简体   繁体   中英

How to specify informative prior only on one specific predictor in rstanarm

I'm currently trying to fit a Bayesian multilevel model using rstanarm. I expect - and there's evidence in literature - that the coefficient of my main predictor lies between 0.15 and 0.65. Therefore, I'd like to set an informative prior but only for this variable and keeping the weakly informative default for the others. Up to now I have:

   stan_glmer(isei_r ~ 1 + maxisei_cntr + agea + as.factor(gender) + as.factor(emp_status) + (1 + maxisei_cntr | country), 
            data = ess,
            seed = 349, 
            prior = normal(0.40, 0.25, autoscale=F))  

But in this way it applies the informative prior to all my covariates. Is it possible to specify an informative prior only for one predictor?

Thanks

Here is what I use to automate this process:

data = iris
formula = Sepal.Length ~ Species + Sepal.Width + Petal.Length

# get the coefficient terms, including expanded dummy columns for factors
formula_terms = colnames(
    model.matrix(
        formula,
        data=data
    )
)
# remove intercept (it can specified separately)
formula_terms = setdiff(formula_terms, '(Intercept)')

# create named vectors mapping some of the coefficients to priors
priors_location = c(
    'Speciesversicolor'=-1,
    'Speciesvirginica'=-1.5,
    'Petal.Length'=1
)
priors_scale = c(
    'Speciesversicolor'=5,
    'Speciesvirginica'=5,
    'Petal.Length'=8
)

Now, define a helper function, which can be put elsewhere:

priors_or_default_for = function(formula_terms, priors, default) {
    as.array(
        sapply(
            formula_terms,
            function(term) ifelse(term %in% names(priors), priors[term], default)
        )
    )
}

And use it to create the priors vector:

location = priors_or_default_for(formula_terms, priors_location, default=0)
scale = priors_or_default_for(formula_terms, priors_scale, default=10)

fit = rstanarm::stan_glm(
    formula,
    data=data,
    prior=normal(
        location=location,
        scale=scale
    )
)

# (recommended) check that the priors were properly assigned
parameters::parameters(fit)
Parameter Median 89% CI pd % in ROPE Rhat ESS Prior
(Intercept) 2.40 [ 1.97, 2.79] 100% 0% 1.000 2556.00 Normal (5.84 +- 2.07)
Speciesversicolor -0.95 [-1.31, -0.61] 100% 0% 1.002 1274.00 Normal (-1.00 +- 5.00)
Speciesvirginica -1.39 [-1.86, -0.93] 100% 0% 1.003 1285.00 Normal (-1.50 +- 5.00)
Sepal.Width 0.43 [ 0.29, 0.55] 100% 0% 1.001 1803.00 Normal (0.00 +- 10.00)
Petal.Length 0.77 [ 0.67, 0.88] 100% 0% 1.002 1359.00 Normal (1.00 +- 8.00)

Yes, it is but you need to pass a vector of prior locations and / or scales whose size(s) are equal to the number of coefficients (excluding the intercept). So, in your example it might be something like:

stan_glmer(..., prior = normal(location = c(0, 0.4, 0, 0),
                               scale = c(10, 0.25, 10, 10))

The technical post webpages of this site follow the CC BY-SA 4.0 protocol. If you need to reprint, please indicate the site URL or the original address.Any question please contact:yoyou2525@163.com.

 
粤ICP备18138465号  © 2020-2024 STACKOOM.COM