Stan/RStan cheat sheets

Stan

Stan reference

Code blocks

• functions | data | transformed data | parameters | transformed parameters | model | generated quantities
• semicolon after statements (variable assignments, sampling statements)
• variable definitions at the beginning of blocks; new local variables can be defined at the beginning of a new {}-block
• dont use dots in variable names

Containers

• matrix[M, N] x | real x[M, N] | vector[M] v | int<lower=1> R[N] (and more)
• a vector[K] x[N] defines a collection of N vectors, each of length K; in R, store as matrix(..., nrow=N, ncol=K) or array(..., dim=c(N,K))

Loops, conditionals

• for (i in 1:N) { ... }
• while (i <= N) { ... }
• if (i == 1) { ... } else if (i == 2) { ... } else { ... }

Sampling statements

• x ~ normal(mu, sigma)
• target += normal_lpdf(x | mu, sigma)

RStan

RStan reference | Vignette

Compile model

• model = stan_model(file='model.stan') (model code in file)
• model = stan_model(model_code=code) (model code in string variable)

Sample from a model

• data is reused from the workspace, argument data=list(...) is optional
• sample = sampling(model, ...) (sample, using compiled object from stan_model)
• sample = sampling(model, pars='mu') (sample, only save parameter mu)
• sample = stan(file=...) (compile from file and sample)
• sample = stan(model_code=...) (compile from string and sample)
• sample = stan(fit=...) (sample, using an object returned by previous call of stan)
• sample = stan(..., data=list(...)) (pass data explicitly)

Reproducibility

• stan(... , init=init, seed=123)
• where init = list(list(mu=...,sigma=...), list(mu=..., sigma=...), ...) where length of list = number of chains, or init = function(chain_id) list(mu=..., sigma=...)

Parallel sampling

• stan(... , chains=8, cores=8)
• sflist2stanfit(parallel::mclapply(1:4, mc.cores=4, function(i) stan(..., chains=1, chain_id=i)))

Chain summary statistics

• traceplot(sample, pars=c('mu', 'sigma'))
• stan_diag(sample) (acceptance rate, log-posterior)
• stan_scat(sample, c('mu', 'sigma')) (scatter plot)
• stan_rhat(sample) (Rhat statistic)
• stan_mcse(sample) (Monte-Carlo std error / posterior sd)
• stan_ess(sample) (effective sample size / sample size)
• stan_ac(sample) (autocorrelation)
• pairs(sample) (scatter plot matrix)

Summary/print

• print(sample)
• summary(sample) (like print for each chain)
• show(sample)
• plot(sample, pars=c('mu', 'sigma')) (interval plots)
• stan_hist(sample, pars=...) (histograms)
• stan_dens(sample, pars=...) (density plots)

Extracting draws

• extract(sample, pars=c('mu', 'sigma')) (returns list, not including warmup)
• extract(sample, permuted=FALSE) (returns 3d array (step, chain, variable), w/o warmup)
• extract(sample, permuted=FALSE, inc_warmup=TRUE)

Misc

• leave-one-out cross validation: R package loo
• to sample from the prior, set length of data N=0 and data x=numeric()
• suppress C++ compiler warnings: withr::with_makevars(new=c(CXXFLAGS='-w'), assignment='+=', code={fit = stan(...)})