3  Comparisons

In this vignette, we introduce “comparisons”, defined as:

Compare the predictions made by a model for different regressor values (e.g., college graduates vs. others): contrasts, differences, risk ratios, odds, etc.

The comparisons() function is extremely flexible, and it allows users to estimate a vast array of quantities of interest. To describe those quantities, we will break the problem up in 4 steps:

  1. Quantity
  2. Grid
  3. Average
  4. Hypothesis

These steps can be combined and mixed and matched to define and compute many different estimands.

3.1 Simple example: Titanic

Consider a logistic regression model estimated using the Titanic mortality data:

library(marginaleffects)

dat <- "https://vincentarelbundock.github.io/Rdatasets/csv/Stat2Data/Titanic.csv"
dat <- read.csv(dat)
dat$PClass[dat$PClass == "*"] <- NA
mod <- glm(Survived ~ PClass * SexCode * Age, data = dat, family = binomial)

3.1.1 Step 1: Quantity

The question that interests us is:

How does the probability of survival (outcome) change if a passenger travels in 1st class vs. 3rd class?

Since we are comparing two predicted outcomes, we will use comparisons(). To indicate that our focal variable is PClass and that we are interested in the comparison between 1st and 3rd class, we will use the variables argument:

comparisons(mod,
  variables = list(PClass = c("1st", "3rd"))) # Step 1: Quantity

3.1.2 Step 2: Grid

In GLM models, most quantities of interest are conditional, in the sense that they will typically depend on the values of all the predictors in the model. Therefore, we need to decide where in the predictor space we want to evaluate the quantity of interest described above.

By default, comparisons() will compute estimates for every row of the original dataset that was used to fit a model. There are 1313 observations in the titanic dataset. Therefore, if we just execute the code in the previous section, we will obtain 1313 estimates of the difference between the probability of survival in 3rd and 1st class:

comparisons(mod,
  variables = list(PClass = c("1st", "3rd"))) # Step 1: Quantity
#> 
#>    Term  Contrast Estimate Std. Error     z Pr(>|z|)    S  2.5 % 97.5 %
#>  PClass 3rd - 1st   -0.496     0.0610 -8.13  < 0.001 51.0 -0.616 -0.376
#>  PClass 3rd - 1st   -0.472     0.1247 -3.79  < 0.001 12.7 -0.716 -0.228
#>  PClass 3rd - 1st   -0.353     0.0641 -5.51  < 0.001 24.7 -0.478 -0.227
#>  PClass 3rd - 1st   -0.493     0.0583 -8.45  < 0.001 55.0 -0.607 -0.379
#>  PClass 3rd - 1st   -0.445     0.1452 -3.07  0.00216  8.9 -0.730 -0.161
#> --- 746 rows omitted. See ?avg_comparisons and ?print.marginaleffects --- 
#>  PClass 3rd - 1st   -0.377     0.0703 -5.36  < 0.001 23.5 -0.515 -0.239
#>  PClass 3rd - 1st   -0.384     0.0726 -5.30  < 0.001 23.0 -0.527 -0.242
#>  PClass 3rd - 1st   -0.412     0.0821 -5.02  < 0.001 20.9 -0.573 -0.251
#>  PClass 3rd - 1st   -0.399     0.0773 -5.16  < 0.001 22.0 -0.550 -0.247
#>  PClass 3rd - 1st   -0.361     0.0661 -5.47  < 0.001 24.4 -0.490 -0.232
#> Columns: rowid, term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high, predicted_lo, predicted_hi, predicted, Survived, PClass, SexCode, Age 
#> Type:  response

Notice that the contrast between 3rd and 1st is different from row to row. This reflects the fact that, in our model, moving from 1st to 3rd would have a different effect on the predicted probability of survival for different individuals.

We can be more specific in our query. Instead of using the empirical distribution as our “grid”, we can specify exactly where we want to evaluate the comparison in the predictor space, by using the newdata argument and the datagrid() function. For example, say I am interested in:

The effect of moving from 1st to 3rd class on the probability of survival for a 50 year old man and a 50 year old woman.

I can type:

cmp <- comparisons(mod,
  variables = list(PClass = c("1st", "3rd")),  # Step 1: Quantity
  newdata = datagrid(Age = 50, SexCode = 0:1)) # Step 2: Grid
cmp
#> 
#>    Term  Contrast Age SexCode Estimate Std. Error     z Pr(>|z|)    S  2.5 %  97.5 % PClass
#>  PClass 3rd - 1st  50       0   -0.184     0.0535 -3.45   <0.001 10.8 -0.289 -0.0796    3rd
#>  PClass 3rd - 1st  50       1   -0.511     0.1242 -4.12   <0.001 14.7 -0.755 -0.2679    3rd
#> 
#> Columns: rowid, term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high, Age, SexCode, predicted_lo, predicted_hi, predicted, Survived, PClass 
#> Type:  response

We now know that moving from 1st to 3rd changes by -0.184 the probability of survival for 50 year old men (SexCode=0), and by -0.511 the probability of survival for 50 year old women (SexCode=1).

3.1.3 Step 3: Averaging

Again, by default comparisons() estimates quantities for all the actually observed units in our dataset. Sometimes, it is convenient to marginalize those conditional estimates, in order to obtain an “average contrast”:

avg_comparisons(mod,                          # Step 3: Average
  variables = list(PClass = c("1st", "3rd"))) # Step 1: Quantity
#> 
#>    Term  Contrast Estimate Std. Error    z Pr(>|z|)    S  2.5 % 97.5 %
#>  PClass 3rd - 1st   -0.396     0.0425 -9.3   <0.001 66.0 -0.479 -0.312
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

Alternatively, we could also take the average, but just of the two estimates that we computed above for the 50 year old man and 50 year old woman.

avg_comparisons(mod,                           # Step 3: Average
  variables = list(PClass = c("1st", "3rd")),  # Step 1: Quantity
  newdata = datagrid(Age = 50, SexCode = 0:1)) # Step 2: Grid
#> 
#>    Term  Contrast Estimate Std. Error     z Pr(>|z|)    S 2.5 % 97.5 %
#>  PClass 3rd - 1st   -0.348     0.0676 -5.15   <0.001 21.8 -0.48 -0.215
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

Notice that this is exactly the same as the average in the estimates from the previous section, which we had stored as cmp:

cmp$estimate
#> [1] -0.1844289 -0.5113098

mean(cmp$estimate)
#> [1] -0.3478694

3.1.4 Hypothesis

Finally, imagine we are interested in this question:

Does moving from 1st to 3rd class have a bigger effect on the probability of survival for 50 year old men, or for 50 year old women?

To answer this, we use the hypothesis argument:

comparisons(mod,
  variables = list(PClass = c("1st", "3rd")),  # Step 1: Quantity
  newdata = datagrid(Age = 50, SexCode = 0:1), # Step 2: Grid
  hypothesis = "b1 = b2")                      # Step 4: Hypothesis
#> 
#>   Term Estimate Std. Error    z Pr(>|z|)   S  2.5 % 97.5 %
#>  b1=b2    0.327      0.135 2.42   0.0156 6.0 0.0618  0.592
#> 
#> Columns: term, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

This result maps directly onto the estimates we had above. It is the difference in the contrast for 50-men and 50-women:

diff(cmp$estimate)
#> [1] -0.3268809

This result can be interpreted as a “difference-in-differences”: Moving from 1st to 3rd has a much larger negative effect on the probability of survival for a 50 year old woman than for a 50 year old man. This difference is statistically significant.

We can do a similar comparison, but instead of fixing a conditional grid, we can average over subgroups of the empirical distribution, using the by argument:

avg_comparisons(mod,
  variables = list(PClass = c("1st", "3rd")),  # Step 1: Quantity
  by = "SexCode",                              # Step 3: Average
  hypothesis = "b1 = b2")                      # Step 4: Hypothesis
#> 
#>   Term Estimate Std. Error    z Pr(>|z|)   S    2.5 % 97.5 %
#>  b1=b2    0.162     0.0845 1.91   0.0558 4.2 -0.00402  0.327
#> 
#> Columns: term, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

3.1.5 Manual computation

Now we show how to use the base R predict() function to compute some of the same quantities as above. This exercise may be clarifying for some users.

grid_50_1_3 <- data.frame(Age = 50, SexCode = 1, PClass = "3rd")
grid_50_1_1 <- data.frame(Age = 50, SexCode = 1, PClass = "1st")
grid_50_0_3 <- data.frame(Age = 50, SexCode = 0, PClass = "3rd")
grid_50_0_1 <- data.frame(Age = 50, SexCode = 0, PClass = "1st")


yhat_50_1_3 <- predict(mod, newdata = grid_50_1_3, type = "response")
yhat_50_1_1 <- predict(mod, newdata = grid_50_1_1, type = "response")
yhat_50_0_3 <- predict(mod, newdata = grid_50_0_3, type = "response")
yhat_50_0_1 <- predict(mod, newdata = grid_50_0_1, type = "response")

## prediction on a grid
predictions(mod, newdata = datagrid(Age = 50, SexCode = 1, PClass = "3rd"))
#> 
#>  Age SexCode PClass Estimate Pr(>|z|)   S 2.5 % 97.5 %
#>   50       1    3rd    0.446    0.661 0.6 0.235  0.679
#> 
#> Columns: rowid, estimate, p.value, s.value, conf.low, conf.high, Survived, Age, SexCode, PClass 
#> Type:  invlink(link)
yhat_50_1_3
#>         1 
#> 0.4463379

## contrast on a grid
comparisons(mod,
  variables = list(PClass = c("1st", "3rd")),
  newdata = datagrid(Age = 50, SexCode = 0:1))
#> 
#>    Term  Contrast Age SexCode Estimate Std. Error     z Pr(>|z|)    S  2.5 %  97.5 % PClass
#>  PClass 3rd - 1st  50       0   -0.184     0.0535 -3.45   <0.001 10.8 -0.289 -0.0796    3rd
#>  PClass 3rd - 1st  50       1   -0.511     0.1242 -4.12   <0.001 14.7 -0.755 -0.2679    3rd
#> 
#> Columns: rowid, term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high, Age, SexCode, predicted_lo, predicted_hi, predicted, Survived, PClass 
#> Type:  response

yhat_50_0_3 - yhat_50_0_1
#>          1 
#> -0.1844289
yhat_50_1_3 - yhat_50_1_1
#>          1 
#> -0.5113098

## difference-in-differences 
comparisons(mod,
  variables = list(PClass = c("1st", "3rd")),
  newdata = datagrid(Age = 50, SexCode = 0:1),
  hypothesis = "b1 = b2")
#> 
#>   Term Estimate Std. Error    z Pr(>|z|)   S  2.5 % 97.5 %
#>  b1=b2    0.327      0.135 2.42   0.0156 6.0 0.0618  0.592
#> 
#> Columns: term, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

(yhat_50_0_3 - yhat_50_0_1) - (yhat_50_1_3 - yhat_50_1_1)
#>         1 
#> 0.3268809

## average of the empirical distribution of contrasts
avg_comparisons(mod, variables = list(PClass = c("1st", "3rd")), by = "SexCode")
#> 
#>    Term              Contrast SexCode Estimate Std. Error     z Pr(>|z|)    S  2.5 % 97.5 %
#>  PClass mean(3rd) - mean(1st)       0   -0.334     0.0570 -5.86   <0.001 27.7 -0.446 -0.222
#>  PClass mean(3rd) - mean(1st)       1   -0.496     0.0623 -7.95   <0.001 49.0 -0.618 -0.374
#> 
#> Columns: term, contrast, SexCode, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high, predicted_lo, predicted_hi, predicted 
#> Type:  response

grid_empirical_1_3 <- dat |> subset(SexCode == 1) |> transform(PClass = "3rd")
grid_empirical_1_1 <- dat |> subset(SexCode == 1) |> transform(PClass = "1st")
grid_empirical_0_3 <- dat |> subset(SexCode == 0) |> transform(PClass = "3rd")
grid_empirical_0_1 <- dat |> subset(SexCode == 0) |> transform(PClass = "1st")
yhat_empirical_0_1 <- predict(mod, newdata = grid_empirical_0_1, type = "response")
yhat_empirical_0_3 <- predict(mod, newdata = grid_empirical_0_3, type = "response")
yhat_empirical_1_1 <- predict(mod, newdata = grid_empirical_1_1, type = "response")
yhat_empirical_1_3 <- predict(mod, newdata = grid_empirical_1_3, type = "response")
mean(yhat_empirical_0_3, na.rm = TRUE) - mean(yhat_empirical_0_1, na.rm = TRUE)
#> [1] -0.3341426
mean(yhat_empirical_1_3, na.rm = TRUE) - mean(yhat_empirical_1_1, na.rm = TRUE)
#> [1] -0.4956673

3.2 Predictor types

3.2.1 Logical and factor predictors

Consider a simple model with a logical and a factor variable:

library(marginaleffects)

tmp <- mtcars
tmp$am <- as.logical(tmp$am)
mod <- lm(mpg ~ am + factor(cyl), tmp)

The comparisons() function automatically computes contrasts for each level of the categorical variables, relative to the baseline category (FALSE for logicals, and the reference level for factors), while holding all other values at their observed values. The avg_comparisons() does the same, but then marginalizes by taking the average of unit-level estimates:

cmp <- avg_comparisons(mod)
cmp
#> 
#>  Term     Contrast Estimate Std. Error     z Pr(>|z|)    S    2.5 % 97.5 %
#>   am  TRUE - FALSE     2.56       1.30  1.97   0.0485  4.4   0.0167   5.10
#>   cyl 6 - 4           -6.16       1.54 -4.01   <0.001 14.0  -9.1661  -3.15
#>   cyl 8 - 4          -10.07       1.45 -6.93   <0.001 37.8 -12.9136  -7.22
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

The summary printed above says that moving from the reference category 4 to the level 6 on the cyl factor variable is associated with a change of -6.156 in the adjusted prediction. Similarly, the contrast from FALSE to TRUE on the am variable is equal to 2.560.

We can obtain different contrasts by using the comparisons() function. For example:

avg_comparisons(mod, variables = list(cyl = "sequential"))
#> 
#>  Term Contrast Estimate Std. Error     z Pr(>|z|)    S 2.5 % 97.5 %
#>   cyl    6 - 4    -6.16       1.54 -4.01  < 0.001 14.0 -9.17  -3.15
#>   cyl    8 - 6    -3.91       1.47 -2.66  0.00781  7.0 -6.79  -1.03
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

avg_comparisons(mod, variables = list(cyl = "pairwise"))
#> 
#>  Term Contrast Estimate Std. Error     z Pr(>|z|)    S  2.5 % 97.5 %
#>   cyl    6 - 4    -6.16       1.54 -4.01  < 0.001 14.0  -9.17  -3.15
#>   cyl    8 - 4   -10.07       1.45 -6.93  < 0.001 37.8 -12.91  -7.22
#>   cyl    8 - 6    -3.91       1.47 -2.66  0.00781  7.0  -6.79  -1.03
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

avg_comparisons(mod, variables = list(cyl = "reference"))
#> 
#>  Term Contrast Estimate Std. Error     z Pr(>|z|)    S  2.5 % 97.5 %
#>   cyl    6 - 4    -6.16       1.54 -4.01   <0.001 14.0  -9.17  -3.15
#>   cyl    8 - 4   -10.07       1.45 -6.93   <0.001 37.8 -12.91  -7.22
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

For comparison, this code produces the same results using the emmeans package:

library(emmeans)
emm <- emmeans(mod, specs = "cyl")
contrast(emm, method = "revpairwise")
#>  contrast    estimate   SE df t.ratio p.value
#>  cyl6 - cyl4    -6.16 1.54 28  -4.009  0.0012
#>  cyl8 - cyl4   -10.07 1.45 28  -6.933  <.0001
#>  cyl8 - cyl6    -3.91 1.47 28  -2.660  0.0331
#> 
#> Results are averaged over the levels of: am 
#> P value adjustment: tukey method for comparing a family of 3 estimates

emm <- emmeans(mod, specs = "am")
contrast(emm, method = "revpairwise")
#>  contrast     estimate  SE df t.ratio p.value
#>  TRUE - FALSE     2.56 1.3 28   1.973  0.0585
#> 
#> Results are averaged over the levels of: cyl

Note that these commands also work on for other types of models, such as GLMs, on different scales:

mod_logit <- glm(am ~ factor(gear), data = mtcars, family = binomial)

avg_comparisons(mod_logit)
#> 
#>  Term Contrast Estimate Std. Error       z Pr(>|z|)    S 2.5 % 97.5 %
#>  gear    4 - 3    0.667   1.36e-01 4.9e+00   <0.001 20.0   0.4  0.933
#>  gear    5 - 3    1.000   9.95e-06 1.0e+05   <0.001  Inf   1.0  1.000
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

avg_comparisons(mod_logit, type = "link")
#> 
#>  Term Contrast Estimate Std. Error       z Pr(>|z|)   S  2.5 % 97.5 %
#>  gear    4 - 3     21.3       4578 0.00464    0.996 0.0  -8951   8994
#>  gear    5 - 3     41.1       9156 0.00449    0.996 0.0 -17904  17986
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  link

3.2.2 Character predictors

All functions of the marginaleffects package attempt to treat character predictors as factor predictors. However, using factors instead of characters when modeling is strongly encouraged, because they are much safer and faster. This is because factors hold useful information about the full list of levels, which makes them easier to track and handle internally by marginaleffects. Users are strongly encouraged to convert their character variables to factor before fitting their models and using slopes() functions.

3.2.3 Numeric predictors

We can also compute contrasts for differences in numeric variables. For example, we can see what happens to the adjusted predictions when we increment the hp variable by 1 unit (default) or by 5 units about the original value:

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

avg_comparisons(mod)
#> 
#>  Term Contrast Estimate Std. Error     z Pr(>|z|)    S   2.5 %  97.5 %
#>    hp       +1  -0.0682     0.0101 -6.74   <0.001 35.9 -0.0881 -0.0484
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

avg_comparisons(mod, variables = list(hp = 5))
#> 
#>  Term Contrast Estimate Std. Error     z Pr(>|z|)    S 2.5 % 97.5 %
#>    hp       +5   -0.341     0.0506 -6.74   <0.001 35.9 -0.44 -0.242
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

Compare adjusted predictions for a change in the regressor between two arbitrary values:

avg_comparisons(mod, variables = list(hp = c(90, 110)))
#> 
#>  Term Contrast Estimate Std. Error     z Pr(>|z|)    S 2.5 % 97.5 %
#>    hp 110 - 90    -1.36      0.202 -6.74   <0.001 35.9 -1.76 -0.968
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

Compare adjusted predictions when the regressor changes across the interquartile range, across one or two standard deviations about its mean, or from across its full range:

avg_comparisons(mod, variables = list(hp = "iqr"))
#> 
#>  Term Contrast Estimate Std. Error     z Pr(>|z|)    S 2.5 % 97.5 %
#>    hp  Q3 - Q1     -5.7      0.845 -6.74   <0.001 35.9 -7.35  -4.04
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

avg_comparisons(mod, variables = list(hp = "sd"))
#> 
#>  Term                Contrast Estimate Std. Error     z Pr(>|z|)    S 2.5 % 97.5 %
#>    hp (x + sd/2) - (x - sd/2)    -4.68      0.694 -6.74   <0.001 35.9 -6.04  -3.32
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

avg_comparisons(mod, variables = list(hp = "2sd"))
#> 
#>  Term            Contrast Estimate Std. Error     z Pr(>|z|)    S 2.5 % 97.5 %
#>    hp (x + sd) - (x - sd)    -9.36       1.39 -6.74   <0.001 35.9 -12.1  -6.64
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

avg_comparisons(mod, variables = list(hp = "minmax"))
#> 
#>  Term  Contrast Estimate Std. Error     z Pr(>|z|)    S 2.5 % 97.5 %
#>    hp Max - Min    -19.3       2.86 -6.74   <0.001 35.9 -24.9  -13.7
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

3.3 Interactions and Cross-Contrasts

In some contexts we are interested in whether the “effect” of a variable changes, as a function of another variable. A very simple strategy to tackle this question is to estimate a model with a multiplicative interaction like this one:

mod <- lm(mpg ~ am * factor(cyl), data = mtcars)

Calling avg_comparisons() with the by argument shows that the estimated comparisons differ based on cyl:

avg_comparisons(mod, variables = "am", by = "cyl")
#> 
#>  Term          Contrast cyl Estimate Std. Error     z Pr(>|z|)   S 2.5 % 97.5 %
#>    am mean(1) - mean(0)   4     5.18       2.05 2.521   0.0117 6.4  1.15   9.20
#>    am mean(1) - mean(0)   6     1.44       2.32 0.623   0.5336 0.9 -3.10   5.98
#>    am mean(1) - mean(0)   8     0.35       2.32 0.151   0.8799 0.2 -4.19   4.89
#> 
#> Columns: term, contrast, cyl, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high, predicted_lo, predicted_hi, predicted 
#> Type:  response

However, using the hypothesis argument for pairwise contrasts between the above comparisons reveals that the heterogeneity is not statistically significant:

avg_comparisons(mod, variables = "am", by = "cyl", hypothesis = "pairwise")
#> 
#>   Term Estimate Std. Error     z Pr(>|z|)   S 2.5 % 97.5 %
#>  4 - 6     3.73       3.09 1.206    0.228 2.1 -2.33   9.80
#>  4 - 8     4.83       3.09 1.559    0.119 3.1 -1.24  10.89
#>  6 - 8     1.09       3.28 0.333    0.739 0.4 -5.33   7.51
#> 
#> Columns: term, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

In other contexts, we are interested in a “cross-contrast” or “cross-comparisons”; we would like to know what happens when two (or more) predictors change at the same time. To assess this, we can specify the regressors of interest in the variables argument, and set the cross=TRUE:

avg_comparisons(mod, variables = c("cyl", "am"), cross = TRUE)
#> 
#>  Estimate Std. Error      z Pr(>|z|)   S  2.5 % 97.5 % C: am C: cyl
#>     -2.33       2.48 -0.942  0.34596 1.5  -7.19   2.52 1 - 0  6 - 4
#>     -7.50       2.77 -2.709  0.00674 7.2 -12.93  -2.07 1 - 0  8 - 4
#> 
#> Columns: term, contrast_am, contrast_cyl, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

3.4 Quantities of interest

This section compares 4 quantities:

  1. Unit-Level Contrasts
  2. Average Contrast
  3. Contrast at the Mean
  4. Contrast Between Marginal Means

The ideas discussed in this section focus on contrasts, but they carry over directly to analogous types of marginal effects.

3.4.1 Unit-level contrasts

In models with interactions or non-linear components (e.g., link function), the value of a contrast or marginal effect can depend on the value of all the predictors in the model. As a result, contrasts and marginal effects are fundamentally unit-level quantities. The effect of a 1 unit increase in \(X\) can be different for Mary or John. Every row of a dataset has a different contrast and marginal effect.

The mtcars dataset has 32 rows, so the comparisons() function produces 32 contrast estimates:

library(marginaleffects)
mod <- glm(vs ~ factor(gear) + mpg, family = binomial, data = mtcars)
cmp <- comparisons(mod, variables = "mpg")
nrow(cmp)
#> [1] 32

3.4.2 Average contrasts

By default, the slopes() and comparisons() functions compute marginal effects and contrasts for every row of the original dataset. These unit-level estimates can be of great interest, as discussed in another vignette. Nevertheless, one may want to focus on one-number summaries: the avg_*() functions or the by argument compute the “Average Marginal Effect” or “Average Contrast,” by taking the mean of all the unit-level estimates.

avg_comparisons(mod, variables = "mpg")
#> 
#>  Term Contrast Estimate Std. Error    z Pr(>|z|)    S  2.5 % 97.5 %
#>   mpg       +1   0.0608     0.0128 4.74   <0.001 18.8 0.0356  0.086
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

comparisons(mod, variables = "mpg", by = TRUE)
#> 
#>  Term Contrast Estimate Std. Error    z Pr(>|z|)    S  2.5 % 97.5 %
#>   mpg       +1   0.0608     0.0128 4.74   <0.001 18.8 0.0356  0.086
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

which are equivalent to:

mean(cmp$estimate)
#> [1] 0.06080995

We could also show the full distribution of contrasts across our dataset with a histogram:

library(ggplot2)

cmp <- comparisons(mod, variables = "gear")

ggplot(cmp, aes(estimate)) +
    geom_histogram(bins = 30) +
    facet_wrap(~contrast, scale = "free_x") +
    labs(x = "Distribution of unit-level contrasts")

This graph displays the effect of a change of 1 unit in the mpg variable, for each individual in the observed data.

3.4.3 Contrasts at the mean

An alternative which used to be very common but has now fallen into a bit of disfavor is to compute “Contrasts at the mean.” The idea is to create a “synthetic” or “hypothetical” individual (row of the dataset) whose characteristics are completely average. Then, we compute and report the contrast for this specific hypothetical individual.

This can be achieved by setting newdata="mean" or to newdata=datagrid(), both of which fix variables to their means or modes:

comparisons(mod, variables = "mpg", newdata = "mean")
#> 
#>  Term Contrast Estimate Std. Error    z Pr(>|z|)   S  2.5 % 97.5 % gear  mpg
#>   mpg       +1    0.166     0.0629 2.65  0.00811 6.9 0.0432   0.29    3 20.1
#> 
#> Columns: rowid, term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high, predicted_lo, predicted_hi, predicted, vs, gear, mpg 
#> Type:  response

Contrasts at the mean can differ substantially from average contrasts.

The advantage of this approach is that it is very cheap and fast computationally. The disadvantage is that the interpretation is somewhat ambiguous. Often times, there simply does not exist an individual who is perfectly average across all dimensions of the dataset. It is also not clear why the analyst should be particularly interested in the contrast for this one, synthetic, perfectly average individual.

3.4.4 Contrasts between marginal means

Yet another type of contrast is the “Contrast between marginal means.” This type of contrast is closely related to the “Contrast at the mean”, with a few wrinkles. It is the default approach used by the emmeans package for R.

Roughly speaking, the procedure is as follows:

  1. Create a prediction grid with one cell for each combination of categorical predictors in the model, and all numeric variables held at their means.
  2. Make adjusted predictions in each cell of the prediction grid.
  3. Take the average of those predictions (marginal means) for each combination of btype (focal variable) and resp (group by variable).
  4. Compute pairwise differences (contrasts) in marginal means across different levels of the focal variable btype.

The contrast obtained through this approach has two critical characteristics:

  1. It is the contrast for a synthetic individual with perfectly average qualities on every (numeric) predictor.
  2. It is a weighted average of unit-level contrasts, where weights assume a perfectly balanced dataset across every categorical predictor.

With respect to (a), the analyst should ask themselves: Is my quantity of interest the contrast for a perfectly average hypothetical individual? With respect to (b), the analyst should ask themselves: Is my quantity of interest the contrast in a model estimated using (potentially) unbalanced data, but interpreted as if the data were perfectly balanced?

For example, imagine that one of the control variables in your model is a variable measuring educational attainment in 4 categories: No high school, High school, Some college, Completed college. The contrast between marginal is a weighted average of contrasts estimated in the 4 cells, and each of those contrasts will be weighted equally in the overall estimate. If the population of interest is highly unbalanced in the educational categories, then the estimate computed in this way will not be most useful.

If the contrasts between marginal means is really the quantity of interest, it is easy to use the comparisons() to estimate contrasts between marginal means. The newdata determines the values of the predictors at which we want to compute contrasts. We can set newdata="marginalmeans" to emulate the emmeans behavior. For example, here we compute contrasts in a model with an interaction:

dat <- read.csv("https://vincentarelbundock.github.io/Rdatasets/csv/palmerpenguins/penguins.csv")
mod <- lm(bill_length_mm ~ species * sex + island + body_mass_g, data = dat)

avg_comparisons(
    mod,
    newdata = "marginalmeans",
    variables = c("species", "island"))
#> 
#>     Term           Contrast Estimate Std. Error      z Pr(>|z|)     S  2.5 % 97.5 %
#>  island  Dream - Biscoe       -0.434      0.451 -0.962    0.336   1.6 -1.318  0.450
#>  island  Torgersen - Biscoe    0.060      0.467  0.128    0.898   0.2 -0.856  0.976
#>  species Chinstrap - Adelie   10.563      0.418 25.272   <0.001 465.7  9.744 11.382
#>  species Gentoo - Adelie       5.792      0.798  7.257   <0.001  41.2  4.228  7.356
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

Which is equivalent to this in emmeans:

emm <- emmeans(
    mod,
    specs = c("species", "island"))
contrast(emm, method = "trt.vs.ctrl1")
#>  contrast                            estimate    SE  df t.ratio p.value
#>  Chinstrap Biscoe - Adelie Biscoe      nonEst    NA  NA      NA      NA
#>  Gentoo Biscoe - Adelie Biscoe          5.792 0.798 331   7.257  <.0001
#>  Adelie Dream - Adelie Biscoe          -0.434 0.451 331  -0.962  0.8476
#>  Chinstrap Dream - Adelie Biscoe       nonEst    NA  NA      NA      NA
#>  Gentoo Dream - Adelie Biscoe           5.358 1.035 331   5.177  <.0001
#>  Adelie Torgersen - Adelie Biscoe       0.060 0.467 331   0.128  0.9998
#>  Chinstrap Torgersen - Adelie Biscoe   nonEst    NA  NA      NA      NA
#>  Gentoo Torgersen - Adelie Biscoe       5.852 1.070 331   5.468  <.0001
#> 
#> Results are averaged over the levels of: sex 
#> P value adjustment: dunnettx method for 8 tests

The emmeans section of the Alternative Software vignette shows further examples.

The excellent vignette of the emmeans package discuss the same issues in a slightly different (and more positive) way:

The point is that the marginal means of cell.means give equal weight to each cell. In many situations (especially with experimental data), that is a much fairer way to compute marginal means, in that they are not biased by imbalances in the data. We are, in a sense, estimating what the marginal means would be, had the experiment been balanced. Estimated marginal means (EMMs) serve that need.

All this said, there are certainly situations where equal weighting is not appropriate. Suppose, for example, we have data on sales of a product given different packaging and features. The data could be unbalanced because customers are more attracted to some combinations than others. If our goal is to understand scientifically what packaging and features are inherently more profitable, then equally weighted EMMs may be appropriate; but if our goal is to predict or maximize profit, the ordinary marginal means provide better estimates of what we can expect in the marketplace.

3.5 Conditional contrasts

Consider a model with an interaction term. What happens to the dependent variable when the hp variable increases by 10 units?

library(marginaleffects)

mod <- lm(mpg ~ hp * wt, data = mtcars)

plot_comparisons(
    mod,
    variables = list(hp = 10),
    condition = "wt")

3.6 Transformations

So far we have focused on simple differences between adjusted predictions. Now, we show how to use ratios, back transformations, and arbitrary functions to estimate a slew of quantities of interest. Powerful transformations and custom contrasts are made possible by using three arguments which act at different stages of the computation process:

  • comparison
  • transform

Consider the case of a model with a single predictor \(x\). To compute average contrasts, we proceed as follows:

  1. Compute adjusted predictions for each row of the dataset for the observed values \(x\): \(\hat{y}_x\)
  2. Compute adjusted predictions for each row of the dataset for the observed values \(x + 1\): \(\hat{y}_{x+1}\)
  3. comparison: Compute unit-level contrasts by taking the difference between (or some other function of) adjusted predictions: \(\hat{y}_{x+1} - \hat{y}_x\)
  4. Compute the average contrast by taking the mean of unit-level contrasts: \(1/N \sum_{i=1}^N \hat{y}_{x+1} - \hat{y}_x\)
  5. transform: Transform the average contrast or return them as-is.

The comparison argument of the comparisons() function determines how adjusted predictions are combined to create a contrast. By default, we take a simple difference between predictions with hi value of \(x\), and predictions with a lo value of \(x\): function(hi, lo) hi-lo.

The transform argument of the comparisons() function applies a custom transformation to the unit-level contrasts.

The transform argument applies a custom transformation to the final quantity, as would be returned if we evaluated the same call without transform.

3.7 Differences

The default contrast calculate by the comparisons() function is a (untransformed) difference between two adjusted predictions. For instance, to estimate the effect of a change of 1 unit, we do:

library(marginaleffects)

mod <- glm(vs ~ mpg, data = mtcars, family = binomial)

## construct data

mtcars_minus <- mtcars_plus <- mtcars
mtcars_minus$mpg <- mtcars_minus$mpg - 0.5
mtcars_plus$mpg <- mtcars_plus$mpg + 0.5

## adjusted predictions
yhat_minus <- predict(mod, newdata = mtcars_minus, type = "response")
yhat_plus <- predict(mod, newdata = mtcars_plus, type = "response")

## unit-level contrasts
con <- yhat_plus - yhat_minus

## average contrasts
mean(con)
#> [1] 0.05540227

We can use the avg_comparisons() function , or the by argument to obtain the same results:

avg_comparisons(mod)
#> 
#>  Term Contrast Estimate Std. Error    z Pr(>|z|)    S 2.5 % 97.5 %
#>   mpg       +1   0.0554    0.00835 6.64   <0.001 34.9 0.039 0.0718
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

comparisons(mod, by = TRUE)
#> 
#>  Term Contrast Estimate Std. Error    z Pr(>|z|)    S 2.5 % 97.5 %
#>   mpg       +1   0.0554    0.00835 6.64   <0.001 34.9 0.039 0.0718
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

3.8 Difference-in-Differences(-in-Differences)

Going back to our Titanic example:

dat <- "https://vincentarelbundock.github.io/Rdatasets/csv/Stat2Data/Titanic.csv"
dat <- read.csv(dat)
titanic <- glm(Survived ~ PClass * SexCode * Age, data = dat, family = binomial)

In this case, a contrast is a difference between predicted probabilities. We can compute that contrast for different types of individuals:

comparisons(
  titanic,
  variables = "SexCode",
  newdata = datagrid(PClass = c("1st", "3rd")))
#> 
#>     Term Contrast PClass Estimate Std. Error    z Pr(>|z|)    S 2.5 % 97.5 % SexCode  Age
#>  SexCode    1 - 0    1st    0.483     0.0631 7.65   <0.001 45.5 0.359  0.606   0.381 30.4
#>  SexCode    1 - 0    3rd    0.335     0.0634 5.29   <0.001 22.9 0.211  0.459   0.381 30.4
#> 
#> Columns: rowid, term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high, PClass, predicted_lo, predicted_hi, predicted, Survived, SexCode, Age 
#> Type:  response

One we can notice above, is that the gap in predicted probabilities of survival between men and women is larger in 1st class than in 3rd class. Being a woman matters more for your chances of survival if you travel in first class. Is the difference between those contrasts (diff-in-diff) statistically significant?

To answer this question, we can compute a difference-in-difference using the hypothesis argument (see the Hypothesis vignette for details). For example, using b1 and b2 to refer to the contrasts in the first and second rows of the output above, we can test if the difference between the two quantities is different from 0:

comparisons(
  titanic,
  hypothesis = "b1 - b2 = 0",
  variables = "SexCode",
  newdata = datagrid(PClass = c("1st", "3rd")))
#> 
#>     Term Estimate Std. Error    z Pr(>|z|)   S   2.5 % 97.5 %
#>  b1-b2=0    0.148     0.0894 1.65   0.0987 3.3 -0.0276  0.323
#> 
#> Columns: term, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

Now, let’s say we consider more types of individuals:

comparisons(
  titanic,
  variables = "SexCode",
  newdata = datagrid(PClass = c("1st", "3rd"), Age = range))
#> 
#>     Term Contrast PClass   Age Estimate Std. Error      z Pr(>|z|)     S   2.5 % 97.5 % SexCode
#>  SexCode    1 - 0    1st  0.17   0.1081      0.122  0.883   0.3774   1.4 -0.1319  0.348   0.381
#>  SexCode    1 - 0    1st 71.00   0.8795      0.057 15.437   <0.001 176.2  0.7679  0.991   0.381
#>  SexCode    1 - 0    3rd  0.17   0.0805      0.157  0.513   0.6081   0.7 -0.2272  0.388   0.381
#>  SexCode    1 - 0    3rd 71.00   0.4265      0.203  2.101   0.0356   4.8  0.0287  0.824   0.381
#> 
#> Columns: rowid, term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high, PClass, Age, predicted_lo, predicted_hi, predicted, Survived, SexCode 
#> Type:  response

With these results, we could compute a triple difference:

comparisons(
  titanic,
  hypothesis = "(b1 - b3) - (b2 - b4) = 0",
  variables = "SexCode",
  newdata = datagrid(PClass = c("1st", "3rd"), Age = range))
#> 
#>               Term Estimate Std. Error     z Pr(>|z|)   S 2.5 % 97.5 %
#>  (b1-b3)-(b2-b4)=0   -0.425      0.359 -1.19    0.236 2.1 -1.13  0.278
#> 
#> Columns: term, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

3.9 Ratios

Instead of taking simple differences between adjusted predictions, it can sometimes be useful to compute ratios or other functions of predictions. For example, the adjrr function the Stata software package can compute “adjusted risk ratios”, which are ratios of adjusted predictions. To do this in R, we use the comparison argument:

avg_comparisons(mod, comparison = "ratio")
#> 
#>  Term Contrast Estimate Std. Error    z Pr(>|z|)    S 2.5 % 97.5 %
#>   mpg       +1     1.29      0.133 9.69   <0.001 71.4  1.03   1.55
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

This result is the average adjusted risk ratio for an increment of 1, that is, the adjusted predictions when the mpg are incremented by 0.5, divided by the adjusted predictions when mpg is decremented by 0.5.

The comparison accepts different values for common types of contrasts: ‘difference’, ‘ratio’, ‘lnratio’, ‘ratioavg’, ‘lnratioavg’, ‘lnoravg’, ‘differenceavg’. These strings are shortcuts for functions that accept two vectors of adjusted predictions and returns a single vector of contrasts. For example, these two commands yield identical results:

avg_comparisons(mod, comparison = "ratio")
#> 
#>  Term Contrast Estimate Std. Error    z Pr(>|z|)    S 2.5 % 97.5 %
#>   mpg       +1     1.29      0.133 9.69   <0.001 71.4  1.03   1.55
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

avg_comparisons(mod, comparison = function(hi, lo) hi / lo)
#> 
#>  Term Contrast Estimate Std. Error    z Pr(>|z|)    S 2.5 % 97.5 %
#>   mpg       +1     1.29      0.133 9.69   <0.001 71.4  1.03   1.55
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

This mechanism is powerful, because it lets users create fully customized contrasts. Here is a non-sensical example:

avg_comparisons(mod, comparison = function(hi, lo) sqrt(hi) / log(lo + 10))
#> 
#>  Term Contrast Estimate Std. Error    z Pr(>|z|)    S 2.5 % 97.5 %
#>   mpg       +1    0.264     0.0261 10.1   <0.001 77.4 0.213  0.315
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

The same arguments work in the plotting function plot_comparisons() as well, which allows us to plot various custom contrasts. Here is a comparison of Adjusted Risk Ratio and Adjusted Risk Difference in a model of the probability of survival aboard the Titanic:

library(ggplot2)
library(patchwork)
titanic <- "https://vincentarelbundock.github.io/Rdatasets/csv/Stat2Data/Titanic.csv"
titanic <- read.csv(titanic)
mod_titanic <- glm(
    Survived ~ Sex * PClass + Age + I(Age^2),
    family = binomial,
    data = titanic)

avg_comparisons(mod_titanic)
#> 
#>    Term      Contrast Estimate Std. Error      z Pr(>|z|)     S    2.5 %   97.5 %
#>  Age    +1             -0.0065    0.00108  -6.03   <0.001  29.2 -0.00862 -0.00439
#>  PClass 2nd - 1st      -0.2058    0.03954  -5.20   <0.001  22.3 -0.28328 -0.12828
#>  PClass 3rd - 1st      -0.4043    0.03958 -10.21   <0.001  78.9 -0.48187 -0.32670
#>  Sex    male - female  -0.4847    0.03004 -16.14   <0.001 192.2 -0.54355 -0.42580
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

p1 <- plot_comparisons(
    mod_titanic,
    variables = "Age",
    condition = "Age",
    comparison = "ratio") +
    ylab("Adjusted Risk Ratio\nP(Survival | Age + 1) / P(Survival | Age)")

p2 <- plot_comparisons(
    mod_titanic,
    variables = "Age",
    condition = "Age") +
    ylab("Adjusted Risk Difference\nP(Survival | Age + 1) - P(Survival | Age)")

p1 + p2

By default, the standard errors around contrasts are computed using the delta method on the scale determined by the type argument (e.g., “link” or “response”). Some analysts may prefer to proceed differently. For example, in Stata, the adjrr computes adjusted risk ratios (ARR) in two steps:

  1. Compute the natural log of the ratio between the mean of adjusted predictions with \(x+1\) and the mean of adjusted predictions with \(x\).
  2. Exponentiate the estimate and confidence interval bounds.

Step 1 is easy to achieve with the comparison argument described above. Step 2 can be achieved with the transform argument:

avg_comparisons(
    mod,
    comparison = function(hi, lo) log(hi / lo),
    transform = exp)
#> 
#>  Term Contrast Estimate Pr(>|z|)   S 2.5 % 97.5 %
#>   mpg       +1     1.27  0.00937 6.7  1.06   1.53
#> 
#> Columns: term, contrast, estimate, p.value, s.value, conf.low, conf.high 
#> Type:  response

Note that we can use the lnratioavg shortcut instead of defining the function ourselves.

The order of operations in previous command was:

  1. Compute the custom unit-level log ratios
  2. Exponentiate them
  3. Take the average using the avg_comparisons()

There is a very subtle difference between the procedure above and this code:

avg_comparisons(
    mod,
    comparison = function(hi, lo) log(hi / lo),
    transform = exp)
#> 
#>  Term Contrast Estimate Pr(>|z|)   S 2.5 % 97.5 %
#>   mpg       +1     1.27  0.00937 6.7  1.06   1.53
#> 
#> Columns: term, contrast, estimate, p.value, s.value, conf.low, conf.high 
#> Type:  response

Since the exp function is now passed to the transform argument of the comparisons() function, the exponentiation is now done only after unit-level contrasts have been averaged. This is what Stata appears to do under the hood, and the results are slightly different.

comparisons(
    mod,
    comparison = function(hi, lo) log(mean(hi) / mean(lo)),
    transform = exp)
#> 
#>  Term Contrast Estimate Pr(>|z|)    S 2.5 % 97.5 %
#>   mpg       +1     1.14   <0.001 31.9  1.09   1.18
#> 
#> Columns: term, contrast, estimate, p.value, s.value, conf.low, conf.high, predicted_lo, predicted_hi, predicted 
#> Type:  response

Note that equivalent results can be obtained using shortcut strings in the comparison argument: “ratio”, “lnratio”, “lnratioavg”.

comparisons(
    mod,
    comparison = "lnratioavg",
    transform = exp)
#> 
#>  Term Contrast Estimate Pr(>|z|)    S 2.5 % 97.5 %
#>   mpg mean(+1)     1.14   <0.001 31.9  1.09   1.18
#> 
#> Columns: term, contrast, estimate, p.value, s.value, conf.low, conf.high, predicted_lo, predicted_hi, predicted 
#> Type:  response

All the same arguments apply to the plotting functions of the marginaleffects package as well. For example we can plot the Adjusted Risk Ratio in a model with a quadratic term:

library(ggplot2)
dat_titanic <- read.csv("https://vincentarelbundock.github.io/Rdatasets/csv/Stat2Data/Titanic.csv")
mod2 <- glm(Survived  ~ Age, data = dat_titanic, family = binomial)
plot_comparisons(
    mod2,
    variables = list("Age" = 10),
    condition = "Age",
    comparison = "ratio") +
    ylab("Adjusted Risk Ratio\nP(Survived = 1 | Age + 10) / P(Survived = 1 | Age)")

3.10 Forward, Backward, Centered, and Custom Differences

By default, the comparisons() function computes a “centered” difference. For example, if we ask comparisons() to estimate the effect of a 10-unit change in predictor x on outcome y, comparisons() will compare the predicted values with x-5 and x+5.

dat <- mtcars
dat$new_hp <- 49 * (mtcars$hp - min(mtcars$hp)) / (max(mtcars$hp) - min(mtcars$hp)) + 1
mod <- lm(mpg ~ log(new_hp), data = dat)

avg_comparisons(
  mod,
  variables = list(new_hp = 10))
#> 
#>    Term Contrast Estimate Std. Error     z Pr(>|z|)    S 2.5 % 97.5 %
#>  new_hp      +10    -4.06      0.464 -8.74   <0.001 58.6 -4.97  -3.15
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

We can supply arbitrary functions to create custom differences. These functions must accept a vector of values for the predictor of interest, and return a data frame with the same number of rows as the length, and two columns with the values to compare. For example, we can do:

forward_diff <- \(x) data.frame(x, x + 10)
backward_diff <- \(x) data.frame(x - 10, x)
center_diff <- \(x) data.frame(x - 5, x + 5)

avg_comparisons(
  mod,
  variables = list(new_hp = forward_diff))
#> 
#>    Term Contrast Estimate Std. Error     z Pr(>|z|)    S 2.5 % 97.5 %
#>  new_hp   custom     -3.8      0.435 -8.74   <0.001 58.6 -4.65  -2.95
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

avg_comparisons(
  mod,
  variables = list(new_hp = backward_diff))
#> 
#>    Term Contrast Estimate Std. Error     z Pr(>|z|)    S 2.5 % 97.5 %
#>  new_hp   custom    -6.51      0.744 -8.74   <0.001 58.6 -7.97  -5.05
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

avg_comparisons(
  mod,
  variables = list(new_hp = center_diff))
#> 
#>    Term Contrast Estimate Std. Error     z Pr(>|z|)    S 2.5 % 97.5 %
#>  new_hp   custom    -4.06      0.464 -8.74   <0.001 58.6 -4.97  -3.15
#> 
#> Columns: term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response

Notice that the last “centered” difference gives the same results as the default comparisons() call.

3.11 Lognormal hurdle model

With hurdle models, we can fit two separate models simultaneously:

  1. A model that predicts if the outcome is zero or not zero
  2. If the outcome is not zero, a model that predicts what the value of the outcome is

We can calculate predictions and marginal effects for each of these hurdle model processes, but doing so requires some variable transformation since the stages of these models use different link functions.

The hurdle_lognormal() family in brms uses logistic regression (with a logit link) for the hurdle part of the model and lognormal regression (where the outcome is logged before getting used in the model) for the non-hurdled part. Let’s look at an example of predicting GDP per capita (which is distributed exponentially) using life expectancy. We’ll add some artificial zeros so that we can work with a hurdle stage of the model.

library(dplyr)
library(ggplot2)
library(patchwork)
library(brms)
library(marginaleffects)
library(gapminder)

## Build some 0s into the GDP column
set.seed(1234)
gapminder <- gapminder::gapminder |> 
  filter(continent != "Oceania") |> 
  # Make a bunch of GDP values 0
  mutate(prob_zero = ifelse(lifeExp < 50, 0.3, 0.02),
         will_be_zero = rbinom(n(), 1, prob = prob_zero),
         gdpPercap0 = ifelse(will_be_zero, 0, gdpPercap)) |> 
  select(-prob_zero, -will_be_zero)

mod <- brm(
  bf(gdpPercap0 ~ lifeExp,
     hu ~ lifeExp),
  data = gapminder,
  family = hurdle_lognormal(),
  chains = 4, cores = 4, seed = 1234)

We have two different sets of coefficients here for the two different processes. The hurdle part (hu) uses a logit link, and the non-hurdle part (mu) uses an identity link. However, that’s a slight misnomer—a true identity link would show the coefficients on a non-logged dollar value scale. Because we’re using a lognormal family, GDP per capita is pre-logged, so the “original” identity scale is actually logged dollars.

summary(mod)
#>  Family: hurdle_lognormal 
#>   Links: mu = identity; sigma = identity; hu = logit 
#> Formula: gdpPercap0 ~ lifeExp 
#>          hu ~ lifeExp
#>    Data: gapminder (Number of observations: 1680) 
#>   Draws: 4 chains, each with iter = 2000; warmup = 1000; thin = 1;
#>          total post-warmup draws = 4000
#> 
#> Population-Level Effects: 
#>              Estimate Est.Error l-95% CI u-95% CI Rhat Bulk_ESS Tail_ESS
#> Intercept        3.47      0.09     3.29     3.65 1.00     4757     3378
#> hu_Intercept     3.16      0.40     2.37     3.96 1.00     2773     2679
#> lifeExp          0.08      0.00     0.08     0.08 1.00     5112     3202
#> hu_lifeExp      -0.10      0.01    -0.12    -0.08 1.00     2385     2652
#> ...

We can get predictions for the hu part of the model on the link (logit) scale:

predictions(mod, dpar = "hu", type = "link",
            newdata = datagrid(lifeExp = seq(40, 80, 20)))
#> 
#>  lifeExp Estimate 2.5 % 97.5 %
#>       40   -0.817 -1.03 -0.604
#>       60   -2.805 -3.06 -2.555
#>       80   -4.790 -5.34 -4.275
#> 
#> Columns: rowid, estimate, conf.low, conf.high, gdpPercap0, lifeExp 
#> Type:  link

…or on the response (percentage point) scale:

predictions(mod, dpar = "hu", type = "response",
            newdata = datagrid(lifeExp = seq(40, 80, 20)))
#> 
#>  lifeExp Estimate   2.5 % 97.5 %
#>       40  0.30630 0.26231 0.3534
#>       60  0.05703 0.04466 0.0721
#>       80  0.00824 0.00478 0.0137
#> 
#> Columns: rowid, estimate, conf.low, conf.high, gdpPercap0, lifeExp 
#> Type:  response

We can also get slopes for the hu part of the model on the link (logit) or response (percentage point) scales:

slopes(mod, dpar = "hu", type = "link",
                newdata = datagrid(lifeExp = seq(40, 80, 20)))
#> 
#>     Term lifeExp Estimate  2.5 %  97.5 %
#>  lifeExp      40  -0.0993 -0.116 -0.0837
#>  lifeExp      60  -0.0993 -0.116 -0.0837
#>  lifeExp      80  -0.0993 -0.116 -0.0837
#> 
#> Columns: rowid, term, estimate, conf.low, conf.high, lifeExp, predicted_lo, predicted_hi, predicted, tmp_idx, gdpPercap0 
#> Type:  link

slopes(mod, dpar = "hu", type = "response",
                newdata = datagrid(lifeExp = seq(40, 80, 20)))
#> 
#>     Term lifeExp  Estimate    2.5 %    97.5 %
#>  lifeExp      40 -0.021078 -0.02591 -0.016588
#>  lifeExp      60 -0.005321 -0.00615 -0.004561
#>  lifeExp      80 -0.000812 -0.00115 -0.000543
#> 
#> Columns: rowid, term, estimate, conf.low, conf.high, lifeExp, predicted_lo, predicted_hi, predicted, tmp_idx, gdpPercap0 
#> Type:  response

Working with the mu part of the model is trickier. Switching between type = "link" and type = "response" doesn’t change anything, since the outcome is pre-logged:

predictions(mod, dpar = "mu", type = "link",
            newdata = datagrid(lifeExp = seq(40, 80, 20)))
#> 
#>  lifeExp Estimate 2.5 % 97.5 %
#>       40     6.61  6.54   6.69
#>       60     8.18  8.15   8.22
#>       80     9.75  9.69   9.82
#> 
#> Columns: rowid, estimate, conf.low, conf.high, gdpPercap0, lifeExp 
#> Type:  link
predictions(mod, dpar = "mu", type = "response",
            newdata = datagrid(lifeExp = seq(40, 80, 20)))
#> 
#>  lifeExp Estimate 2.5 % 97.5 %
#>       40     6.61  6.54   6.69
#>       60     8.18  8.15   8.22
#>       80     9.75  9.69   9.82
#> 
#> Columns: rowid, estimate, conf.low, conf.high, gdpPercap0, lifeExp 
#> Type:  response

For predictions, we need to exponentiate the results to scale them back up to dollar amounts. We can do this by post-processing the results (e.g. with dplyr::mutate(predicted = exp(predicted))), or we can use the transform argument in predictions() to pass the results to exp() after getting calculated:

predictions(mod, dpar = "mu", 
            newdata = datagrid(lifeExp = seq(40, 80, 20)),
            transform = exp)
#> 
#>  lifeExp Estimate 2.5 % 97.5 %
#>       40      744   694    801
#>       60     3581  3449   3718
#>       80    17215 16110  18410
#> 
#> Columns: rowid, estimate, conf.low, conf.high, gdpPercap0, lifeExp 
#> Type:  response

We can pass transform = exp to plot_predictions() too:

plot_predictions(
  mod,
  dpar = "hu",
  type = "link",
  condition = "lifeExp") +
  labs(y = "hu",
       title = "Hurdle part (hu)",
       subtitle = "Logit-scale predictions") +
plot_predictions(
  mod,
  dpar = "hu",
  type = "response",
  condition = "lifeExp") +
  labs(y = "hu",
       subtitle = "Percentage point-scale predictions") +
plot_predictions(
  mod,
  dpar = "mu",
  condition = "lifeExp") +
  labs(y = "mu",
       title = "Non-hurdle part (mu)",
       subtitle = "Log-scale predictions") +
plot_predictions(
  mod,
  dpar = "mu",
  transform = exp,
  condition = "lifeExp") +
  labs(y = "mu",
       subtitle = "Dollar-scale predictions")

For marginal effects, we need to transform the predictions before calculating the instantaneous slopes. We also can’t use the slopes() function directly—we need to use comparisons() and compute the numerical derivative ourselves (i.e. predict gdpPercap at lifeExp of 40 and 40.001 and calculate the slope between those predictions). We can use the comparison argument to pass the pair of predicted values to exp() before calculating the slopes:

## step size of the numerical derivative
eps <- 0.001

comparisons(
  mod,
  dpar = "mu",
  variables = list(lifeExp = eps),
  newdata = datagrid(lifeExp = seq(40, 80, 20)),
  # rescale the elements of the slope
  # (exp(40.001) - exp(40)) / exp(0.001)
  comparison = function(hi, lo) ((exp(hi) - exp(lo)) / exp(eps)) / eps
)
#> 
#>     Term Contrast lifeExp Estimate  2.5 % 97.5 %
#>  lifeExp   +0.001      40     58.4   55.8     61
#>  lifeExp   +0.001      60    280.9  266.6    296
#>  lifeExp   +0.001      80   1349.4 1222.6   1490
#> 
#> Columns: rowid, term, contrast, estimate, conf.low, conf.high, lifeExp, predicted_lo, predicted_hi, predicted, tmp_idx, gdpPercap0 
#> Type:  response

We can visually confirm that these are the instantaneous slopes at each of these levels of life expectancy:

predictions_data <- predictions(
  mod,
  newdata = datagrid(lifeExp = seq(30, 80, 1)),
  dpar = "mu",
  transform = exp) |>
  select(lifeExp, prediction = estimate)

slopes_data <- comparisons(
  mod,
  dpar = "mu",
  variables = list(lifeExp = eps),
  newdata = datagrid(lifeExp = seq(40, 80, 20)),
  comparison = function(hi, lo) ((exp(hi) - exp(lo)) / exp(eps)) / eps) |>
  select(lifeExp, estimate) |>
  left_join(predictions_data, by = "lifeExp") |>
  # Point-slope formula: (y - y1) = m(x - x1)
  mutate(intercept = estimate * (-lifeExp) + prediction)

ggplot(predictions_data, aes(x = lifeExp, y = prediction)) +
  geom_line(size = 1) + 
  geom_abline(data = slopes_data, aes(slope = estimate, intercept = intercept), 
              size = 0.5, color = "red") +
  geom_point(data = slopes_data) +
  geom_label(data = slopes_data, aes(label = paste0("Slope: ", round(estimate, 1))),
             nudge_x = -1, hjust = 1) +
  theme_minimal()

We now have this in the experiments section

3.12 Visual examples

okabeito <- c('#E69F00', '#56B4E9', '#009E73', '#F0E442', '#0072B2', '#D55E00', '#CC79A7', '#999999', '#000000')
options(ggplot2.discrete.fill = okabeito)
library(ggplot2)
theme_set(theme_minimal())

library(marginaleffects)
library(ggplot2)

set.seed(1024)
n <- 200
d <- data.frame(
  y = rnorm(n),
  cond = as.factor(sample(0:1, n, TRUE)),
  episode = as.factor(sample(0:4, n, TRUE)))

model1 <- lm(y ~ cond * episode, data = d)

p <- predictions(model1, newdata = datagrid(cond = 0:1, episode = 1:3))
ggplot(p, aes(x = cond, y = estimate, shape = episode, color = episode)) +
  geom_point()


## do episodes 1 and 2 differ when `cond=0`
ggplot(p, aes(x = cond, y = estimate, shape = episode, color = episode)) +
  geom_point() +
  geom_segment(aes(x = 1, xend = 1, y = p$estimate[1], yend = p$estimate[2]), color = "black") +
  ggtitle("What is the vertical distance between the linked points?")


comparisons(model1,
  variables = list(episode = 1:2), # comparison of interest
  newdata = datagrid(cond = 0))    # grid
#> 
#>     Term Contrast cond Estimate Std. Error     z Pr(>|z|)   S  2.5 % 97.5 % episode
#>  episode    2 - 1    0    0.241      0.396 0.609    0.542 0.9 -0.535   1.02       0
#> 
#> Columns: rowid, term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high, cond, predicted_lo, predicted_hi, predicted, y, episode 
#> Type:  response

## do cond=0 and cond=1 differ when episode = 1
ggplot(p, aes(x = cond, y = estimate, shape = episode, color = episode)) +
  geom_point() +
  geom_segment(aes(x = 1, xend = 2, y = p$estimate[1], yend = p$estimate[4]), color = okabeito[1]) +
  ggtitle("What is the vertical distance between the linked points?")


comparisons(model1,
  variables = "cond",              # comparison of interest
  newdata = datagrid(episode = 1)) # grid
#> 
#>  Term Contrast episode Estimate Std. Error    z Pr(>|z|)   S  2.5 % 97.5 % cond
#>  cond    1 - 0       1    0.546      0.347 1.57    0.115 3.1 -0.134   1.23    0
#> 
#> Columns: rowid, term, contrast, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high, episode, predicted_lo, predicted_hi, predicted, y, cond 
#> Type:  response

## Is the difference between episode 1 and 2 larger in cond=0 or cond=1? 
## try this without the `hypothesis` argument to see what we are comparing more clearly
ggplot(p, aes(x = cond, y = estimate, shape = episode, color = episode)) +
  geom_point() +
  annotate("rect", xmin = .9, xmax = 1.1, ymin = p$estimate[1], ymax = p$estimate[2], alpha = .2, fill = "green") +
  annotate("rect", xmin = 1.9, xmax = 2.1, ymin = p$estimate[4], ymax = p$estimate[5], alpha = .2, fill = "orange")  +
  ggtitle("Is the green box taller than the orange box?")


comparisons(model1,
  variables = list(episode = 1:2), # comparison of interest
  newdata = datagrid(cond = 0:1),  # grid
  hypothesis = "b1 = b2")          # hypothesis
#> 
#>   Term Estimate Std. Error     z Pr(>|z|)   S  2.5 % 97.5 %
#>  b1=b2    0.413      0.508 0.812    0.417 1.3 -0.583   1.41
#> 
#> Columns: term, estimate, std.error, statistic, p.value, s.value, conf.low, conf.high 
#> Type:  response