Tidybayes spread_draws unexpected output

Interfaces Other
1

Hi,

I’m having an issue with using the R tidybayes::spread_draws function to extract draws of a random slope coefficient in brms model.

It is giving me twice the number of draws I expect, and the second half seem not correspond to the correct coefficient.

I’ve attached my code output here: https://bignardi.co.uk/MiscellaneousRCode/99_spreaddraws_confusion.html

I may have made a mistake somewhere, but I’m trying to understand what I’ve done wrong so I can avoid making mistakes with this in future.

My current hypothesis is that maybe it is extracting some transformed version of the parameter that brms creates in stan?

2

In your sampling call you are specifying 2 chains of 10,000 iterations. The estimated posterior will thus contain 20,000 samples of random effect parameters.

You can reduce the number of samples by specifying an argument to ndraws

4

I don’t think so because i have 5,000 warm up samples. My output says that: total post-warmup draws = 10000.

spread_draws should not include warmup iterations by deafult, and i set brms to not save warm up samples.

5

It seems you have parameters in the form of x[i, j], so your use of the function,

tidybayes::spread_draws(r_pps[pps,condition_contrast])

I think creates a variable condition_contrast in the resulting data frame/tibble and puts all variations of what shows up in x[,__] into that variable. I see two types that show up (intercept, and condition_contrast), so assuming the function behaves the same with brms objects as with cmdstanr/Stan objects that would explain what you think are double and some of them not being what you expect.

2 Likes
6

I think @ssp3nc3r has identified the likely explanation.

@Giacomo to get a better sense of the structure of the output of spread_draws(), you can pipe the output of spread_draws() into the median_qi() function. This will show you what the unique values of the indices of the array were as well as medians and quantile intervals for them.

You could also try spread_rvars() instead of spread_draws(), which outputs posterior::rvar()s instead of many rows of draws, and therefore makes it easier to see the structure of the output data frame.

2 Likes
7

Thank you both! Of course! I missed that when checking the output, but now i remember in the past filter()'ing the draws. I will check out spread_rvars() as I will probably make this same mistake again in future.

1 Like