SDT mixed model: log transformation of parameters within model specification

Interfaces brms
, , ,
1

Hi Community,

I am running a signal detection model using brm on 43 subjects with appx. 700 trials per subject.
I provide a small example dataset with 4 subjects.
I want to add beta power as a regressor, to see whether beta power correlates with the criterion or sensitivity. As the single trial values for beta power are very small, I log transform the values. is it okey to do that in the model syntax:


#make factors

cue_power$ID = as.factor(cue_power$ID) # subject ID
cue_power$isyes = as.factor(cue_power$isyes) # signal or noise trial (1 or 0)
cue_power$sayyes = as.factor(cue_power$sayyes) # detection response (yes or no, 1 or 0)

# fit the model 

glm_power <- brm(sayyes ~ 1 + isyes*log10(beta) + (1 + isyes*log10(beta)| ID),
                        family = bernoulli(link = "probit"),
                        data = cue_power,
                        cores = 10,
                        thin = 2,
                        iter = 4000,
                        file = "SDT_mixedmodel_power200prestimulus.rds",
                        control = list(adapt_delta = 0.99)
)

I get the following error msg:

Chain 4: Rejecting initial value:
Chain 4: Log probability evaluates to log(0), i.e. negative infinity.
Chain 4: Stan can’t start sampling from this initial value.
Chain 4:
Chain 4: Initialization between (-2, 2) failed after 100 attempts.
Chain 4: Try specifying initial values, reducing ranges of constrained values, or reparameterizing the model.
[1] “Error in sampler$call_sampler(args_list[[i]]) : Initialization failed.”
[2] “Zusätzlich: Warnmeldungen:”
[3] "1: Paket ‘ggplot2’ wurde unter R Version 4.1.2 erstellt "
[4] "2: vorhergehender Import ‘lifecycle::last_warnings’ durch ‘rlang::last_warnings’ während des Ladens von ‘tibble’ ersetzt "
[5] "3: vorhergehender Import ‘lifecycle::last_warnings’ durch ‘rlang::last_warnings’ während des Ladens von ‘pillar’ ersetzt "
[1] “error occurred during calling the sampler; sampling not done”

I assume I get this error because of the negative log values. But if I transform my values before model specification I still have negative values. Not sure how to proceed here. Any ideas?

Another small question: Where can I see which value is the base value if I have factorial predictors? Does brms always use the 0 coded variable as base?

Best,

Carina
sdt_power_sample.csv (332.6 KB)

2

@matti?

1 Like
3

It’s usually better to transform the values before entering them into the model.

I would try

glm_power <- brm(
  sayyes ~ 1 + isyes*log10(beta) + (1 + isyes*log10(beta)| ID),
  family = bernoulli(link = "probit"),
  data = cue_power,
  cores = 10,
  thin = 2,
  iter = 4000,
  inits = "0",  # This might help. On the other hand, maybe not!
  file = "SDT_mixedmodel_power200prestimulus.rds",
  control = list(adapt_delta = 0.99)
)

Note also that 10 cores doesn’t do anything unless you also run 10 chains or are using threading.

brms uses R contrasts, so unless you’ve modified something the alphanumerically first factor level is the reference level. see ?contrasts. To make the criterion a more standard metric you should use contrast coding see e.g. Sometimes I R - Bayesian Estimation of Signal Detection Models.

Hope that helps.