Although priors are set exactly as posterior, sampling from priors gives invalid predictions

Modeling
,
4

I see, so essentially, traditional priors over parameters alone are not enough to “mimick” or rather reproduce a model. In other words, two models could have the same parameters, but different covariances which would result in different predictions. That’s an interesting thought! Though a bit despairing in regards to my issue :)

In regards to the model above, following your suggestions, I have tried to:

  1. estimate properly the weird parameter by increasing chains, iterations, delta, max_treedepth, as well as setting a starting value (following this).
inits <- list()
for(i in 1:5) inits[[i]] <- list(Intercept_ndt = -5)

m <- brms::brm(formula, data = iris, iter = 5000, chains = 5, 
               adapt_delta = 0.9, max_treedepth = 15, inits = inits,
               refresh = 0, seed = 33)

Unfortunately that doesn’t help and the parameter just keep getting immense:

      Parameter   Median     MAD Rhat
b_ndt_Intercept -2.6e+13 3.8e+13  1.8
  1. I went for getting the “pairs plot of the model posterior”. I think you are referring to:
pairs(m)

image
image2435×1085 777 KB

I’m not sure what to look for in this plot though… are there any particular or typical patterns that I should be wary of?

  1. I tried to set the priors to this bug number… but it won’t allow me to (the SD is too big)…
Compiling Stan program...
|
Semantic error in 'C:/Users/user/AppData/Local/Temp/RtmpWSLGjV/model-45c70fb5cbd.stan', line 98, column 52 to column 62:
   -------------------------------------------------
    96:    target += normal_lpdf(Intercept | 1.3, 0.18);
    97:    target += normal_lpdf(Intercept_sigma | -2.6, 0.29);
    98:    target += normal_lpdf(Intercept_ndt | -260000000, 3800000000);
                                                             ^
    99:    target += normal_lpdf(sd_1[1] | 0.3, 0.19)
   100:      - 1 * normal_lccdf(0 | 0.3, 0.19);
   -------------------------------------------------

Integer literal cannot be larger than 2_147_483_647.

Well… I’m not sure what options I have left 😅 Thanks though for your input so far!