Multiple Linear Regression - Intro

Notes and in-class exercises

You can download the .qmd file for this activity here and open in R-studio. The rendered version is posted in the course website (Activities tab). I often experiment with the class activities (and see it in live!) and make updates, but I always post the final version before class starts. To be sure you have the most up-to-date copy, please download it once you’ve settled in before class begins.

Notes

Learning goals

By the end of this lesson, you should be familiar with:

  • some limitations of simple linear regression
  • the general goals behind multiple linear regression
  • strategies for visualizing and interpreting multiple linear regression models of \(Y\) vs 2 predictors, 1 quantitative and 1 categorical

Readings and videos

Today is a day to discover ideas, so no readings or videos to go through before class.

Motivation

EXAMPLE 1

Let’s explore some data on penguins. First, enter install.packages("palmerpenguins") in the console (not Rmd). Then load the penguins data. You can find a codebook for these data by typing ?penguins in your console (not qmd).

# Load packages
library(tidyverse)
data(penguins)
penguins <- penguins %>% 
  filter(species != "Adelie", bill_len < 57)

# Check it out
head(penguins)
##   species island bill_len bill_dep flipper_len body_mass    sex year
## 1  Gentoo Biscoe     46.1     13.2         211      4500 female 2007
## 2  Gentoo Biscoe     50.0     16.3         230      5700   male 2007
## 3  Gentoo Biscoe     48.7     14.1         210      4450 female 2007
## 4  Gentoo Biscoe     50.0     15.2         218      5700   male 2007
## 5  Gentoo Biscoe     47.6     14.5         215      5400   male 2007
## 6  Gentoo Biscoe     46.5     13.5         210      4550 female 2007

Our goal is to build a model that we can use to get good predictions of penguins’ flipper (“arm”) lengths.

Consider 2 simple linear regression models of flipper_len by penguin sex and species:

# Model 1: R-squared = 0.1205
summary(lm(flipper_len ~ sex, penguins))
## 
## Call:
## lm(formula = flipper_len ~ sex, data = penguins)
## 
## Residuals:
##    Min     1Q Median     3Q    Max 
## -27.22 -10.22   3.78   8.78  17.37 
## 
## Coefficients:
##             Estimate Std. Error t value Pr(>|t|)    
## (Intercept)  205.220      1.197 171.434  < 2e-16 ***
## sexmale        8.408      1.679   5.007  1.3e-06 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 11.42 on 183 degrees of freedom
##   (4 observations deleted due to missingness)
## Multiple R-squared:  0.1205, Adjusted R-squared:  0.1157 
## F-statistic: 25.07 on 1 and 183 DF,  p-value: 1.297e-06

# Model 2: R-squared = 0.7014
summary(lm(flipper_len ~ species, penguins))
## 
## Call:
## lm(formula = flipper_len ~ species, data = penguins)
## 
## Residuals:
##     Min      1Q  Median      3Q     Max 
## -18.045  -5.045  -1.045   3.955  15.955 
## 
## Coefficients:
##               Estimate Std. Error t value Pr(>|t|)    
## (Intercept)   196.0448     0.8065  243.07   <2e-16 ***
## speciesGentoo  21.0372     1.0039   20.96   <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 6.602 on 187 degrees of freedom
## Multiple R-squared:  0.7014, Adjusted R-squared:  0.6998 
## F-statistic: 439.2 on 1 and 187 DF,  p-value: < 2.2e-16

How might we improve our predictions of flipper_len using only these 2 predictors? What do you think is a reasonable range of possible values for the new R-squared?

EXAMPLE 2

Consider a simple linear regression model of flipper_len by bill_len:

penguins %>% 
  ggplot(aes(y = flipper_len, x = bill_len)) + 
  geom_point() + 
  geom_smooth(method = "lm", se = FALSE)

Thoughts? What’s going on here? How does this highlight the limitations of a simple linear regression model?

EXAMPLE 3

The cps dataset contains employment information collected by the U.S. Current Population Survey (CPS) in 2018. We can use these data to explore wages among 18-34 year olds. The original codebook is here.

# Import data
cps <- read_csv("https://mac-stat.github.io/data/cps_2018.csv") %>% 
  select(-education, -hours) %>% 
  filter(age >= 18, age <= 34) %>% 
  filter(wage < 250000)
# Check it out
head(cps)
## # A tibble: 6 × 6
##    wage   age marital industry   health    education_level
##   <dbl> <dbl> <chr>   <chr>      <chr>     <chr>          
## 1 75000    33 single  management fair      bachelors      
## 2 33000    19 single  management very_good bachelors      
## 3 43000    33 married management good      bachelors      
## 4 50000    32 single  management excellent HS             
## 5 14400    28 single  service    excellent HS             
## 6 33000    31 married management very_good bachelors

We can use a simple linear regression model to summarize the relationship of wage with marital status:

# Build the model
wage_mod <- lm(wage ~ marital, data = cps)

# Summarize the model
coef(summary(wage_mod))
##                Estimate Std. Error   t value     Pr(>|t|)
## (Intercept)    46145.23    921.062  50.10002 0.000000e+00
## maritalsingle -17052.37   1127.177 -15.12839 5.636068e-50

What do you / don’t you conclude from this model? How does it highlight the limitations of a simple linear regression model?

Reflection: Why are multiple regression models so useful?

We can put more than 1 predictor into a regression model! Adding predictors to models…

  • Predictive viewpoint: Helps us better predict the response
  • Descriptive viewpoint: Helps us better understand the isolated (causal) effect of a variable by holding constant confounders

Multiple linear regression model formula

In general, a multiple linear regression model of \(Y\) with multiple predictors \((X_1, X_2, ..., X_p)\) is represented by the following formula:

\[E[Y \mid X_1, X_2, ..., X_p] = \beta_0 + \beta_1 X_1 + \beta_2 X_2 + ... + \beta_p X_p\]

Exercises

In the coming weeks, we’ll explore how to visualize, interpret, build, and evaluate multiple linear regression models.

First, we’ll explore some foundations using a model of penguin flipper_len by just 2 predictors: bill_len (quantitative) and species (categorical).

Exercise 1: Visualizing the relationship

We’ve learned how to visualize the relationship of flipper_len by bill_len alone:

penguins %>% 
  ggplot(aes(y = flipper_len, x = bill_len)) + 
  geom_point()

  1. How might we change the scatterplot points to also indicate information about penguin species?
penguins %>% 
  ggplot(aes(y = flipper_len, x = bill_len, color = species)) + 
  geom_point()

penguins %>% 
  ggplot(aes(y = flipper_len, x = bill_len, shape = species)) + 
  geom_point()

Exercise 2: Visualizing the model

We’ve also learned that a simple linear regression model of flipper_len by bill_len alone can be represented by a line:

penguins %>% 
  ggplot(aes(y = flipper_len, x = bill_len)) + 
  geom_point() + 
  geom_smooth(method = "lm", se = FALSE)

  1. Reflecting on your plot of flipper_len by bill_len and species in Exercise 1, how do you think a multiple regression model of flipper_len using both of these predictors would be represented? Check your intuition below by modifying the code below to include species in this plot, as you did in Exercise 1.
penguins %>% 
  ggplot(aes(y = flipper_len, x = bill_len, color = species)) + 
  geom_point() + 
  geom_smooth(method = "lm", se = FALSE)

Exercise 3: Intuition

Your plot in Exercise 2 demonstrated that the multiple linear regression model of flipper_len by bill_len and species is represented by 2 lines.

Let’s interpret the punchlines!

For each question, provide an answer along with evidence from the model lines that supports your answer.

  1. What’s the relationship between flipper_len and species, no matter a penguin’s bill_len?

Response: Put your response here.

  1. What’s the relationship between flipper_len and bill_len, no matter a penguin’s species?

Response: Put your response here.

  1. Does the rate of increase in flipper_len with bill_len differ between the two species?

Response: Put your response here.

Exercise 4: Model formula

Of course, there’s a formula behind the multiple regression model. We can obtain this using the usual lm() function.

# Build the model
penguin_mod <- lm(flipper_len ~ bill_len + species, data = penguins)

# Summarize the model
coef(summary(penguin_mod))
##                 Estimate Std. Error  t value     Pr(>|t|)
## (Intercept)   127.753693  6.1174521 20.88348 1.194472e-50
## bill_len        1.402367  0.1249665 11.22194 1.194932e-22
## speciesGentoo  22.848036  0.7938292 28.78206 1.981732e-70

  1. In the lm() function, how did we communicate that we wanted to model flipper_len by both bill_len and species? bill_len + species

  2. Complete the following model formula:

    E[flipper_len | bill_len, speciesGentoo] = 127.753693 + 1.402367 * bill_len + 22.848036 * speciesGentoo

Exercise 5: Sub-model formulas

Ok. We now have a single formula for the model.

And we observed earlier that this formula is represented by two lines: one describing the relationship between flipper_len and bill_len for Chinstrap penguins and the other for Gentoo penguins.

Let’s bring these ideas together.

Utilize the model formula to obtain the equations of these two lines, i.e. to obtain the sub-model formulas for the 2 species. Hint: Plug speciesGentoo = 1 and speciesGentoo = 0

Chinstrap: flipper_len = 127.753693 + 1.402367* bill_len

Gentoo: flipper_len = (127.753693 +22.848036) + 1.402367 * bill_len

Exercise 6: coefficients – physical interpretation

Reflecting on Exercise 5, let’s interpret what the model coefficients tell us about the physical properties of the two 2 sub-model lines. Choose the correct option given in parentheses:

  1. The intercept coefficient, 127.75, is the intercept of the line for (Chinstrap / Gentoo) penguins.

  2. The bill_len coefficient, 1.40, is the (intercept / slope) of both lines.

  3. The speciesGentoo coefficient, 22.85, indicates that the (intercept / slope) of the line for Gentoo is 22.85mm higher than the (intercept / slope) of the line for Chinstrap. Similarly, since the lines are parallel, the line for Gentoo is 22.85mm higher than the line for Chinstrap at any bill_len.

Exercise 7: coefficients – contextual interpretation

Next, interpret each coefficient in a contextually meaningful way. What do they tell us about penguin flipper lengths?!

  1. Interpret 127.75 (intercept of the Chinstrap line).

  2. Interpret 1.40 (slope of both lines). For both Chinstrap and Gentoo penguins, we expect…

  3. Interpret 22.85. At any bill_len, we expect…

Exercise 8: Prediction

Now that we better understand the model, let’s use it to predict flipper lengths! Recall the model summary and visualization:

coef(summary(penguin_mod))
##                 Estimate Std. Error  t value     Pr(>|t|)
## (Intercept)   127.753693  6.1174521 20.88348 1.194472e-50
## bill_len        1.402367  0.1249665 11.22194 1.194932e-22
## speciesGentoo  22.848036  0.7938292 28.78206 1.981732e-70

penguins %>% 
  ggplot(aes(y = flipper_len, x = bill_len, color = species)) + 
  geom_point() + 
  geom_smooth(method = "lm", se = FALSE)

  1. Predict the flipper length of a Chinstrap penguin with a 50mm long bill. Make sure your calculation is consistent with the plot.
127.75 + 1.40*___ + 22.85*___
## Error in parse(text = input): <text>:1:16: unexpected input
## 1: 127.75 + 1.40*__
##                    ^
  1. Predict the flipper length of a Gentoo penguin with a 50mm long bill. Make sure your calculation is consistent with the plot.
127.75 + 1.40*___ + 22.85*___
## Error in parse(text = input): <text>:1:16: unexpected input
## 1: 127.75 + 1.40*__
##                    ^
  1. Use the predict() function to confirm your predictions in parts a and b.
# Confirm the calculation in part a
predict(penguin_mod,
        newdata = data.frame(bill_len = ___, species = "___"))

# Confirm the calculation in part b
predict(penguin_mod,
        newdata = data.frame(bill_len = ___, species = "___"))
## Error in parse(text = input): <text>:3:42: unexpected input
## 2: predict(penguin_mod,
## 3:         newdata = data.frame(bill_len = __
##                                             ^

Exercise 9: R-squared

Finally, recall that improving our predictions was one motivation for multiple linear regression (using 2 predictors instead of 1). To this end, consider the R-squared values of the simple linear regression models that use just one predictor at a time:

mod_bill <- lm(flipper_len ~ bill_len, data = penguins)
summary(mod_bill)
## 
## Call:
## lm(formula = flipper_len ~ bill_len, data = penguins)
## 
## Residuals:
##     Min      1Q  Median      3Q     Max 
## -30.441 -10.998   3.015   8.942  19.881 
## 
## Coefficients:
##             Estimate Std. Error t value Pr(>|t|)    
## (Intercept) 177.4971    13.6676  12.987   <2e-16 ***
## bill_len      0.6712     0.2850   2.355   0.0195 *  
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 11.91 on 187 degrees of freedom
## Multiple R-squared:  0.02881,    Adjusted R-squared:  0.02362 
## F-statistic: 5.548 on 1 and 187 DF,  p-value: 0.01954

mod_species <- lm(flipper_len ~ species, data = penguins)
summary(mod_species)
## 
## Call:
## lm(formula = flipper_len ~ species, data = penguins)
## 
## Residuals:
##     Min      1Q  Median      3Q     Max 
## -18.045  -5.045  -1.045   3.955  15.955 
## 
## Coefficients:
##               Estimate Std. Error t value Pr(>|t|)    
## (Intercept)   196.0448     0.8065  243.07   <2e-16 ***
## speciesGentoo  21.0372     1.0039   20.96   <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 6.602 on 187 degrees of freedom
## Multiple R-squared:  0.7014, Adjusted R-squared:  0.6998 
## F-statistic: 439.2 on 1 and 187 DF,  p-value: < 2.2e-16

  1. If you had to use only 1 of our 2 predictors, which would give the better predictions of flipper_len?

  2. What do you guess is the R-squared of our multiple regression model that uses both of these predictors? Why?

  3. Check your intuition. How does the R-squared of our multiple regression model compare to that of the 2 separate simple linear regression models?

summary(penguin_mod)
## 
## Call:
## lm(formula = flipper_len ~ bill_len + species, data = penguins)
## 
## Residuals:
##      Min       1Q   Median       3Q      Max 
## -15.4763  -3.2260  -0.1525   3.1581  15.5303 
## 
## Coefficients:
##               Estimate Std. Error t value Pr(>|t|)    
## (Intercept)   127.7537     6.1175   20.88   <2e-16 ***
## bill_len        1.4024     0.1250   11.22   <2e-16 ***
## speciesGentoo  22.8480     0.7938   28.78   <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Residual standard error: 5.112 on 186 degrees of freedom
## Multiple R-squared:  0.8219, Adjusted R-squared:   0.82 
## F-statistic: 429.3 on 2 and 186 DF,  p-value: < 2.2e-16





Solutions

Exercise 1: Visualizing the relationship

  1. There are multiple options!
penguins %>% 
  ggplot(aes(y = flipper_len, x = bill_len, color = species)) + 
  geom_point()


penguins %>% 
  ggplot(aes(y = flipper_len, x = bill_len, shape = species)) + 
  geom_point()

Exercise 2: Visualizing the model

penguins %>% 
  ggplot(aes(y = flipper_len, x = bill_len, color = species)) + 
  geom_point() + 
  geom_smooth(method = "lm", se = FALSE)

Exercise 3: Intuition

  1. Gentoo tend to have longer flippers.
  2. Flipper length is positively associated with bill length.
  3. No. the lines are parallel / have the same slopes.

Exercise 4: Model formula

  1. bill_len + species
  2. E[flipper_len | bill_len, speciesGentoo] = 127.75 + 1.40 * bill_len + 22.85 * speciesGentoo

Exercise 5: Sub-model formulas

Chinstrap: flipper_len = 127.75 + 1.40 bill_len

Gentoo: flipper_len = (127.75 + 22.85) + 1.40 bill_len = 150.6 + 1.40 bill_len

Exercise 6: coefficients – physical interpretation

  1. The intercept coefficient, 127.75, is the intercept of the line for Chinstrap penguins.

  2. The bill_len coefficient, 1.40, is the slope of both lines.

  3. The speciesGentoo coefficient, 22.85, indicates that the intercept of the line for Gentoo is 22.85mm higher than the intercept of the line for Chinstrap. Similarly, since the lines are parallel, the line for Gentoo is 22.85mm higher than the line for Chinstrap at any bill_len.

Exercise 7: coefficients – contextual interpretation

  1. For both Chinstrap and Gentoo penguins, average flipper lengths increase by 1.40mm for every additional mm in bill length.

  2. At any bill_len, the average flipper length for Gentoo penguins is 22.85mm longer than that for Chinstrap penguins.

Exercise 8: Prediction

# a
127.75 + 1.40*50 + 22.85*0
## [1] 197.75

# b
127.75 + 1.40*50 + 22.85*1
## [1] 220.6

# c
predict(penguin_mod,
        newdata = data.frame(bill_len = 50, 
                             species = "Chinstrap"))
##       1 
## 197.872
predict(penguin_mod,
        newdata = data.frame(bill_len = 50, 
                             species = "Gentoo"))
##        1 
## 220.7201

Exercise 9: R-squared

  1. species
  2. no wrong answer
  3. It’s higher than the R-squared when we use either predictor alone!