Predicting with a Fitted Ridge Model in R

From a Fitted Ridge Model to Predictions

A fitted ridge model is useful only if it can be applied to new cases. In R, the predict method for a glmnet fit does this calculation: it combines the fitted coefficients with a matrix of new predictor values. A prediction call can look simple, but its result is only meaningful when the penalty value, predictor columns, and output scale are all understood.

This tutorial focuses on numeric predictions from a fitted ridge model at a chosen value of lambda. It shows how to check the input and output dimensions, interpret response-scale predictions, and verify the calculations independently. The examples use hypothetical data and fitted coefficient values.

What the Prediction Call Requires

For an ordinary Gaussian ridge model, suppose the fitted conditional mean has the form

Here \(p\) is the number of predictors, \(x_j\) is the value of predictor \(j\) for the new case, and the coefficients depend on the selected \( \lambda \). The model is fitted using a squared-error criterion with a ridge penalty on the predictor coefficients; the intercept is not penalized. For interpreting the fitted mean, the usual linear-model conditions include a correctly specified conditional mean and errors with mean zero. Additional conditions, such as constant error variance or normal errors, matter for some inferential procedures, but are not needed simply to calculate a fitted prediction.

In R, a typical call is predict(fit, newx = x_new, s = 0.4, type = "response"). The object fit is a fitted glmnet model. The object x_new must be a numeric matrix with one row per new case and one column for each predictor used in fitting. Its columns must have the same meaning and order as the training matrix. The s argument identifies the requested penalty value. When type = "response" is used with a Gaussian model, the result is on the outcome's original response scale.

If fitting used standardize = TRUE, predictions still use new predictor values in the same units as the training matrix. The fitted model applies its internal standardization consistently. If the analyst manually transformed the training predictors before fitting, the new predictors must receive that same transformation before prediction. predict.glmnet does not infer what a column means from its name; matching the number of columns alone is not enough.

Why the Output Has This Form

For a Gaussian model with an identity link, the response-scale prediction is the fitted linear predictor. For one new case, the model computes the intercept plus the sum of each predictor value multiplied by its fitted coefficient. With multiple new cases, the same calculation is performed row by row. With multiple penalty values, each value selects a different coefficient vector, so the calculations form separate columns.

Stacking these prediction vectors for \(K\) requested penalty values gives an \(n_{\text{new}}\)-by-\(K\) matrix. This explains both the dimensions and the contents of the result. In particular, a single requested penalty value generally still gives a one-column matrix rather than a plain vector.

Worked Predictions for a Gaussian Ridge Fit

Response Scale and Link Scale

For Gaussian ridge regression, the identity link means the fitted linear predictor and response-scale prediction have the same numerical value. That equivalence does not hold for every family. For example, in a binomial model the linear predictor is a log-odds value, while a response-scale prediction is a probability. The requested type therefore matters when the model has a non-identity link.

For a binomial model with linear predictor \( \eta \), the inverse-logit transformation converts the linear predictor to the estimated probability \(p\):

Choosing Lambda and Avoiding Prediction Errors

The value supplied through s is the penalty at which predictions are requested; it is not automatically the value that produced the best predictive performance. The choice of \( \lambda \) should come from the modeling decision already made, such as a specified value or a value selected using a separate evaluation procedure. If a cross-validation fit object is used, values such as lambda.min can identify a selected penalty; predictions should still be checked to ensure the intended value and object were supplied.

A requested value of \( \lambda \) need not be one of the exact values stored in the fitted model's path. glmnet can interpolate coefficients for requested values between fitted values. For a question that supplies a particular penalty, use that value in s; do not silently replace it with a nearby entry from the fit.

The most consequential errors are often data-preparation errors rather than arithmetic errors. A matrix with the right dimensions can still be wrong if its columns are reversed or if a factor's dummy columns do not match the training design. Similarly, manually standardized training predictors require new data transformed using the same training means and scales. Check column names and order before prediction, and avoid relying on a data frame that may silently coerce columns to an unintended type.

In practice, a reliable prediction check has three parts: confirm the fitted model and \( \lambda \), confirm that every new-data column matches the training input, and confirm that the result has the expected scale and dimensions. These checks make a short R call auditable and reduce the risk of reporting a technically valid but incorrectly interpreted number. This treatment is aligned with the SOA Exam PA (Predictive Analytics) syllabus.