# Why isn't simulation showing that ridge regression better than linear model

Cross Validated Asked by andy_dorsey on January 1, 2022

I am learning about ridge regression. I was under the impression that ridge regression is valuable because it provides better out of sample predictive accuracy than standard linear models. For example, see the bottom of page 217 in this well known statistical learning text: http://faculty.marshall.usc.edu/gareth-james/ISL/ISLR%20Seventh%20Printing.pdf. I tried setting up a short simulation to demonstrate it, but my results aren’t showing that ridge models are superior.

First, I simulated the exact multiarm design using DeclareDesign in R (the only difference is I boosted the N = 300). I then set up a simulation where I simulated a data set 1,000 times, split it into a test and training data set, and then fit a linear model and ridge regression model to the training data set. I then examined how well each model predicted responses in the test data set. Surprisingly, I don’t show that the linear model does any worse. I must be going wrong somewhere, right? Below is my code – it doesn’t take long to run and I’d appreciate any tips on where I might have went wrong.

# Add libraries
library(DeclareDesign)
library(ridge)
library(tidyverse)
library(fastDummies)

# Use DeclareDesign to get function that can simulate data
N <- 300
outcome_means <- c(0.5, 1, 2, 0.5)
sd_i <- 1
outcome_sds <- c(0, 0, 0, 0)

population <- declare_population(N = N, u_1 = rnorm(N, 0, outcome_sds[1L]),
u_2 = rnorm(N, 0, outcome_sds[2L]), u_3 = rnorm(N, 0, outcome_sds[3L]),
u_4 = rnorm(N, 0, outcome_sds[4L]), u = rnorm(N) * sd_i)
potential_outcomes <- declare_potential_outcomes(formula = Y ~ (outcome_means +
u_1) * (Z == "1") + (outcome_means + u_2) * (Z == "2") +
(outcome_means + u_3) * (Z == "3") + (outcome_means +
u_4) * (Z == "4") + u, conditions = c("1", "2", "3", "4"),
assignment_variables = Z)
estimand <- declare_estimands(ate_Y_2_1 = mean(Y_Z_2 - Y_Z_1), ate_Y_3_1 = mean(Y_Z_3 -
Y_Z_1), ate_Y_4_1 = mean(Y_Z_4 - Y_Z_1), ate_Y_3_2 = mean(Y_Z_3 -
Y_Z_2), ate_Y_4_2 = mean(Y_Z_4 - Y_Z_2), ate_Y_4_3 = mean(Y_Z_4 -
Y_Z_3))
assignment <- declare_assignment(num_arms = 4, conditions = c("1", "2", "3",
"4"), assignment_variable = Z)
reveal_Y <- declare_reveal(assignment_variables = Z)
estimator <- declare_estimator(handler = function(data) {
estimates <- rbind.data.frame(ate_Y_2_1 = difference_in_means(formula = Y ~
Z, data = data, condition1 = "1", condition2 = "2"),
ate_Y_3_1 = difference_in_means(formula = Y ~ Z, data = data,
condition1 = "1", condition2 = "3"), ate_Y_4_1 = difference_in_means(formula = Y ~
Z, data = data, condition1 = "1", condition2 = "4"),
ate_Y_3_2 = difference_in_means(formula = Y ~ Z, data = data,
condition1 = "2", condition2 = "3"), ate_Y_4_2 = difference_in_means(formula = Y ~
Z, data = data, condition1 = "2", condition2 = "4"),
ate_Y_4_3 = difference_in_means(formula = Y ~ Z, data = data,
condition1 = "3", condition2 = "4"))
names(estimates)[names(estimates) == "N"] <- "N_DIM"
estimates$$estimator_label <- c("DIM (Z_2 - Z_1)", "DIM (Z_3 - Z_1)", "DIM (Z_4 - Z_1)", "DIM (Z_3 - Z_2)", "DIM (Z_4 - Z_2)", "DIM (Z_4 - Z_3)") estimates$$estimand_label <- rownames(estimates)
estimates$$estimate <- estimates$$coefficients
estimates$term <- NULL return(estimates) }) multi_arm_design <- population + potential_outcomes + assignment + reveal_Y + estimand + estimator # Get holding matrix for R2 values rsq_values <- matrix(nrow = 1000, ncol = 2) # Simulate for (i in 1:100){ # Get simulated data set input_data <- draw_data(multi_arm_design) # Format data for analysis input_data <- input_data %>% fastDummies::dummy_cols(select_columns = "Z", remove_first_dummy = TRUE) %>% select(Y:Z_4) # Prep training and test data #set.seed(206) # set seed to replicate results training_index <- sample(1:nrow(input_data), 0.7*nrow(input_data)) # indices for 70% training data - arbitrary training_data <- input_data[training_index, ] # training data test_data <- input_data[-training_index, ] # test data # Fit linear model lm_mod <- lm(Y ~ ., data = training_data) # Fit ridge regression ridge_mod <- linearRidge(Y ~ ., data = training_data) # Get actual (from test data) and fitted values for each model actual <- test_data$Y
lm_predicted <- predict(lm_mod, test_data) # predict linear model on test data
ridge_predicted <- predict(ridge_mod, test_data) # predict ridge model on test data

# See how well linear model from training data fits test data (expressed as R2)
lm_rss <- sum((lm_predicted - actual) ^ 2)
lm_tss <- sum((actual - mean(actual)) ^ 2)
rsq_values[i, 1] <- lm_rsq

# See how well ridge model from training data fits test data (expressed as R2)
ridge_rss <- sum((ridge_predicted - actual) ^ 2)
ridge_tss <- sum((actual - mean(actual)) ^ 2)
rsq_values[i, 2] <- ridge_rsq
}

# Make matrix into data frame
rsq_values <- data.frame(rsq_values)

# Summarize R2 values for linear model
summary(rsq_values$X1) # Summarize R2 values for ridge model summary(rsq_values$X2)


You are doing nothing wrong. Ridge regression, the LASSO, and other penalized-coefficient regressions yield biased estimations. The idea is that maybe accepting a little bias will greatly reduce the variance.

However, there is nothing in how ridge regression, the LASSO, etc. are formulated that guarantees they will perform better at predictions of out of sample. Sometimes a simple linear model informed by theory and created by an analyst who knows the problem domain can trounce a model selected by ridge regression. This is true across problem domains and in all sorts of circumstances.

This is, essentially, a question about model selection. There is no need for code; the issue is not specific to your data or method of inference. Your findings illustrate that model selection (or what ML/AI people call feature selection) is not a solved problem.

Answered by kurtosis on January 1, 2022

## Related Questions

### Error metric to compare ratios derived from a binary prediction task

0  Asked on November 9, 2021

### Machine learning to select features to create predictive model

0  Asked on November 9, 2021 by atrag

### Summary statistics to collect for data that is too large in volume

0  Asked on November 9, 2021 by ryu576

### Endogeneity testing using correlation test

3  Asked on November 9, 2021 by sabiste

### find upper and lower bounds on average causal effect $theta$

1  Asked on November 9, 2021 by moneyball

### How to correct for unequal observation time in poisson regression

1  Asked on November 9, 2021 by thomas-richardson

### On gradient boosting and types of encodings

0  Asked on November 9, 2021

### RFE: Pre-define a specificity threshold

1  Asked on November 9, 2021

### Best software for image segmentation for time-series images?

1  Asked on November 9, 2021 by tshast2

### semantic value of a convolution neural network

1  Asked on November 9, 2021

### Getting Different R-squared for Same Model Equation

0  Asked on November 9, 2021

### How to maximize the steady state transition probability for a state in a Markov chain by altering that state’s outgoing transition probabilities?

0  Asked on November 9, 2021 by rob-goldstone

### Find out if increase in a series of values is statistically significant

1  Asked on November 9, 2021

### Covariance between sample mean of two simple random samples

0  Asked on November 6, 2021 by ana-pau-de-la-fuente

### Distribution of the maximum of two correlated normal variables

2  Asked on November 6, 2021

### How to formulate a classification problem with time series element

1  Asked on November 6, 2021 by rishu

### Asymmetric error measure for forecasts

1  Asked on November 6, 2021 by user235039

### How to model variance of a heteroskedastic dataset

1  Asked on November 6, 2021 by yilei-huang

### What analysis do I do? (Wage, education, gender)

2  Asked on November 6, 2021 by marvdste

### Is random seed a hyper-parameter to tune in training deep neural network?

1  Asked on November 6, 2021