This is the latest in my series of screencasts demonstrating how to use the tidymodels packages, from just starting out to tuning more complex models with many hyperparameters. Today’s screencast walks through how to use bootstrap resampling, with this week’s #TidyTuesday dataset on CEO departures. ?

Here is the code I used in the video, for those who prefer reading instead of or in addition to video.

Explore data

Our modeling goal is to estimate how involuntary CEO departures are changing with time. Let’s start by reading in the data.

library(tidyverse)

departures_raw <- read_csv("https://raw.githubusercontent.com/rfordatascience/tidytuesday/master/data/2021/2021-04-27/departures.csv")

How are involuntary departures changing with time? What about the rest of the CEO departures?

departures_raw %>%
  filter(departure_code < 9) %>%
  mutate(involuntary = if_else(departure_code %in% 3:4, "involuntary", "other")) %>%
  filter(fyear > 1995, fyear < 2019) %>%
  count(fyear, involuntary) %>%
  ggplot(aes(fyear, n, color = involuntary)) +
  geom_line(size = 1.2, alpha = 0.5) +
  geom_point(size = 2) +
  geom_smooth(method = "lm", lty = 2) +
  scale_y_continuous(limits = c(0, NA)) +
  labs(x = NULL, y = "Number of CEO departures", color = NULL)

Looks like proportionally more departures are involuntary over time, but that is what we’ll work on estimating. Let’s create a data set to use for modeling.

departures <- departures_raw %>%
  filter(departure_code < 9) %>%
  mutate(involuntary = if_else(departure_code %in% 3:4, "involuntary", "other")) %>%
  filter(fyear > 1995, fyear < 2019)

departures

## # A tibble: 6,942 x 20
##    dismissal_datase… coname  gvkey fyear co_per_rol exec_fullname departure_code
##                <dbl> <chr>   <dbl> <dbl>      <dbl> <chr>                  <dbl>
##  1            559043 SONICB… 27903  2002         -1 L. Gregory B…              7
##  2                12 AMERIC…  1045  1997          1 Robert L. Cr…              5
##  3                13 AMERIC…  1045  2002          3 Donald J. Ca…              3
##  4                31 ABBOTT…  1078  1998          6 Duane L. Bur…              5
##  5                43 ADVANC…  1161  2001         11 Walter Jerem…              5
##  6                51 AETNA …  1177  1997         16 Ronald Edwar…              5
##  7                63 AHMANS…  1194  1997         22 Charles R. R…              7
##  8                65 AIR PR…  1209  2000         28 Harold A. Wa…              5
##  9                76 ALBERT…  1239  2007         34 Howard B. Be…              5
## 10                78 ALBERT…  1240  2000         38 Gary Glenn M…              3
## # … with 6,932 more rows, and 13 more variables: ceo_dismissal <dbl>,
## #   interim_coceo <chr>, tenure_no_ceodb <dbl>, max_tenure_ceodb <dbl>,
## #   fyear_gone <dbl>, leftofc <dttm>, still_there <chr>, notes <chr>,
## #   sources <chr>, eight_ks <chr>, cik <dbl>, _merge <chr>, involuntary <chr>

Bootstrapping a model

We can count up the two kinds of departures per financial year and fit the model once, for the whole data set.

library(broom)

df <- departures %>%
  count(fyear, involuntary) %>%
  pivot_wider(names_from = involuntary, values_from = n)

mod <- glm(cbind(involuntary, other) ~ fyear, data = df, family = "binomial")
summary(mod)

## 
## Call:
## glm(formula = cbind(involuntary, other) ~ fyear, family = "binomial", 
##     data = df)
## 
## Deviance Residuals: 
##     Min       1Q   Median       3Q      Max  
## -2.9858  -1.2075  -0.1947   0.7302   3.6816  
## 
## Coefficients:
##               Estimate Std. Error z value Pr(>|z|)    
## (Intercept) -33.236731   8.949722  -3.714 0.000204 ***
## fyear         0.015875   0.004459   3.560 0.000370 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## (Dispersion parameter for binomial family taken to be 1)
## 
##     Null deviance: 78.421  on 22  degrees of freedom
## Residual deviance: 65.722  on 21  degrees of freedom
## AIC: 200.86
## 
## Number of Fisher Scoring iterations: 4

tidy(mod, exponentiate = TRUE)

## # A tibble: 2 x 5
##   term        estimate std.error statistic  p.value
##   <chr>          <dbl>     <dbl>     <dbl>    <dbl>
## 1 (Intercept) 3.68e-15   8.95        -3.71 0.000204
## 2 fyear       1.02e+ 0   0.00446      3.56 0.000370

When we use exponentiate = TRUE, we get the model coefficients on the linear scale instead of the logistic scale.

What we want to do is fit a model like this a whole bunch of times, instead of just once. Let’s create bootstrap resamples.

library(rsample)

set.seed(123)
ceo_folds <- bootstraps(departures, times = 1e3)
ceo_folds

## # Bootstrap sampling 
## # A tibble: 1,000 x 2
##    splits              id           
##    <list>              <chr>        
##  1 <split [6942/2543]> Bootstrap0001
##  2 <split [6942/2557]> Bootstrap0002
##  3 <split [6942/2509]> Bootstrap0003
##  4 <split [6942/2554]> Bootstrap0004
##  5 <split [6942/2542]> Bootstrap0005
##  6 <split [6942/2530]> Bootstrap0006
##  7 <split [6942/2509]> Bootstrap0007
##  8 <split [6942/2553]> Bootstrap0008
##  9 <split [6942/2586]> Bootstrap0009
## 10 <split [6942/2625]> Bootstrap0010
## # … with 990 more rows

Now we need to make a function to count up the departures by year and type, fit our model, and return the coefficients we want.

fit_binom <- function(split) {
  df <- analysis(split) %>%
    count(fyear, involuntary) %>%
    pivot_wider(names_from = involuntary, values_from = n)

  mod <- glm(cbind(involuntary, other) ~ fyear, data = df, family = "binomial")
  tidy(mod, exponentiate = TRUE)
}

We can apply that function to all our bootstrap resamples with purrr::map().

boot_models <- ceo_folds %>% mutate(coef_info = map(splits, fit_binom))
boot_models

## # Bootstrap sampling 
## # A tibble: 1,000 x 3
##    splits              id            coef_info       
##    <list>              <chr>         <list>          
##  1 <split [6942/2543]> Bootstrap0001 <tibble [2 × 5]>
##  2 <split [6942/2557]> Bootstrap0002 <tibble [2 × 5]>
##  3 <split [6942/2509]> Bootstrap0003 <tibble [2 × 5]>
##  4 <split [6942/2554]> Bootstrap0004 <tibble [2 × 5]>
##  5 <split [6942/2542]> Bootstrap0005 <tibble [2 × 5]>
##  6 <split [6942/2530]> Bootstrap0006 <tibble [2 × 5]>
##  7 <split [6942/2509]> Bootstrap0007 <tibble [2 × 5]>
##  8 <split [6942/2553]> Bootstrap0008 <tibble [2 × 5]>
##  9 <split [6942/2586]> Bootstrap0009 <tibble [2 × 5]>
## 10 <split [6942/2625]> Bootstrap0010 <tibble [2 × 5]>
## # … with 990 more rows

Explore results

What did we find? We can compute bootstrap confidence intervals with int_pctl().

percentile_intervals <- int_pctl(boot_models, coef_info)
percentile_intervals

## # A tibble: 2 x 6
##   term          .lower .estimate      .upper .alpha .method   
##   <chr>          <dbl>     <dbl>       <dbl>  <dbl> <chr>     
## 1 (Intercept) 6.03e-23 0.0000273 0.000000246   0.05 percentile
## 2 fyear       1.01e+ 0 1.02      1.03          0.05 percentile

We can also visualize the results as well.

boot_models %>%
  unnest(coef_info) %>%
  filter(term == "fyear") %>%
  ggplot(aes(estimate)) +
  geom_vline(xintercept = 1, lty = 2, color = "gray50", size = 2) +
  geom_histogram() +
  labs(
    x = "Annual increase in involuntary CEO departures",
    title = "Over this time period, CEO departures are increasingly involuntary",
    subtitle = "Each passing year corresponds to a departure being 1-2% more likely to be involuntary"
  )