4  Modelling and regression

4.1 R packages needed for this chapter

library(easystats)
library(tidyverse)
library(rstanarm)  # optional!

4.2 What’s modelling?

Read this great introduction by modelling by Russel Poldrack. Actually, the whole book is nice Poldrack (2022).

An epitome of modelling is this, let’s call it the fundamental modelling equation, a bit grandiose but at the point, see Equation 4.1.

The data can be separated in the model’s prediction and the rest (the “error”), i.e., what’s unaccounted for by the model.

\[ \text{data} = \text{model} + \text{error} \tag{4.1}\]

A more visual account of our basic modelling equation is depicted in Figure 4.1.

flowchart LR
  X --> Y
  error --> Y
Figure 4.1: A more visual account of our basic modelling equation

4.3 Regression as the umbrella tool for modelling

One regression

Source: Image Flip

Alternatively, venture into the forest of statistical tests as outlined e.g. here, at Uni Muenster. Proceed at your own peril.

You may want to ponder on this image of a decision tree of which test to choose, see Figure Figure 4.2.

Figure 4.2: Choose your test carefully

4.3.1 Common statistical tests are linear models

As Jonas Kristoffer Lindeløv tells us, we can formulate most statistical tests as a linear model, ie., a regression.

Common statistical tests as linear models

4.3.2 How to find the regression line

In the simplest case, regression analyses can be interpreted geometrically as a line in a 2D coordinate system, see Figure Figure 4.3.

Figure 4.3: Least Square Regression

Source: Orzetoo, CC-SA, Wikimedia

Put simple, we are looking for the line which is in the “middle of the points”. More precisely, we place the line such that the squared distances from the line to the points is minimal, see Figre Figure 4.3.

Consider Figure Figure 4.4, from this source by Roback & Legler (2021). It visualizes not only the notorious regression line, but also sheds light on regression assumptions, particularly on the error distribution.

Figure 4.4: Regression and some of its assumptions

Among the assumptions of the linear model are:

  • linearity of the function
  • variance of \(y\) remains constant across range of \(x\)
  • normality of residuals

4.3.3 The linear model

Here’s the canonical form of the linear model.

Consider a model with \(k\) predictors:

\[y = \beta_0 + \beta_1 x_1 + \ldots + \beta_k x_k + \epsilon\]

4.3.4 Algebraic derivation

For the mathematical inclined, check out this derivation of the simple case regression model. Note that the article is written in German, but your browser can effortlessly translate into English. Here’s a similar English article from StackExchange.

4.4 In all its glory

Let’s depict the residuals, s. Figure 4.5.

Figure 4.5: Residuals as deviations from the predicted value

4.5 First model: one metric predictor

First, let’s load some data:

data(mtcars)
glimpse(mtcars)
Rows: 32
Columns: 11
$ mpg  <dbl> 21.0, 21.0, 22.8, 21.4, 18.7, 18.1, 14.3, 24.4, 22.8, 19.2, 17.8,…
$ cyl  <dbl> 6, 6, 4, 6, 8, 6, 8, 4, 4, 6, 6, 8, 8, 8, 8, 8, 8, 4, 4, 4, 4, 8,…
$ disp <dbl> 160.0, 160.0, 108.0, 258.0, 360.0, 225.0, 360.0, 146.7, 140.8, 16…
$ hp   <dbl> 110, 110, 93, 110, 175, 105, 245, 62, 95, 123, 123, 180, 180, 180…
$ drat <dbl> 3.90, 3.90, 3.85, 3.08, 3.15, 2.76, 3.21, 3.69, 3.92, 3.92, 3.92,…
$ wt   <dbl> 2.620, 2.875, 2.320, 3.215, 3.440, 3.460, 3.570, 3.190, 3.150, 3.…
$ qsec <dbl> 16.46, 17.02, 18.61, 19.44, 17.02, 20.22, 15.84, 20.00, 22.90, 18…
$ vs   <dbl> 0, 0, 1, 1, 0, 1, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0,…
$ am   <dbl> 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0,…
$ gear <dbl> 4, 4, 4, 3, 3, 3, 3, 4, 4, 4, 4, 3, 3, 3, 3, 3, 3, 4, 4, 4, 3, 3,…
$ carb <dbl> 4, 4, 1, 1, 2, 1, 4, 2, 2, 4, 4, 3, 3, 3, 4, 4, 4, 1, 2, 1, 1, 2,…

4.5.1 Frequentist

Define and fit the model:

lm1_freq <- lm(mpg ~ hp, data = mtcars)

Get the parameter values:

parameters(lm1_freq)
Parameter Coefficient SE CI CI_low CI_high t df_error p
(Intercept) 30.0988605 1.6339210 0.95 26.7619488 33.4357723 18.421246 30 0e+00
hp -0.0682283 0.0101193 0.95 -0.0888947 -0.0475619 -6.742388 30 2e-07

Plot the model parameters:

plot(parameters(lm1_freq))

4.5.2 Bayesian

lm1_bayes <- stan_glm(mpg ~ hp, data = mtcars)

SAMPLING FOR MODEL 'continuous' NOW (CHAIN 1).
Chain 1: 
Chain 1: Gradient evaluation took 0.003812 seconds
Chain 1: 1000 transitions using 10 leapfrog steps per transition would take 38.12 seconds.
Chain 1: Adjust your expectations accordingly!
Chain 1: 
Chain 1: 
Chain 1: Iteration:    1 / 2000 [  0%]  (Warmup)
Chain 1: Iteration:  200 / 2000 [ 10%]  (Warmup)
Chain 1: Iteration:  400 / 2000 [ 20%]  (Warmup)
Chain 1: Iteration:  600 / 2000 [ 30%]  (Warmup)
Chain 1: Iteration:  800 / 2000 [ 40%]  (Warmup)
Chain 1: Iteration: 1000 / 2000 [ 50%]  (Warmup)
Chain 1: Iteration: 1001 / 2000 [ 50%]  (Sampling)
Chain 1: Iteration: 1200 / 2000 [ 60%]  (Sampling)
Chain 1: Iteration: 1400 / 2000 [ 70%]  (Sampling)
Chain 1: Iteration: 1600 / 2000 [ 80%]  (Sampling)
Chain 1: Iteration: 1800 / 2000 [ 90%]  (Sampling)
Chain 1: Iteration: 2000 / 2000 [100%]  (Sampling)
Chain 1: 
Chain 1:  Elapsed Time: 0.066 seconds (Warm-up)
Chain 1:                0.063 seconds (Sampling)
Chain 1:                0.129 seconds (Total)
Chain 1: 

SAMPLING FOR MODEL 'continuous' NOW (CHAIN 2).
Chain 2: 
Chain 2: Gradient evaluation took 2.2e-05 seconds
Chain 2: 1000 transitions using 10 leapfrog steps per transition would take 0.22 seconds.
Chain 2: Adjust your expectations accordingly!
Chain 2: 
Chain 2: 
Chain 2: Iteration:    1 / 2000 [  0%]  (Warmup)
Chain 2: Iteration:  200 / 2000 [ 10%]  (Warmup)
Chain 2: Iteration:  400 / 2000 [ 20%]  (Warmup)
Chain 2: Iteration:  600 / 2000 [ 30%]  (Warmup)
Chain 2: Iteration:  800 / 2000 [ 40%]  (Warmup)
Chain 2: Iteration: 1000 / 2000 [ 50%]  (Warmup)
Chain 2: Iteration: 1001 / 2000 [ 50%]  (Sampling)
Chain 2: Iteration: 1200 / 2000 [ 60%]  (Sampling)
Chain 2: Iteration: 1400 / 2000 [ 70%]  (Sampling)
Chain 2: Iteration: 1600 / 2000 [ 80%]  (Sampling)
Chain 2: Iteration: 1800 / 2000 [ 90%]  (Sampling)
Chain 2: Iteration: 2000 / 2000 [100%]  (Sampling)
Chain 2: 
Chain 2:  Elapsed Time: 0.064 seconds (Warm-up)
Chain 2:                0.058 seconds (Sampling)
Chain 2:                0.122 seconds (Total)
Chain 2: 

SAMPLING FOR MODEL 'continuous' NOW (CHAIN 3).
Chain 3: 
Chain 3: Gradient evaluation took 2.2e-05 seconds
Chain 3: 1000 transitions using 10 leapfrog steps per transition would take 0.22 seconds.
Chain 3: Adjust your expectations accordingly!
Chain 3: 
Chain 3: 
Chain 3: Iteration:    1 / 2000 [  0%]  (Warmup)
Chain 3: Iteration:  200 / 2000 [ 10%]  (Warmup)
Chain 3: Iteration:  400 / 2000 [ 20%]  (Warmup)
Chain 3: Iteration:  600 / 2000 [ 30%]  (Warmup)
Chain 3: Iteration:  800 / 2000 [ 40%]  (Warmup)
Chain 3: Iteration: 1000 / 2000 [ 50%]  (Warmup)
Chain 3: Iteration: 1001 / 2000 [ 50%]  (Sampling)
Chain 3: Iteration: 1200 / 2000 [ 60%]  (Sampling)
Chain 3: Iteration: 1400 / 2000 [ 70%]  (Sampling)
Chain 3: Iteration: 1600 / 2000 [ 80%]  (Sampling)
Chain 3: Iteration: 1800 / 2000 [ 90%]  (Sampling)
Chain 3: Iteration: 2000 / 2000 [100%]  (Sampling)
Chain 3: 
Chain 3:  Elapsed Time: 0.066 seconds (Warm-up)
Chain 3:                0.06 seconds (Sampling)
Chain 3:                0.126 seconds (Total)
Chain 3: 

SAMPLING FOR MODEL 'continuous' NOW (CHAIN 4).
Chain 4: 
Chain 4: Gradient evaluation took 1.7e-05 seconds
Chain 4: 1000 transitions using 10 leapfrog steps per transition would take 0.17 seconds.
Chain 4: Adjust your expectations accordingly!
Chain 4: 
Chain 4: 
Chain 4: Iteration:    1 / 2000 [  0%]  (Warmup)
Chain 4: Iteration:  200 / 2000 [ 10%]  (Warmup)
Chain 4: Iteration:  400 / 2000 [ 20%]  (Warmup)
Chain 4: Iteration:  600 / 2000 [ 30%]  (Warmup)
Chain 4: Iteration:  800 / 2000 [ 40%]  (Warmup)
Chain 4: Iteration: 1000 / 2000 [ 50%]  (Warmup)
Chain 4: Iteration: 1001 / 2000 [ 50%]  (Sampling)
Chain 4: Iteration: 1200 / 2000 [ 60%]  (Sampling)
Chain 4: Iteration: 1400 / 2000 [ 70%]  (Sampling)
Chain 4: Iteration: 1600 / 2000 [ 80%]  (Sampling)
Chain 4: Iteration: 1800 / 2000 [ 90%]  (Sampling)
Chain 4: Iteration: 2000 / 2000 [100%]  (Sampling)
Chain 4: 
Chain 4:  Elapsed Time: 0.062 seconds (Warm-up)
Chain 4:                0.057 seconds (Sampling)
Chain 4:                0.119 seconds (Total)
Chain 4:

Actually, we want to suppress some overly verbose output of the sampling, so add the argument refresh = 0:

lm1_bayes <- stan_glm(mpg ~ hp, data = mtcars, refresh = 0)

Get the parameter values:

parameters(lm1_bayes)
Parameter Median CI CI_low CI_high pd Rhat ESS Prior_Distribution Prior_Location Prior_Scale
(Intercept) 30.0590424 0.95 26.7439165 33.3124516 1 1.000771 2982.289 normal 20.09062 15.0673701
hp -0.0679057 0.95 -0.0882979 -0.0472594 1 1.000275 3533.338 normal 0.00000 0.2197599

Plot the model parameters:

plot(parameters(lm1_bayes))

4.5.3 Model performance

r2(lm1_freq)
# R2 for Linear Regression
       R2: 0.602
  adj. R2: 0.589
r2(lm1_bayes)
# Bayesian R2 with Compatibility Interval

  Conditional R2: 0.583 (95% CI [0.379, 0.753])

4.5.4 Model check

Here’s a bunch of typical model checks in the Frequentist sense.

check_model(lm1_freq)

And here are some Bayesian flavored model checks.

check_model(lm1_bayes)

4.5.5 Get some predictions

lm1_pred <- estimate_relation(lm1_freq)
lm1_pred
hp Predicted SE CI_low CI_high
52.00000 26.550990 1.1766139 24.148024 28.95396
83.44444 24.405590 0.9358933 22.494241 26.31694
114.88889 22.260189 0.7548971 20.718484 23.80190
146.33333 20.114789 0.6828911 18.720139 21.50944
177.77778 17.969389 0.7518697 16.433866 19.50491
209.22222 15.823989 0.9310065 13.922620 17.72536
240.66667 13.678588 1.1707841 11.287528 16.06965
272.11111 11.533188 1.4412478 8.589767 14.47661
303.55556 9.387788 1.7280486 5.858642 12.91693
335.00000 7.242387 2.0242544 3.108308 11.37647

More details on the above function can be found on the respective page at the easystats site.

4.5.6 Plot the model

plot(lm1_pred)

4.6 More of this

More technical details for gauging model performance and model quality, can be found on the site of the R package “performance at the easystats site.

4.7 Lab

Run a simple regression on your own research data. Present the results. Did you encounter any glitches?

4.8 Bayes-members only

Bayes statistics provide a distribution as the result of the analysis, the posterior distribution, which provides us with quite some luxury.

As the posterior distribution manifests itself by a number of samples, we can easily filter and manipulate this sample distribution in order to ask some interesing questions.

See

lm1_bayes_tibble <- as_tibble(lm1_bayes)  # cast as a tibble (table)

head(lm1_bayes_tibble)  # show the first few rows
(Intercept) hp sigma
29.24798 -0.0585108 4.148574
31.44328 -0.0793732 3.583081
31.66962 -0.0810650 3.406428
29.41583 -0.0611019 4.993273
28.85437 -0.0658251 3.520692
31.20567 -0.0744505 4.846585

4.8.1 Asking for probabilites

What’s the probability that the effect of hp is negative?

lm1_bayes_tibble %>% 
  count(hp < 0)
hp < 0 n
TRUE 4000

Feel free to ask similar questions!

4.8.2 Asking for quantiles

With a given probability of, say 90%, how large is the effect of hp?

quantile(lm1_bayes_tibble$hp, .9)
       90% 
-0.0547369

What’s the smallest 95% percent interval for the effect of hp?

hdi(lm1_bayes)
Parameter CI CI_low CI_high Effects Component
(Intercept) 0.95 26.8254064 33.3821880 fixed conditional
hp 0.95 -0.0874363 -0.0469058 fixed conditional

In case you prefer 89% intervals (I do!):

hdi(lm1_bayes, ci = .89)
Parameter CI CI_low CI_high Effects Component
(Intercept) 0.89 27.4210227 32.6653275 fixed conditional
hp 0.89 -0.0847388 -0.0522295 fixed conditional

4.8.3 Model specification

In Bayes statistics, it is customary to specify the model in something like the following way:

\[\begin{aligned} y_i &\sim N(\mu_i,\sigma)\\ \mu_i &= \beta_0 + \beta_1 x_i\\ \beta_0, \beta_1 &\sim N(0, 1) \\ \sigma &\sim E(1) \end{aligned}\]

In this specification, \(N\) refers to the normal distribution, and \(E\) to the exponential distribution. Furthermore, this model assumes that the \(X\) and \(Y\) are given in standard units.

4.8.4 Prediction interval

A prediction interval answers the following question

How large is the uncertainty in \(y\) associated with a given obersation? What interval of values should I expect for a randomly chosen observation?

For example, what’s the uncertainty attached to the fuel economy of a car with 100 hp?

estimate_prediction(model = lm1_bayes, 
                    data = tibble(hp = 100) )
hp Predicted SE CI_low CI_high
100 23.26178 4.156918 15.40819 31.53782

4.8.5 … And more

We could even ask intriguing questions such as

Given the model, and given two random observations, one from the experimental group and one from the control group, what is the probability that observation 1 has a higher value in \(Y\) than observation 2 has?

Note that we are not only asking for “typical” observations as predicted by the model but we are also taking into account the uncertainty of the prediction for each group. Hence, this kind of questions is likely to yield more realistic (and less clear-cut) answers than just asking for the typical value. In other words, such analyses draw on the posterior predictive distribution.

4.9 Multiple metric predictors

Assume we have a theory that dictates that fuel economy is a (causal) function of horse power and engine displacement.

lm2_freq <- lm(mpg ~ hp + disp, data = mtcars)
parameters(lm2_freq)
Parameter Coefficient SE CI CI_low CI_high t df_error p
(Intercept) 30.7359042 1.3315661 0.95 28.0125457 33.4592628 23.082522 29 0.0000000
hp -0.0248401 0.0133855 0.95 -0.0522165 0.0025363 -1.855746 29 0.0736791
disp -0.0303463 0.0074049 0.95 -0.0454909 -0.0152016 -4.098159 29 0.0003063

Similarly for Bayes inference:

lm2_bayes <- stan_glm(mpg ~ hp + disp, data = mtcars)

Results

parameters(lm2_bayes)
Parameter Median CI CI_low CI_high pd Rhat ESS Prior_Distribution Prior_Location Prior_Scale
(Intercept) 30.7393959 0.95 27.9953578 33.4954849 1.000 0.9999147 4336.162 normal 20.09062 15.0673701
hp -0.0243992 0.95 -0.0525050 0.0036740 0.957 1.0021956 2041.955 normal 0.00000 0.2197599
disp -0.0307138 0.95 -0.0458102 -0.0153927 1.000 1.0014999 2001.424 normal 0.00000 0.1215712 
plot(parameters(lm2_bayes))

r2(lm2_bayes)
# Bayesian R2 with Compatibility Interval

  Conditional R2: 0.733 (95% CI [0.580, 0.845])

Depending on the value of disp the prediction of mpg from hp will vary:

lm2_pred <- estimate_relation(lm2_freq)
plot(lm2_pred)

4.10 One nominal predictor

lm3a <- lm(mpg ~ am, data = mtcars)
parameters(lm3a)
Parameter Coefficient SE CI CI_low CI_high t df_error p
(Intercept) 17.147368 1.124602 0.95 14.85062 19.44411 15.247492 30 0.000000
am 7.244939 1.764422 0.95 3.64151 10.84837 4.106127 30 0.000285 
lm3a_means <- estimate_means(lm3a, at = "am = c(0, 1)")
lm3a_means
am Mean SE CI_low CI_high
0 17.14737 1.124602 14.85062 19.44411
1 24.39231 1.359578 21.61568 27.16894

If we were not to specify the values of am which we would like to get predictions for, the default of the function would select 10 values, spread across the range of am. For numeric variables, this is usually fine. However, for nominal variables - and am is in fact a nominally scaled variable - we insist that we want predictions for the levels of the variable only, that is for 0 and 1.

However, unfortunately, the plot needs a nominal variable if we are to compare groups. In our case, am is considered a numeric variables, since it consists of numbers only. The plot does not work, malheureusement:

plot(lm3a_means)
Error in `rlang::sym()`:
! Can't convert a character vector to a symbol.

We need to transform am to a factor variable. That’s something like a string. If we hand over a factor() to the plotting function, everything will run smoothly. Computationwise, no big differences:

mtcars2 <-
  mtcars %>% 
  mutate(am_f = factor(am))

lm3a <- lm(mpg ~ am_f, data = mtcars2)
parameters(lm3a)
Parameter Coefficient SE CI CI_low CI_high t df_error p
(Intercept) 17.147368 1.124602 0.95 14.85062 19.44411 15.247492 30 0.000000
am_f1 7.244939 1.764422 0.95 3.64151 10.84837 4.106127 30 0.000285 
lm3a_means <- estimate_means(lm3a)
plot(lm3a_means)

Note that we should have converted am to a factor variable before fitting the model. Otherwise, the plot won’t work.

Here’s a more hand-crafted version of the last plot, see Fig. Figure 4.6.

ggplot(mtcars2) +
  aes(x = am_f, y = mpg) +
  geom_violin() +
  geom_jitter(width = .1, alpha = .5) +
  geom_pointrange(data = lm3a_means,
                  color = "orange",
                  aes(ymin = CI_low, ymax = CI_high, y = Mean)) +
  geom_line(data = lm3a_means, aes(y = Mean, group = 1))
Figure 4.6: Means per level of am

4.11 One metric and one nominal predictor

mtcars2 <-
  mtcars %>% 
  mutate(cyl = factor(cyl))

lm4 <- lm(mpg ~ hp + cyl, data = mtcars2)
parameters(lm4)
Parameter Coefficient SE CI CI_low CI_high t df_error p
(Intercept) 28.6501182 1.5877870 0.95 25.3976840 31.9025524 18.044056 28 0.0000000
hp -0.0240388 0.0154079 0.95 -0.0556005 0.0075228 -1.560163 28 0.1299540
cyl6 -5.9676551 1.6392776 0.95 -9.3255631 -2.6097471 -3.640418 28 0.0010921
cyl8 -8.5208508 2.3260749 0.95 -13.2855993 -3.7561022 -3.663188 28 0.0010286 
lm4_pred <- estimate_relation(lm4)
plot(lm4_pred)

4.12 Watch out for Simpson

Beware! Model estimates can swing wildly if you add (or remove) some predictor from your model. See this post for an demonstration.

4.13 What about correlation?

Correlation is really a close cousin to regression. In fact, regression with standardized variables amounts to correlation.

Let’s get the correlation matrix of the variables in involved in lm4.

lm4_corr <- 
  mtcars %>% 
  select(mpg, hp, disp) %>% 
  correlation()

lm4_corr
Parameter1 Parameter2 r CI CI_low CI_high t df_error p Method n_Obs
mpg hp -0.7761684 0.95 -0.8852686 -0.5860994 -6.742388 30 2e-07 Pearson correlation 32
mpg disp -0.8475514 0.95 -0.9233594 -0.7081376 -8.747151 30 0e+00 Pearson correlation 32
hp disp 0.7909486 0.95 0.6106794 0.8932775 7.080122 30 1e-07 Pearson correlation 32 
plot(summary(lm4_corr))

4.14 Exercises

  1. mtcars simple 1
  2. mtcars simple 2
  3. mtcars simple 3

🧑‍🎓 I want more!

👨‍🏫 Checkout all exercises tagged with “regression” on datenwerk. Pro-Tipp: Use the translation function of your browers to translate the webpage into your favorite language.

4.15 Case study

4.16 Lab

Get your own data, and build a simple model reflecting your research hypothesis. If you are lacking data (or hypothesis) get something close to it.

4.17 Going further

An accessible treatment of regression is provided by Ismay & Kim (2020).

Roback & Legler (2021) provide a more than introductory account of regression while being accessible. A recent but already classic book (if this is possible) is the book by Gelman et al. (2021). You may also benefit from Poldrack (2022) (open access).

For a gentle introduction to the basics of modelling, see ModernDive Chap. 5.0 (Ismay & Kim, 2020), and get the R code here. For slightly more advanced topics on linear regression such as multiple regression and interaction, see ModernDive Chap. 6, and get the R code here.

4.18 Debrief

Science!