---
title: "Interactions in Linear and Logistic Models"
subtitle: |
  | Additional Topic 
  | Math 219
author: "Yurk"
format: 
  revealjs:
    theme: beige
editor: source
---



```{r}
#| include: false
#| echo: false

library(tidyverse)
library(infer)
library(openintro)
library(broom)
library(HH)
library(GLMsData)
```

## Iris

-   The `iris` dataset has 150 observations of 5 variables
-   Previously we fit a parallel-slopes / additive model to the relationship between petal width and petal length for 3 different species, *Iris setosa*, *Iris versicolor*, and *Iris virginica*

```{r}
#| include: TRUE
#| echo: TRUE

lm(Petal.Length ~ Petal.Width + Species, data = iris) |>
  tidy()
```

------------------------------------------------------------------------

```{r}
#| include: true
#| echo: false
#| fig-cap: "Scatter plot of petal length vs. petal width colored by species, along with parallel slopes model."

par_ir_lines <- tibble(b0 = c(1.21, 2.91, 3.49), b1 = c(1.02, 1.02, 1.02),
                      Species = c("setosa", "versicolor", "virginica"))

iris |>
  ggplot(aes(Petal.Width, Petal.Length, color = Species)) +
  geom_point() +
  geom_abline(data = par_ir_lines,
              mapping = aes(slope = b1, intercept = b0, color = Species)) +
  theme_minimal()
```

::: {style="font-size: 25px"}
$$\widehat{Petal.Length}=\left\{\begin{array}{cl}1.21+1.02\times Petal.Width, & \textrm{if } Species = ``setosa''\\2.91+1.02\times Petal.Width, & \textrm{if } Species = ``versicolor''\\3.49+1.02\times Petal.Width, & \textrm{if } Species = ``virginica''\end{array}\right.$$
:::

## Interaction

- We can include an interaction between species and petal width in the model
- There is an interaction if different species have different relationships between the response (petal length) and petal width
- Unlike the additive model, a model with an interaction has a different slope for each species

---

- There is a significant interaction between species and petal width

```{r}
#| include: TRUE
#| echo: TRUE

lm(Petal.Length ~ Petal.Width * Species, data = iris) |>
  tidy()
```

---

```{r}
#| include: true
#| echo: false
#| fig-cap: "Scatter plot of petal length vs. petal width colored by species, along with model with interaction."

iris |>
  ggplot(aes(Petal.Width, Petal.Length, color = Species)) +
  geom_point() +
  geom_smooth(method = "lm", se = FALSE) +
  theme_minimal()
```

::: {style="font-size: 25px"}
$$\widehat{Petal.Length}=\left\{\begin{array}{cl}1.33+0.55\times Petal.Width, & \textrm{if } Species = ``setosa''\\1.78+1.09\times Petal.Width, & \textrm{if } Species = ``versicolor''\\4.24+0.65\times Petal.Width, & \textrm{if } Species = ``virginica''\end{array}\right.$$
:::

## Palmer Penguins

```{r}
#| include: false
#| echo: false

library(palmerpenguins)
data(penguins, package = "palmerpenguins")
penguins <- penguins |>
  dplyr::select(-c(island, year)) |>
  drop_na()
```

- `penguins` dataset [^4]
- Measurements for three species of penguins from Palmer Archipelago
- Previously, we considered an additive model predicting body mass using bill depth, flipper length, and sex

[^4]: The `penguins` dataset is from the `palmerpenguins` package.

## Three-way interaction

- With three predictors we can evaluate whether or not there is evidence of a **three-way interaction**
- A three-way interaction is more difficult to interpret than a two-way interaction

::: {style="font-size: 35px"}

```{r}
#| include: true
#| echo: true

lm(body_mass_g ~ bill_depth_mm * sex * flipper_length_mm, data = penguins) |> 
  tidy()
```

:::

---

- The three-way interaction is not significant, so we drop it from the model and consider the possible two-way interactions
- Of these, the only two-way interaction that is significant is between flipper length and bill depth
- Drop the others from the model

```{r}
lm(body_mass_g ~ bill_depth_mm * sex +  flipper_length_mm * sex + bill_depth_mm * flipper_length_mm, data = penguins) |> 
  tidy()
```

---

- The negative coefficient for the interaction between flipper length and bill depth indicates that the rate at which body mass increases with bill depth decreases as flipper length increases

```{r}
lm(body_mass_g ~ sex + bill_depth_mm * flipper_length_mm, data = penguins) |> 
  tidy()
```

---

The model

::: {style="font-size: 30px"}

$$\begin{array}{rcl}\widehat{body\_mass\_g} &=& -28121 + 498\times sexmale \\ & & + 1434 \times bill\_depth\_mm \\ & & + 164\times flipper\_length\_mm \\ & & - 7.34 \times bill\_depth\_mm\times flipper\_length\_mm \end{array}$$
:::

can also be written as

::: {style="font-size: 30px"}

$$\begin{array}{rcl}\widehat{body\_mass\_g} &=& -28121 + 498\times sexmale \\ & & + 164\times flipper\_length\_mm \\ & & +(1434 - 7.34 \times flipper\_length\_mm)\times bill\_depth\_mm\end{array}$$

:::

## Discrimination in Hiring

-   Does perceived race or sex of an applicant affect job application callback rates?
-   Randomly assigned a name to each resume
-   Name implied applicant's race (Black or White) and sex (male or female)

---

- Previously we fit a logistic model to predict the probability of receiving a call back using job city, years experience, honors, and race
- All of these predictors are statistically significant, including race

```{r}
#| include: true
#| echo: true

glm(received_callback ~ job_city + years_experience + honors + race,
    family = binomial, data = resume) |>
  tidy()
```

---

- Is there evidence of an interaction between race and any of the other predictors?
- For example, does the relationship between the probability of receiving a call back and years of experience depend on the race of the applicant?
- Let's test all of the possible two-way interactions involving race

---

- We do not find convincing evidence of an interaction between race an any of the other predictors
- We would proceed using the earlier model without interactions

```{r}
#| include: true
#| echo: true

glm(received_callback ~ job_city*race + years_experience*race +
      honors*race, family = binomial, data = resume) |>
  tidy()
```