Prior predictive check - asymmetry in estimation of uncertainty?

Interfaces brms
1

I am using brms (2.1.2, windows, rstudio) and am running a prior predictive check for data from an experiment, and was playing around with the parameters, specifically the slope and don’t understand the behaviour:

modelTEST1<- brm(outcome ~ 1 + predictor , 
               data  = dat, warmup = 1000,
               prior = c( prior(normal(0,1000), class="Intercept"),
                          prior(cauchy(0,100), class = "sigma"),
                          prior(normal(0,10000000), class = "b") ), 
               iter  = 3000, chains = 2, 
               sample_prior = "only",
               seed  = 221, control = list(adapt_delta = 0.97),
               cores = 2)

plot(conditional_effects(modelTEST1)). This yields a graph (see attached)

Rplot
Rplot817×585 6.25 KB

What I don’t understand is why there is this asymmetry in the graph? Why is there more uncertainty of the estimate of predictor=0 than predictor =1? It appears that the larger the slope, the larger the asymmetry - but why?

Sorry if I am missing something really basic, but am totally stumped!

Thanks

2

I’m not an expert with brms but looking at your graph, it seems like the model is interpreting your ConditionASINT variable as a continuous predictor, when in fact it seems more like a factor with two levels (0, 1). Is that correct? If so, you need to specify ConditionASINT as a factor prior to running the model, otherwise you will get weird output.

Hope this helps!. Also, posting your code might help others to diagnose any potential problems.

3

Thanks so much for answering - actually I recoded it as a factor and the same thing happens - the interval on the left is greater than the interval on the right

Rplot02
Rplot02817×1118 4.73 KB

I would have thought that both intervals should be identical, since I am calling brms with "sample_prior = “only” "

I’m not sure what you mean by “posting your code”. The code was in the post:

modelTEST1<- brm(outcome ~ 1 + predictor , 
               data  = dat, warmup = 1000,
               prior = c( prior(normal(0,1000), class="Intercept"),
                          prior(cauchy(0,100), class = "sigma"),
                          prior(normal(0,10000000), class = "b") ), 
               iter  = 3000, chains = 2, 
               sample_prior = "only",
               seed  = 221, control = list(adapt_delta = 0.97),
               cores = 2)

(sorry if I am making some mistake with the formatting or etiquette of asking questions - this is my first time posting on here)

Thankyou

4

I just realized the problem - because there is an unequal number of data in both conditions, it expects different amounts of variation in both conditions

2 Likes
5

Yeah, sounds like you got your solution! Also, what’s the range of your data? The “normal(0, 100000000)” prior may be unnecessarily vague.