Explicit subtraction of effects in brms

Interfaces brms
1

I wrote a model in Stan to fit the goal difference goal_diff of a football match as a function of the difference in strength between the home team and the away team. This is the data:

data {
  int<lower=1> N;  // number of observations
  int<lower=1> T;  // number of teams
  int<lower=1,upper=T> home[N];  // home team
  int<lower=1,upper=T> away[N];  // away team
  vector[N] goal_diff;  // goal difference
}

And this is the relevant part of the model:

model {
  real goal_diff_hat;
  for (i in 1:N) {
    goal_diff_hat = Intercept + b[home[i]] - b[away[i]];
    target += double_exponential_lpdf(goal_diff[i] | goal_diff_hat, s);
  }
}

It works fine, and has nice results. Evidently, two data ints are used to estimate a single effect vector. I would like to fit this model with brms, as it’s easier to iterate over different models in it. Initially I thought I could do something with the multi-membership mm helper, but the explicit subtraction in b[home[i]] - b[away[i]] is necessary to correctly estimate b. Is it possible to code this model with a brms formula?

2

You can use negative weights in mm() to achieve this.

3

I had overlooked this! Thanks, Paul.

4

Actually, with brms 2.13.0, when I try to reproduce the mm example with negative weights:

dat <- data.frame(
  y = rnorm(100), x1 = rnorm(100), x2 = rnorm(100),
  g1 = sample(1:10, 100, TRUE), g2 = sample(1:10, 100, TRUE)
)
dat$w1 <- rep(1, 100)
dat$w2 <- rep(-1, 100)

fit <- brm(
  y ~ x1 + (1|mm(g1, g2, weights = cbind(w1, w2))),
  data = dat
)

I get:

Error: Cannot scale negative weights.

With traceback:

7: stop(..., call. = FALSE)
6: stop2("Cannot scale negative weights.")
5: data_gr_local(x, data = data, ranef = ranef)
4: data_predictor.brmsterms(bterms, data = data, prior = prior, 
       data2 = data2, ranef = ranef, basis = basis)
3: data_predictor(bterms, data = data, prior = prior, data2 = data2, 
       ranef = ranef, basis = basis)
2: .make_standata(bterms, data = data, prior = prior, data2 = data2, 
       stanvars = stanvars)
1: brm(y ~ x1 + (1 | mm(g1, g2, weights = cbind(w1, w2))), data = dat)
5

Set scale = FALSE in mm()

6

I had realized this just after posting, so I tried to delete the post. I canceled the deletion to keep this thread coherent. Thanks again!