> R2_6 :  0.9808 
#> R2_7 :  0.9966 
#> R2_8 :  0.9966 
#> R2_9 :  0.9778

# Case B: Linear regression without intercept (Values diverge)
model_no_int <- lm(y ~ x - 1, data = df1)
results <- r2(model_no_int)
results
#> R2_1 :  0.9777 
#> R2_2 :  1.0836 
#> R2_3 :  1.0830 
#> R2_4 :  0.9783 
#> R2_5 :  0.9808 
#> R2_6 :  0.9808 
#> R2_7 :  0.9961 
#> R2_8 :  0.9961 
#> R2_9 :  0.9717

Observation: In Case B, notice that $R^2_2$ and $R^2_3$ exceed 1.0. This demonstrates why choosing the correct definition is critical for models without an intercept.

The r2() function returns a list object. While the output is formatted for readability, you can easily access individual values for further analysis or reporting.

# Accessing specific R2 values from the result object
results$r2_1
#>      r2_1 
#> 0.9776853

results$r2_9
#>      r2_9 
#> 0.9717156

# You can also use it in your custom functions or data frames
my_val <- results$r2_1

To complement $R^2$ analysis, use comp_fit() to evaluate models via standard error metrics such as RMSE, MAE, and MSE.

comp_fit(model_no_int)
#> RMES :  3.9008 
#> MAE :  3.6520 
#> MSE :  18.2593

For details, refer to the documentation for each function.

Kvalseth, T. O. (1985). Cautionary Note about $R^2$. The American Statistician, 39(4), 279-285. DOI: 10.1080/00031305.1985.10479448