# load packages
library(tidyverse)
library(tidymodels)
library(knitr)
rail_trail <-
read_csv("https://sta221-fa25.github.io/data/rail_trail.csv")Variance Inflation Factors
These notes come from Prof. Maria Tackett’s course notes for STA221.
Here we explain the connection between the Variance Inflation Factor (VIF) and \(\mathbf{C} = (\mathbf{X}^\mathsf{T}\mathbf{X})^{-1}\). This explanation is motivated by Chapter 3 of Montgomery, Peck, and Vining (2021).
Unit length scaling
We have talked about standardizing predictors, such that
\[x_{ij_{std}} = \frac{x_{ij} - \bar{x_j}}{s_{x_{j}}}\] such that \(\bar{x}_j\) is the mean and \(s_{x_{j}}\) is the standard deviation of the predictor \(x_j\).
The standardized predictors have a mean of 0 and variance of 1.
Another common type of scaling is unit length scaling. We will denote these scaled predictors as \(w_j\). We apply this scaling on both the predictor and response variable in the following way.
\[w_{ij} = \frac{x_{ij} - \bar{x_j}}{\sqrt{s_{jj}}}\] where \(s_{jj} = \sum_{i = 1}^{n}(x_{ij} - \bar{x}_j)^2\)
The scaled response variable, denoted \(y^0\) is
\[y_i^0 = \frac{y_i - \bar{y}}{\sqrt{SST}}\]
where \(SST\) is the sum of squares total, \(\sum_{i=1}^n(y_i - \bar{y})^2\)
We will use the scaled predictor and response variable to show the relationship between \(\mathbf{C} = (\mathbf{X}^\mathsf{T}\mathbf{X})^{-1}\) and the formula for VIF. More specifically, that
\[ C_{jj} = VIF_{j} = \frac{1}{1 - R^2_j} \]
When we use the scaled predictors and response.
Variance inflation factor
We will use the rail_trail data from the notes to illustrate this connection. We will focus on the predictors hightemp, avgtemp, and precip.
We begin by creating unit length scaled versions of the variables.
hightemp_norm <- sum((rail_trail$hightemp - mean(rail_trail$hightemp))^2)
avgtemp_norm <- sum((rail_trail$avgtemp - mean(rail_trail$avgtemp))^2)
precip_norm <- sum((rail_trail$precip - mean(rail_trail$precip))^2)
volume_norm <- sum((rail_trail$volume - mean(rail_trail$volume))^2)
rail_trail <- rail_trail |>
mutate(hightemp_scaled = (hightemp - mean(hightemp)) / hightemp_norm^.5,
avgtemp_scaled = (avgtemp - mean(avgtemp)) / avgtemp_norm^.5,
precip_scaled = (precip - mean(precip)) / precip_norm^.5,
volume_scaled = (volume - mean(volume)) / volume_norm^.5
)The matrix \(\mathbf{W}^\mathsf{T}\mathbf{W}\) is equivalent to the correlation matrix for these predictors.
# use -1 to remove the intercept for the correlation matrix
W <- model.matrix(volume_scaled ~ hightemp_scaled + avgtemp_scaled + precip_scaled - 1, data = rail_trail)
t(W)%*%W hightemp_scaled avgtemp_scaled precip_scaled
hightemp_scaled 1.0000000 0.9196439 0.1343172
avgtemp_scaled 0.9196439 1.0000000 0.2725832
precip_scaled 0.1343172 0.2725832 1.0000000When we fit a model using the unit length scaling for the response and predictor variables, we would expect \(Var(\hat{\beta}_j) / \hat{\sigma}^2_{\epsilon} \approx 1\). As we see below, however, these values are greater than 1.
trail_model_scaled <- lm(volume_scaled ~ hightemp_scaled + avgtemp_scaled + precip_scaled , data = rail_trail)
beta_se <- tidy(trail_model_scaled)$std.error
sigma <- glance(trail_model_scaled)$sigma
beta_se^2 / sigma^2[1] 0.01111111 7.16188175 7.59715405 1.19343051(You can ignore the first element, which represents the intercept).
These values show how much the standard errors of the coefficients are inflated given the correlation between the predictors (the off diagonal elements of \(\mathbf{W}^\mathsf{T}\mathbf{W}\)). The amount by which the standard errors are inflated are called the variance inflation factors (VIF).
Under this model using the unit-length-scaled predictors and response, we see these variance inflation factors are equal to the diagonal elements of \(\mathbf{C} = (\mathbf{W}^\mathsf{T}\mathbf{W})^{1}\).
C <- solve(t(W) %*% W)
diag(C)hightemp_scaled avgtemp_scaled precip_scaled
7.161882 7.597154 1.193431 Thus,
\[C_{jj} = VIF_{j} = \frac{1}{1 - R^2_j}\]
References
Montgomery, Douglas C, Elizabeth A Peck, and G Geoffrey Vining. 2021. Introduction to Linear Regression Analysis. John Wiley & Sons.