When evaluating the fit of a statistical model, particularly within the realm of generalized linear models and logistic regression, one frequently encounters the concept of pseudo r squared. Unlike the familiar R-squared value from ordinary least squares regression, which directly measures the proportion of variance explained, this metric serves a distinct purpose. It provides a means to assess how well a model with predictors performs compared to a model with only the intercept, adapting a core concept to non-linear frameworks where the traditional definition does not apply directly.
Understanding the Motivation Behind Pseudo Measures
The foundation of any r-squared variant lies in comparing the deviance of a specific model to the deviance of a null model. In standard linear regression, the total sum of squares is partitioned into explained and unexplained components. For models estimated via maximum likelihood, such as those using logit or probit links, the likelihood function replaces the sum of squares. The logic remains similar: a model that predicts the outcome well will have a likelihood closer to the saturated model. Consequently, these metrics are often referred to as "pseudo" because they mimic the logic of R-squared rather than measuring the exact same quantity.
Key Types and Calculation Methods
Several formulas exist for computing this metric, each with slight variations in interpretation. One common approach is based on the likelihood ratio chi-square statistic, dividing the difference between the null deviance and model deviance by the null deviance. Another popular version, often associated with Cox and Snell, uses a formula that attempts to bound the value between 0 and 1, though it rarely reaches the upper limit. A third variant, proposed by Nagelkerke, scales the Cox-Snell measure to ensure a maximum value of 1, making it easier to compare across different studies or model types.
Interpretation Guidelines and Limitations
It is crucial to approach the numerical value of a pseudo r squared with caution. A value of .40 in a logistic model might represent an excellent fit, while the same number in a linear model would be considered very poor. The baseline for what constitutes a "good" fit is entirely context-dependent and varies by field. Furthermore, adding more predictors will almost always increase the value, even if those predictors are irrelevant. Therefore, these metrics should never be used in isolation; they are most effective when combined with hypothesis tests for individual coefficients, information criteria like AIC or BIC, and validation techniques such as cross-validation.
Distinguishing from Traditional R-Squared
A significant point of confusion arises from the difference between this pseudo measure and the standard coefficient of determination. The traditional R-squared is tied to the variance of the observed data, specifically the proportion of the total variation in the dependent variable that is predictable from the independent variables. Because pseudo r squared values are derived from the log-likelihood function, they do not correspond to variance explained. They are best understood as relative indices of model fit. This distinction is vital for researchers transitioning from linear modeling to generalized linear models.
