Latent mean centering (latent covariate models) in brms

Interfaces brms
,
21

The switcheroo idea can run, but you need to add the second non linear equation within the nlf() function. This is because in brms, when you add nl=T, it only treate the main (first) equation as non linear. Then any other non-linear equations need to be specify by nlf

But when I tested this, the model doesnt converged for a couple of tries

m4_latent_formula <- bf(
  # urge model
  urge ~ alphaU + phiU*(urge1 - alphaU) + betaU*(dep - alphaD),
  alphaU ~ 1 + (1 | person),
  phiU ~ 1 + (1 | person),
  betaU ~ 1 + (1 | person),
  
  # dep model
  nlf(dep ~ alphaD + phiD*(dep1 - alphaD) + betaD*(urge - alphaU)) ,
  alphaD ~ 1 + (1 | person),
  phiD~ 1 + (1 | person),
  betaD ~ 1 + (1 | person),
  nl = TRUE
) +
  gaussian()
3 Likes
22

Thanks for this @Mauricio_Garnier-Villarre! I’ve modified this code a bit and gave it a shot.

🎉⭐️ It works! ⭐️🎉

(I think 🤣)

Huge thanks to everyone who helped with this!

I don’t think there is a reason to write a whole other model for dep, because I just want to get the mean (for the sole purpose of using it in the first non-linear formula). So I edited your suggestion a bit to just estimate the mean with a nlf() spec:

m4_latent_formula <- bf(
  urge ~ alpha + phi*(urge1 - alpha) + beta*(dep - depB),
  alpha ~ 1 + (1 | person),
  phi ~ 1 + (1 | person),
  beta ~ 1 + (1 | person),
  nlf(depB ~ depBI),
  depBI ~ 1 + (1 | person),
  nl = TRUE
) +
  gaussian()

p <- get_prior(m4_latent_formula, data = m4_data) %>%
  mutate(
    prior = case_when(
      class == "b" & coef == "Intercept" ~ "normal(0, 1)",
      class == "sd" & coef == "Intercept" ~ "student_t(7, 0, 1)",
      TRUE ~ prior
    )
  )

m4_latent <- brm(
  m4_latent_formula,
  data = m4_data,
  # Prior is required
  prior = p,
  control = list(adapt_delta = 0.99),
  file = "m4_latent_v3"
)
variable Result (brms) Authors
b_alpha_Intercept (α) -0.11 [-0.32, 0.10] -0.01 [-0.18, 0.16]
b_phi_Intercept (ϕ) 0.21 [0.18, 0.25] 0.21 [0.17, 0.24]
b_beta_Intercept (β) 0.79 [0.61, 0.97] 0.80 [0.61, 0.95]
b_depBI_Intercept (DepB) -0.10 [-0.24, 0.04] 0.01 [-0.02, 0.04]
var_alpha_Intercept (σα2) 0.56 [0.41, 0.78] 0.60 [0.44, 0.83]
var_phi_Intercept (σϕ)2 0.02 [0.01, 0.03] 0.02 [0.01, 0.03]
var_beta_Intercept (σβ2) 0.76 [0.58, 1.04] 0.79 [0.61, 0.95]
var_depBI_Intercept (σDepB2) 0.01 [0.00, 0.05] 0.01 [0.00, 0.01]
sigma (σ2) 1.14 [1.10, 1.19] 1.14 [1.09, 1.19]

So the real “trick” here is to wrap the predicted parameter in nlf(), and this should be a general solution.

The results are not identical to the MPlus ones reported in the target paper, but I am not too worried about that, because there will be some differences in the priors and I could run longer chains etc. I’m going to test this a bit more before marking your post as a solution.

Amazing work. Thank you!

10 Likes
23

OK.

I’ve done some checks and this works afaict. I was cheeky and marked my own post as the solution (it was closest to my original question and works with the original code provided), but want to recognize that I wouldn’t have been able to do so without @Mauricio_Garnier-Villarre’s example code!

I’ve started a GitHub repository working on a tutorial-style manuscript for a psychology audience (not statisticians) on this topic with Joran Jongerling (who might be @Joran here [? 😀]). It is currently private but as I said before, I would be glad to invite people who would like to contribute to the manuscript as coauthors. @Mauricio_Garnier-Villarre @simonbrauer @e.m.mccormick please let me know if you’d like to join—I currently list you only in the acknowledgements section of draft 0.0001😀.

In any case this is marked as solved, huge thanks to everyone!

1 Like