Cross validation

Prof. Maria Tackett

Mar 05, 2026

Announcements

• Lab 05 due TODAY at 11:59pm

• HW 03 due Tuesday, March 17 at 11:59pm

• Statistics experience due April 15

• Please provide mid-semester feedback by Friday: https://duke.qualtrics.com/jfe/form/SV_3lvkRQbz7PMuJVA

• Exam 02 date proposed date change

  • In-class: Thursday, April 9

  • Take-home: Thursday, April 9 - Saturday, April 11

  • Please email me by tomorrow if you have any concerns

Topics

• Training and testing sets
• Cross validation

Computational setup

# load packages
library(tidyverse)
library(tidymodels)
library(patchwork)
library(knitr)
library(kableExtra)

# set default theme and larger font size for ggplot2
ggplot2::theme_set(ggplot2::theme_bw(base_size = 20))

Data: Restaurant tips

Which variables help us predict the amount customers tip at a restaurant? To answer this question, we will use data collected in 2011 by a student at St. Olaf who worked at a local restaurant.

# A tibble: 169 × 4
     Tip Party Meal   Age   
   <dbl> <dbl> <fct>  <fct> 
 1  2.99     1 Dinner Yadult
 2  2        1 Dinner Yadult
 3  5        1 Dinner SenCit
 4  4        3 Dinner Middle
 5 10.3      2 Dinner SenCit
 6  4.85     2 Dinner Middle
 7  5        4 Dinner Yadult
 8  4        3 Dinner Middle
 9  5        2 Dinner Middle
10  1.58     1 Dinner SenCit
# ℹ 159 more rows

Variables

Predictors:

• Party: Number of people in the party
• Meal: Time of day (Lunch, Dinner, Late Night)
• Age: Age category of person paying the bill (Yadult, Middle, SenCit)
• Payment: Payment type (Cash, Credit, Credit/CashTip)

Response vs. predictors

Training vs. testing sets

• The training set (i.e., the data used to fit the model) does not have the capacity to be a good arbiter of performance.

• It is not an independent piece of information; predicting the training set can only reflect what the model already knows.

• Suppose you give a class a test, then give them the answers, then provide the same test. The student scores on the second test do not accurately reflect what they know about the subject; these scores would probably be higher than their results on the first test.

• We can reserve some data for a testing set that can be used to evaluate the model performance

Training and testing sets

Create training and testing sets using functions from the resample R package (part of tidymodels)

Step 1: Create an initial split:

set.seed(210)
tips_split <- initial_split(tips, prop = 0.75) #prop = 3/4 by default

Step 2: Save training data

tips_train <- training(tips_split)
dim(tips_train)

[1] 126  13

Step 3: Save testing data

tips_test <- testing(tips_split)
dim(tips_test)

[1] 43 13

Application exercise

📋 sta210-sp26.github.io/ae/ae-09-model-compare.html

Cross validation

Spending our data

• We have already established that the idea of data spending where the test set was recommended for obtaining an unbiased estimate of performance.
• However, we usually need to understand the effectiveness of the model before using the test set.
• Typically we can’t decide on which final model to take to the test set without making model assessments.
• Remedy: Resampling to make model assessments on training data in a way that can generalize to new data.

Resampling for model assessment

Resampling is only conducted on the training set. The test set is not involved. For each iteration of resampling, the data are partitioned into two subsamples:

• The model is fit with the analysis set. Model fit statistics such as Adj. \(R^2\) and \(R^2\) are computed based on this model fit.
• The model is evaluated with the assessment set.

Resampling for model assessment

Image source: Kuhn and Silge. Tidy modeling with R.

Analysis and assessment sets

• Analysis set is analogous to training set.
• Assessment set is analogous to test set.
• The terms analysis and assessment avoids confusion with initial split of the data.
• These data sets are mutually exclusive.

Cross validation

More specifically, v-fold cross validation – commonly used resampling technique:

• Randomly split your training data into v partitions
• Use v-1 partitions for analysis, and the remaining 1 partition for analysis (model fit + model fit statistics)
• Repeat v times, updating which partition is used for assessment each time

Let’s give an example where v = 3…

To get started…

Split data into training and test sets

set.seed(210)
tips_split <- initial_split(tips, prop = 0.75)
tips_train <- training(tips_split)
tips_test <- testing(tips_split)

To get started…

Specify model

tips_spec <- linear_reg()

tips_spec

Linear Regression Model Specification (regression)

Computational engine: lm 

To get started…

Create workflow

tips_wflow1 <- workflow() |>
  add_model(tips_spec) |>
  add_formula(Tip ~ Age + Party)

tips_wflow1

══ Workflow ════════════════════════════════════════════════════════════════════
Preprocessor: Formula
Model: linear_reg()

── Preprocessor ────────────────────────────────────────────────────────────────
Tip ~ Age + Party

── Model ───────────────────────────────────────────────────────────────────────
Linear Regression Model Specification (regression)

Computational engine: lm 

Cross validation, step 1

Randomly split your training data into 3 partitions:

Tips: Split training data

folds <- vfold_cv(tips_train, v = 3)
folds

#  3-fold cross-validation 
# A tibble: 3 × 2
  splits          id   
  <list>          <chr>
1 <split [84/42]> Fold1
2 <split [84/42]> Fold2
3 <split [84/42]> Fold3

Cross validation, steps 2 and 3

• Use v-1 partitions for analysis, and the remaining 1 partition for assessment
• Repeat v times, updating which partition is used for assessment each time

Cross validation: Tips data

• Question
• Submit

There are 126 observations in the training data. How many observations are in a single assessment set?

🔗 https://forms.office.com/r/n5wYDEKnFX

Tips: Fit resamples

tips_fit_rs1 <- tips_wflow1 |>
  fit_resamples(resamples = folds)

tips_fit_rs1

# Resampling results
# 3-fold cross-validation 
# A tibble: 3 × 4
  splits          id    .metrics         .notes          
  <list>          <chr> <list>           <list>          
1 <split [84/42]> Fold1 <tibble [2 × 4]> <tibble [0 × 4]>
2 <split [84/42]> Fold2 <tibble [2 × 4]> <tibble [0 × 4]>
3 <split [84/42]> Fold3 <tibble [2 × 4]> <tibble [0 × 4]>

Cross validation, now what?

• We’ve fit a bunch of models
• Now it’s time to use them to collect metrics (e.g., RMSE, \(R^2\) ) on each model and use them to evaluate model fit and how it varies across folds

Collect metrics from CV

# Produces summary across all CV
collect_metrics(tips_fit_rs1)

# A tibble: 2 × 6
  .metric .estimator  mean     n std_err .config        
  <chr>   <chr>      <dbl> <int>   <dbl> <chr>          
1 rmse    standard   2.14      3  0.195  pre0_mod0_post0
2 rsq     standard   0.623     3  0.0470 pre0_mod0_post0

Note: These are calculated using the assessment data

Deeper look into results

cv_metrics1 <- collect_metrics(tips_fit_rs1, summarize = FALSE) 

cv_metrics1

# A tibble: 6 × 5
  id    .metric .estimator .estimate .config        
  <chr> <chr>   <chr>          <dbl> <chr>          
1 Fold1 rmse    standard       1.92  pre0_mod0_post0
2 Fold1 rsq     standard       0.713 pre0_mod0_post0
3 Fold2 rmse    standard       2.52  pre0_mod0_post0
4 Fold2 rsq     standard       0.554 pre0_mod0_post0
5 Fold3 rmse    standard       1.97  pre0_mod0_post0
6 Fold3 rsq     standard       0.603 pre0_mod0_post0

Better presentation of results

cv_metrics1 |>
  mutate(.estimate = round(.estimate, 3)) |>
  pivot_wider(id_cols = id, names_from = .metric, values_from = .estimate) |>
  kable(col.names = c("Fold", "RMSE", "R-sq"))

Fold	RMSE	R-sq
Fold1	1.915	0.713
Fold2	2.525	0.554
Fold3	1.967	0.603

Cross validation in practice

• To illustrate how CV works, we used v = 3:

  • Analysis sets are 2/3 of the training set
  • Each assessment set is a distinct 1/3
  • The final resampling estimate of performance averages each of the 3 replicates

• This was useful for illustrative purposes, but v is often 5 or 10; we generally prefer 10-fold cross-validation as a default

Example model selection workflow

• Exploratory data analysis

• Using training data…

  • Fit and evaluate candidate models using cross validation

  • Select the best fit model

  • Check model conditions and diagnostics

  • Repeat as needed until you’ve landed on final model

• Evaluate the final model performance using the test set

Tip

See Section 10.5.2 for more detailed workflow

Data analysis workflow

Source: Introduction to Regression Analysis

Recap

• Training and testing sets
• Cross validation

Next class

☀️ Have a good spring break ☀️