Interactions in Regression Models — Student Handout

Follow-along script and exercises

Author

Your name here

Published

October 5, 2026

Keywords

Allen Bush-Beaupré, ecological statistics, agricultural entomology, integrated pest management, Bayesian statistics, statistical consulting, GLMM, Université de Sherbrooke, CREUS, Bishop’s University

Before you start

Run the chunk below once to install everything you need. It’s written to avoid the most common glmmTMB installation failures:

  • Binary vs. source mismatch: glmmTMB links against TMB and Matrix, which must be built with compatible ABI versions. Installing all three together from the same repository snapshot (rather than upgrading one in isolation) avoids “package was installed with a different version” load-time errors.
  • Stale compiled objects after an R or Matrix upgrade: if you’ve upgraded R recently and see an error like Error: package or namespace load failed for 'glmmTMB' ... unable to load shared object, reinstalling TMB and glmmTMB from source with --no-multiarch (via install.packages(type = "source")) usually fixes it.
  • Missing compiler toolchain: source installs of TMB/glmmTMB need a working C++ compiler. On Windows this means Rtools; on macOS, Xcode command line tools. The check below tells you if compilation is unavailable before you hit a cryptic error.
required_pkgs <- c("tidyverse", "glmmTMB", "marginaleffects", "patchwork", "GGally",
                    "distributional", "ggdist", "tinytable", "performance", "TMB", "Matrix")

# Reinstalling Matrix/TMB/glmmTMB together (rather than one at a time) avoids the
# most common cause of glmmTMB load failures: an ABI mismatch between glmmTMB's
# compiled code and the Matrix/TMB versions it was linked against.
missing_pkgs <- required_pkgs[!vapply(required_pkgs, requireNamespace, logical(1), quietly = TRUE)]

if (length(missing_pkgs) > 0) {
  install.packages(missing_pkgs)
}

# Verify glmmTMB actually loads and can fit a trivial model - this is the step
# that catches ABI mismatches that "successful" installation can still leave behind.
glmmtmb_ok <- tryCatch({
  library(glmmTMB)
  test_fit <- glmmTMB(mpg ~ wt, data = mtcars)
  TRUE
}, error = function(e) {
  message("glmmTMB failed to load/fit: ", conditionMessage(e))
  FALSE
})

if (!glmmtmb_ok) {
  message("Reinstalling TMB and glmmTMB from source to resolve a likely ABI mismatch...")
  install.packages(c("TMB", "glmmTMB"), type = "source")

  # Re-check after the source reinstall
  glmmtmb_ok <- tryCatch({
    library(glmmTMB)
    glmmTMB(mpg ~ wt, data = mtcars)
    TRUE
  }, error = function(e) {
    message("Still failing: ", conditionMessage(e))
    message("If this persists, check that you have a working C++ compiler installed ",
            "(Rtools on Windows, Xcode command line tools on macOS), then restart R and re-run this chunk.")
    FALSE
  })
}

stopifnot("glmmTMB did not install/load successfully - see messages above" = glmmtmb_ok)

Setup

Run this at the start of every session (after the install chunk above has succeeded once).

library(tidyverse)       # Data manipulation and plotting
library(glmmTMB)         # Generalized linear (mixed) models
library(marginaleffects) # Compute marginal/conditional effects
library(patchwork)       # Combine multiple plots
library(GGally)          # Pairwise plots
library(distributional)  # Simulate mixture/gamma/normal distributions (Part 3)
library(ggdist)          # stat_slab() for predicted-distribution plots (Part 2)
library(tinytable)       # tt(): simple, nice-looking results tables
library(performance)     # r2(), r2_tjur(): model R-squared measures

# Set ggplot2 theme
black_theme <- theme(
  axis.line = element_line(linewidth = 2, lineend = "round", color = "white"),
  panel.grid = element_blank(),
  panel.background = element_rect(fill = "black", color = NA),
  axis.ticks = element_blank(),
  axis.text = element_text(size = 21, face = "bold", color = "white"),
  axis.title = element_text(size = 21, face = "bold", color = "white"),
  plot.title = element_text(size = 25, face = "bold", color = "white"),
  plot.background = element_rect(fill = "black", color = NA),
  legend.background = element_rect(fill = "black", color = NA),
  legend.text = element_text(size = 21, color = "white"),
  legend.title = element_text(size = 21, face = "bold", color = "white"),
  legend.position = "bottom"
)

theme_set(black_theme)

# Set global default for geom_point shape
update_geom_defaults("point", list(shape = 1, size = 3, alpha = 0.8))

# silence marginaleffects warnings about glmmTMB
options(marginaleffects_safe = FALSE)

set.seed(333)

Coding style & functions used throughout

Before diving into the case studies, here’s a quick reference for the conventions and workhorse functions used repeatedly in this handout.

The native pipe, |>. Wherever you see data |> mutate(...) |> ggplot(...), read it left-to-right as “take data, and then apply mutate(), and then pass the result to ggplot()”. It’s equivalent to nesting the calls (ggplot(mutate(data, ...))), just easier to read and edit in sequence. We use the base-R pipe (|>), not the older magrittr pipe (%>%) — for our purposes here they behave the same way.

Simulating data with tibble() and random-draw functions. Every case study starts by simulating data rather than loading a real dataset. We pick “true” parameter values (an intercept, one or more slopes) up front, then generate observations from a known random process:

  • rnorm(n, mean, sd) — draws from a Normal distribution (continuous outcomes, e.g. weight, petal length)
  • rbinom(n, size = 1, prob) — draws from a Bernoulli/Binomial distribution (presence/absence outcomes)
  • rpois(n, lambda) — draws from a Poisson distribution (count outcomes, e.g. abundance)
  • runif(n, min, max) — draws from a Uniform distribution (often used to simulate a predictor, not an outcome)
  • distributional::dist_*() + generate() — an alternative way to simulate predictors (e.g. dist_beta(), dist_gamma(), dist_mixture()) used in Part 3

Simulating this way means we know the ground truth (the exact intercept and slopes used to generate the data), so we can directly check whether a fitted model recovers it — a habit worth carrying into your own work whenever you want to sanity-check a modeling approach before applying it to real data.

Fitting models with glmmTMB(). Our model-fitting workhorse throughout is glmmTMB(response ~ predictors, data = ..., family = ...). Key things to know:

  • y ~ x1 * x2 expands to y ~ x1 + x2 + x1:x2 — main effects for both predictors and their interaction, all in one shorthand
  • family = gaussian() for continuous outcomes, binomial(link = "logit") for presence/absence, poisson(link = "log") for counts, truncated_poisson for a hurdle model’s count-when-present part
  • dispformula = ~ x lets predictors affect the variance of the outcome, not just its mean (used in the distributional-regression section of Part 2)
  • zi = ~ x adds a zero-inflation sub-model: a binomial part that models whether a value is a “structural zero” (used in the hurdle-model section of Part 3)

The marginaleffects package. Raw model coefficients (from summary()) are often not the quantity you actually want to report — especially with interactions or non-linear link functions. marginaleffects converts coefficients into direct, interpretable answers:

  • predictions() — the model-implied value of the outcome at specific covariate values (a conditional prediction)
  • avg_predictions() — the same, but averaged over a distribution of covariate values (a marginal prediction)
  • slopes() — the model-implied effect (derivative/slope) of one predictor on the outcome, at specific covariate values
  • avg_slopes() — the same, averaged over a distribution of covariate values
  • datagrid(...) — builds the covariate grid at which predictions()/slopes() are evaluated; arguments can be fixed values, functions like mean or fivenum (the five-number summary), or vectors of values
  • datagrid(..., grid_type = "counterfactual") — pairs each focal value with every observed row of the data, so avg_predictions()/avg_slopes() genuinely average over the observed distribution of the other predictors, instead of holding them fixed at a single value (e.g., their mean)
  • by = "group" — instead of collapsing (averaging) over a variable, keep results broken out by its levels
  • type = "response" vs. type = "link" — get a predictor’s effect on the natural (probability, count) scale vs. the scale the model is additive on; for hurdle/zero-inflated models, type = "conditional", "response", and "zprob" select which model component to report
  • hypotheses(hypothesis = "b1 - b2 = 0") — tests linear combinations of estimates, e.g. “is the slope for group A different from the slope for group B?”

set.seed(). Called once near the top of this script so that “random” simulation is actually reproducible — running the same code twice gives identical numbers. If your results differ from a classmate’s despite identical code, check whether you both used the same seed at the same point in the script.

Part 1 — Categorical × Numerical Interactions

What is an interaction?

An interaction means the effect of one predictor depends on the value of another.

  • No interaction: snow depth affects weight the same way for every sheep
  • Interaction: the effect of snow depth on weight differs by sex

For a categorical moderator (e.g., sex) and a continuous predictor (e.g., snow depth): the interaction means the slope differs between categories.

Conditional vs. marginal effects

  • Conditional effects: the effect of a predictor at specific values of other predictors (e.g., the slope of snow depth for males)
  • Marginal effects: the effect averaged across the distribution of other predictors (e.g., the average slope of snow depth across the whole population)

Both are legitimate; they answer different questions. Interactions are exactly where the distinction starts to matter.

Case study: Bighorn sheep

Ram Mountain, Alberta. Question: how does winter snow depth relate to spring mass in Bighorn sheep, and does it differ by sex? Males and females segregate in winter — females use lower, less snowy habitat.

n_individuals_population <- 1000
sexes <- c("male", "female")
male_intercept <- 220; female_intercept <- 180
sd_weight <- 15
male_loss <- -0.8; female_loss <- -2.5

bighorn_population <- tibble(
  sex = rep(sexes, n_individuals_population / length(sexes)),
  snow_depth = runif(n_individuals_population,
    min = ifelse(sex == "male", 35, 0),
    max = ifelse(sex == "male", 70, 35)
  ),
  weight = rnorm(n_individuals_population,
    mean = ifelse(sex == "male",
      male_intercept + male_loss * snow_depth,
      female_intercept + female_loss * snow_depth),
    sd = sd_weight
  )
)

Built-in features: different intercepts, different slopes, and different snow-depth distributions by sex.

Simpson’s Paradox

Pooling males and females together and fitting a single trend line can hide — or even reverse — the true relationship.

ggplot(bighorn_population, aes(x = snow_depth, y = weight)) +
  geom_point(aes(color = sex)) +
  geom_smooth(method = "lm", se = FALSE, color = "red", linewidth = 1)

Ignoring sex hides the true within-sex trends.

Pooled: snow depth looks positively associated with weight — even though both sexes lose weight with more snow.

Why? Sex is a confounder

  • Within each sex: more snow → lower weight (true negative relationship)
  • Between sexes: males see more snow and are heavier (confounded)
  • Pooled data compares heavy, high-snow males to light, low-snow females → spurious positive trend
ggplot(bighorn_population) +
  geom_point(aes(x = snow_depth, y = weight, color = sex)) +
  geom_smooth(aes(x = snow_depth, y = weight), method = "lm", se = FALSE, color = "red") +
  geom_smooth(aes(x = snow_depth, y = weight, color = sex), method = "lm", se = FALSE)

Within-sex trends are negative, as expected.

Exercise 1

Using bighorn_population:

  1. Fit weight ~ sex * snow_depth with glmmTMB(). Call it mod_mass_true.
  2. From the summary, identify which coefficient represents the male slope of snow depth. (Hint: it’s not printed directly — you have to add two rows together.)
  3. In one sentence, explain why a naive weight ~ snow_depth model (ignoring sex) would mislead a manager trying to understand the effect of a harsh winter. Can you think of an example of such a situation in your own field or research?

(~5 minutes — discuss with a neighbor)

Sampling from the population

In practice we can’t measure everyone. Two sampling scenarios:

  • Scenario 1: balanced trap, 50 males / 50 females
  • Scenario 2: biased trap, 80 males / 20 females
n_individuals <- 100
bighorn_mass_1 <- bighorn_population |>
  group_by(sex) |>
  slice_sample(n = n_individuals / 2, replace = FALSE) |>
  ungroup()

mod_mass_1 <- glmmTMB(weight ~ sex * snow_depth,
                       data = bighorn_mass_1, family = gaussian())
summary(mod_mass_1)$coefficients$cond
                     Estimate Std. Error    z value     Pr(>|z|)
(Intercept)        179.711531  3.4913367  51.473560 0.000000e+00
sexmale             39.011436 11.0092824   3.543504 3.948474e-04
snow_depth          -2.495525  0.1717861 -14.526933 8.180169e-48
sexmale:snow_depth   1.686797  0.2685643   6.280794 3.368472e-10
attr(,"ddf")
[1] "asymptotic"

Reading the coefficients:

  • Intercept: female weight at snow_depth = 0
  • sexmale: male − female intercept difference
  • snow_depth: female slope
  • sexmale:snow_depth: male − female slope difference

Useful for hypothesis tests, not directly interpretable for “what’s the effect for males?” — that’s what marginaleffects is for.

Q1: Do sexes differ in spring mass?

predictions(mod_mass_1, by = "sex")

    sex Estimate Std. Error    z Pr(>|z|)   S 2.5 % 97.5 %
 female      137       1.88 73.0   <0.001 Inf   133    141
 male        178       1.88 95.1   <0.001 Inf   175    182

Type: response
predictions(mod_mass_1, by = "sex") |>
  hypotheses(hypothesis = "b1 - b2 = 0")

 Hypothesis Estimate Std. Error     z Pr(>|z|)     S 2.5 % 97.5 %
    b1-b2=0    -41.5       2.65 -15.7   <0.001 181.2 -46.7  -36.3

This is a marginal prediction: the model-implied mean for each sex, averaged over the observed snow-depth distribution in the sample.

Q2: Does snow depth affect mass equally by sex?

slopes(mod_mass_1, variables = "snow_depth", by = "sex", vcov = TRUE) |>
  ggplot(aes(x = estimate, xmax = conf.high, xmin = conf.low, y = sex, color = sex)) +
  geom_pointrange(size = 1) +
  geom_vline(xintercept = female_loss, color = "red") +
  geom_vline(xintercept = male_loss, color = "blue") +
  theme(legend.position = "none")

Slopes of snow depth by sex, with true simulated values as vertical lines.
slopes(mod_mass_1, variables = "snow_depth", by = "sex", vcov = TRUE) |>
  hypotheses(hypothesis = "b1 - b2 = 0")

 Hypothesis Estimate Std. Error     z Pr(>|z|)    S 2.5 % 97.5 %
    b1-b2=0    -1.69      0.269 -6.28   <0.001 31.5 -2.21  -1.16

Q3: What’s the population-average slope?

avg_slopes() computes the slope for every observation, then averages:

avg_slopes(mod_mass_1, variables = "snow_depth", vcov = TRUE)

 Estimate Std. Error     z Pr(>|z|)     S 2.5 % 97.5 %
    -1.65      0.134 -12.3   <0.001 113.1 -1.92  -1.39

Term: snow_depth
Type: response
Comparison: dY/dX
mean(c(male_loss, female_loss))  # true simulated average
[1] -1.65

Marginal vs conditional, visualized

avg_pred_mass_1 <- avg_predictions(mod_mass_1, variables = "snow_depth")
pred_sex_mass_1 <- predictions(mod_mass_1, newdata = bighorn_mass_1)

ggplot() +
  geom_ribbon(data = avg_pred_mass_1,
              aes(x = snow_depth, ymin = conf.low, ymax = conf.high),
              color = "lightgreen", fill = "lightgreen", alpha = 0.4) +
  geom_line(data = avg_pred_mass_1, aes(x = snow_depth, y = estimate),
            color = "lightgreen") +
  geom_ribbon(data = pred_sex_mass_1,
              aes(x = snow_depth, ymin = conf.low, ymax = conf.high, color = sex, fill = sex),
              alpha = 0.8) +
  geom_line(data = pred_sex_mass_1, aes(x = snow_depth, y = estimate, color = sex)) +
  geom_point(data = bighorn_mass_1, aes(y = weight, x = snow_depth, color = sex)) +
  guides(alpha = "none", fill = "none") +
  labs(title = "Green = marginal; red/blue = conditional (by sex)") +
  coord_cartesian(ylim = c(50, 220))

Green = marginal; red/blue = conditional (by sex).

Scenario 2: biased sampling (80% male)

bighorn_mass_2 <- bind_rows(
  bighorn_population |> filter(sex == "male") |> slice_sample(n = 80),
  bighorn_population |> filter(sex == "female") |> slice_sample(n = 20)
)
mod_mass_2 <- glmmTMB(weight ~ sex * snow_depth, data = bighorn_mass_2, family = gaussian())
summary(mod_mass_2)
 Family: gaussian  ( identity )
Formula:          weight ~ sex * snow_depth
Data: bighorn_mass_2

      AIC       BIC    logLik -2*log(L)  df.resid 
    845.8     858.9    -417.9     835.8        95 


Dispersion estimate for gaussian family (sigma^2):  250 

Conditional model:
                   Estimate Std. Error z value Pr(>|z|)    
(Intercept)        182.7351     7.5600  24.171  < 2e-16 ***
sexmale             35.8867    12.7414   2.817  0.00485 ** 
snow_depth          -2.5263     0.3534  -7.149 8.76e-13 ***
sexmale:snow_depth   1.7521     0.4027   4.351 1.36e-05 ***
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
avg_slopes(mod_mass_2, variables = "snow_depth", vcov = TRUE)

 Estimate Std. Error     z Pr(>|z|)    S 2.5 % 97.5 %
    -1.12       0.17 -6.62   <0.001 34.7 -1.46 -0.792

Term: snow_depth
Type: response
Comparison: dY/dX

Conditional (by-sex) slopes stay close to truth. But the marginal slope is biased toward the overrepresented, weaker-effect sex (males).

Correcting for known sampling bias

If we know the true population sex ratio (50:50), we can average over a balanced newdata grid instead of our biased sample.

newdata_balanced <- bighorn_mass_2 |>
  group_by(sex) |>
  reframe(snow_depth = seq(min(snow_depth), max(snow_depth), length.out = 50)) |>
  ungroup()

Exercise 2

Using the model that was fit to the entire population and the one with a biased sample, use the avg_predictions() function along with a balanced and an unbalanced prediction grid to create a plot comparing the marginal predictions between the true population, biased population predictions, and corrected population predictions. There should be three trend lines.

(~15 minutes)

Part 1 takeaways

  1. Interactions → conditional effects that depend on other variables
  2. Simpson’s Paradox is a real risk when pooling across confounded groups
  3. Marginal ≠ conditional; both are valid but answer different questions
  4. Raw model coefficients rarely equal your research quantity of interest — use marginaleffects
  5. Marginal effects inherit your sample’s covariate distribution unless you override it with newdata

Part 2 — Numerical × Numerical Interactions

Two numerical predictors

With two continuous predictors, an interaction means: the slope of one predictor changes as the other predictor changes — not a separate slope per group, but a sliding one.

A raw coefficient is only “the slope when the other predictor equals zero” — meaningful only if zero is a meaningful value. This motivates centering predictors on their mean, so the intercept and main-effect coefficients describe average conditions.

Case study: petal length

Simulated flowers; petal length as a function of temperature and rain, with an interaction. Predictors are mean-centered: the intercept becomes the predicted response at average conditions rather than at (biologically meaningless) zero temperature/rain.

n_flowers <- 200
temperature <- rnorm(n_flowers, mean = 10, sd = 15)
rain <- rgamma(n_flowers, shape = 2)
avg_petal_length <- 25.6; sd_petal_length <- 4.5
temperature_effect <- 0.4; rain_effect <- 0.9; interaction_effect <- 0.2

flower_measurements_1 <- tibble(
  temp_centered = temperature - mean(temperature),
  rain_centered = rain - mean(rain),
  petal_length = rnorm(n_flowers,
    mean = avg_petal_length +
      temperature_effect * temp_centered +
      rain_effect * rain_centered +
      interaction_effect * temp_centered * rain_centered,
    sd = sd_petal_length)
)

mod_flower_1 <- glmmTMB(petal_length ~ temp_centered * rain_centered,
                         data = flower_measurements_1)
summary(mod_flower_1)$coefficients$cond
                              Estimate Std. Error   z value     Pr(>|z|)
(Intercept)                 24.7315526 0.34191019 72.333477 0.000000e+00
temp_centered                0.4441477 0.02209133 20.105064 6.663435e-90
rain_centered                1.0517379 0.26351075  3.991253 6.572525e-05
temp_centered:rain_centered  0.2181736 0.01616323 13.498145 1.603630e-41
attr(,"ddf")
[1] "asymptotic"

Every term has a small p-value — but what do these numbers mean for our research question? Time to unpack them.

Slope of one predictor at a single value of the other

# Effect of temperature when rain_centered = 0 (i.e., average rain)
slopes(mod_flower_1, newdata = datagrid(rain_centered = 0),
       variables = "temp_centered")

 rain_centered Estimate Std. Error    z Pr(>|z|)     S 2.5 % 97.5 %
             0    0.444     0.0221 20.1   <0.001 296.2 0.401  0.487

Term: temp_centered
Type: response
Comparison: dY/dX

This matches the temp_centered coefficient directly — because rain is centered, “at rain = 0” means “at average rain.”

Slope across the range of the moderator

slopes_by_temp <- slopes(mod_flower_1,
  newdata = datagrid(temp_centered = seq(min(flower_measurements_1$temp_centered),
                                          max(flower_measurements_1$temp_centered), length.out = 30)),
  variables = "rain_centered", by = "temp_centered")

trend_slope_temp <- coef(lm(estimate ~ temp_centered, data = slopes_by_temp))[["temp_centered"]]

slopes_by_temp |>
  ggplot() +
  geom_hline(yintercept = 0, color = "blue", linewidth = 1.5, linetype = 2) +
  geom_ribbon(aes(x = temp_centered + mean(temperature), ymin = conf.low, ymax = conf.high),
    alpha = 0.4, fill = "lightgrey") +
  geom_line(aes(x = temp_centered + mean(temperature), y = estimate),
    lineend = "round", linewidth = 2, color = "white") +
  annotate("text", x = -Inf, y = Inf, hjust = -0.2, vjust = 1.5,
    label = paste0("slope = ", round(trend_slope_temp, 3)),
    color = "white", size = 6, fontface = "bold") +
  scale_x_continuous(breaks = seq(-25, 50, by = 5)) +
  labs(y = "Effect of rain on petal length", x = "Temperature values")

Effect of rain on petal length, across observed temperatures.

The interaction coefficient is exactly the slope of this line — how much the effect of rain changes per unit of temperature.

Visualizing predictions at chosen moderator values

Common choice: the five-number summary of the moderator.

predictions(mod_flower_1,
  newdata = datagrid(
    temp_centered = seq(min(flower_measurements_1$temp_centered),
                         max(flower_measurements_1$temp_centered), length.out = 30),
    rain_centered = fivenum
  )) |>
  as_tibble() |>
  mutate(rain = round(rain_centered + mean(rain), 2)) |>
  ggplot() +
  geom_line(aes(y = estimate, x = temp_centered + mean(temperature), color = as.factor(rain)),
            lineend = "round", linewidth = 1.5) +
  geom_ribbon(aes(ymin = conf.low, ymax = conf.high,
                  x = temp_centered + mean(temperature), fill = as.factor(rain)), alpha = 0.2) +
  labs(y = "Predicted petal length", x = "Temperature", color = "Rain", fill = "Rain") +
  coord_cartesian(ylim = c(0, 50))

Predicted petal length across temperature, at five reference rain levels.

Response surfaces — use with caution

predictions(mod_flower_1,
  newdata = datagrid(
    temp_centered = seq(min(flower_measurements_1$temp_centered),
                         max(flower_measurements_1$temp_centered), length.out = 30),
    rain_centered = seq(min(flower_measurements_1$rain_centered),
                         max(flower_measurements_1$rain_centered), length.out = 30)
  )) |>
  as_tibble() |>
  mutate(rain = rain_centered + mean(rain), temp = temp_centered + mean(temperature)) |>
  ggplot(aes(x = rain, y = temp, z = estimate)) +
  stat_contour_filled() +
  labs(x = "Rain", y = "Temperature", fill = "Predicted petal length")

Contour bins are arbitrary and hide uncertainty — good for exploration, risky for publication-quality inference. Prefer line-and-ribbon plots when you can.

Exercise 3

Using mod_flower_1:

  1. Compute the slope of temperature across the observed range of rain (mirror the rain-across-temperature plot above).
  2. At what value of rain does the temperature effect become non-significant (CI crosses zero), if at all?
  3. Rewrite the research question this model could answer as a single sentence, identifying which variable is the exposure and which is the moderator.

(~10 minutes)

Interactions beyond the mean: distributional regression

Interactions don’t have to act only on the average response. glmmTMB’s dispformula lets predictors affect the standard deviation too — useful when variance itself is ecologically meaningful (e.g., reduced trait variance under stress).

log_sd_intercept <- log(sd_petal_length)
log_sd_temp_slope <- 0.03

flower_measurements_2 <- tibble(
  temp_centered = temperature - mean(temperature),
  rain_centered = rain - mean(rain),
  petal_length = rnorm(n_flowers,
    mean = avg_petal_length + temperature_effect * temp_centered,
    sd = exp(log_sd_intercept + log_sd_temp_slope * temp_centered))
)
ggplot(flower_measurements_2, aes(x = temp_centered + mean(temperature), y = petal_length)) +
  geom_smooth(method = 'lm',     
              lineend = "round",
              linewidth = 2,
              color = "white",
              se = F) + 
  geom_point(color = "white") +
  labs(x = "Temperature",
       y = "Observed petal length")

The spread around the trend increases with temperature.
mod_flower_2 <- glmmTMB(petal_length ~ temp_centered,
                         dispformula = ~ temp_centered,
                         data = flower_measurements_2)

Checking the dispersion model

predictions(mod_flower_2, newdata = flower_measurements_2, type = "disp") |>
  as_tibble() -> pred_sd_temp

true_sd_temp <- flower_measurements_2 |>
  mutate(true_sd = exp(log_sd_intercept + log_sd_temp_slope * temp_centered))

ggplot() +
  geom_line(data = true_sd_temp,
            aes(x = temp_centered + mean(temperature), y = true_sd, color = "Simulated"),
            linewidth = 2, linetype = "dashed") +
  geom_ribbon(data = pred_sd_temp,
              aes(ymin = conf.low, ymax = conf.high, x = temp_centered + mean(temperature)),
              fill = "red", alpha = 0.2) +
  geom_line(data = pred_sd_temp,
            aes(x = temp_centered + mean(temperature), y = estimate, color = "Estimated"),
            linewidth = 2, lineend = "round") +
  scale_color_manual(values = c("Simulated" = "green", "Estimated" = "red"), name = "Standard deviation") +
  labs(x = "Temperature", y = "Standard deviation")

Recovered vs. simulated standard deviation of petal length.

type = "disp" retrieves predictions/slopes for the dispersion sub-model, exactly as type = "conditional" does for the mean.

Simulating and fitting interactions on the standard deviation

log_sd_rain_slope <- 0.02
log_sd_interaction_effect <- 0.005

flower_measurements_3 <- tibble(
  temp_centered = temperature - mean(temperature),
  rain_centered = rain - mean(rain),
  petal_length = rnorm(
    n = n_flowers,
    mean = avg_petal_length +
      temperature_effect * temp_centered +
      rain_effect * rain_centered,
    sd = exp(log_sd_intercept +
      log_sd_temp_slope * temp_centered +
      log_sd_rain_slope * rain_centered +
      log_sd_interaction_effect * temp_centered * rain_centered))
)

mod_flower_measurements_3 <- glmmTMB(petal_length ~ temp_centered + rain_centered,
                                     dispformula = ~ temp_centered + rain_centered + temp_centered * rain_centered,
                                     data = flower_measurements_3)

We can interpret the interactive effects of rain and temperature on the standard deviation of petal length in the same way that we would for the average (see the Numerical x Numerical Interactions tutorial for worked examples). Another way to communicate the results: since the two parameters of the Normal distribution act together to generate the observations, we can predict new distributions of observations for different predictor values.

Predicting new observations

predicted_observations <- predictions(mod_flower_measurements_3,
            newdata = datagrid(
              temp_centered = fivenum,
              rain_centered = fivenum
            ),
            type = "conditional") |>
  as_tibble() |>
  select(temp_centered, rain_centered, estimate) |>
  rename(mu = estimate) |>
  left_join(predictions(mod_flower_measurements_3,
            newdata = datagrid(
              temp_centered = fivenum,
              rain_centered = fivenum
            ),
            type = "disp") |>
  as_tibble() |>
  select(temp_centered, rain_centered, estimate) |>
  rename(sd = estimate)) |>
  rowwise() |>
  mutate(temp_centered = round(temp_centered),
         rain_centered = round(rain_centered),
         pred_obs = dist_normal(mu = mu, sd = sd))

# Build facet labels dynamically from the actual rounded temp_centered values
# (hardcoding them would silently show NA facet labels if the simulated data changes)
temp_labels <- setNames(
  paste0("temp = ", sort(unique(predicted_observations$temp_centered))),
  sort(unique(predicted_observations$temp_centered))
)

ggplot(predicted_observations, aes(xdist = pred_obs, color = as.factor(rain_centered))) +
  stat_slab(fill = NA, normalize = "xy") +
  scale_color_brewer(palette = "Accent") +
  facet_wrap(~ as.factor(temp_centered), scales = "free", labeller = as_labeller(temp_labels)) +
  labs(x = "Predicted petal length", color = "Rain (centered)") +
  theme(axis.line.y = element_blank(),
        axis.text.y = element_blank(),
        axis.title.y = element_blank(),
        strip.text = element_text(color = "black", face = "bold", size = 12))

Predicted distributions of petal length at five reference rain and temperature levels.

Part 2 takeaways

  1. With two numerical predictors, the interaction coefficient is the rate of change of one slope per unit of the other — not a separate slope for each group
  2. Centering predictors makes main-effect coefficients interpretable as “effect at average conditions”
  3. Plot slopes across the moderator’s range, not just the coefficient table
  4. Five-number-summary prediction plots are usually more honest than response surfaces (they retain uncertainty)
  5. Interactions can act on variance, not just the mean, via distributional regression (dispformula)