In wlandau/targets-stan: An Example Targets Pipeline for Bayesian Data Analysis

knitr::opts_chunk$set(echo = TRUE)
options(tidyverse.quiet = TRUE)

This report contains the validation results of a small Bayesian model. Here, we summarize the results computed in earlier targets of the pipeline. We reference our targets with tar_load() and tar_read(). This ensures

1. Because of the tar_render() function from the tarchetypes package (see _targets.R) targets automatically detects the dependencies of this report and rebuilds it when its dependencies change.
2. We can run the report by itself if the targets are already in the _targets/ data store.

Continuous covariate

library(targets)
library(tidyverse)
tar_load(fit_continuous)

Our results are in a data frame with one row per simulated dataset and columns with information about our fitted models.

fit_continuous

If we implemented the model in stan/model.stan correctly, then roughly 90% of model fits should cover the true beta parameter that generated the data in 90% credible intervals.

mean(fit_continuous$cover_beta)

The posterior median of beta should be reasonably close to the true value.

ggplot(fit_continuous) +
  geom_point(aes(x = beta_true, y = median)) +
  geom_abline(intercept = 0, slope = 1) +
  theme_gray(16)

We should also check convergence diagnostics. rhat should ideally be close to 1.

ggplot(fit_continuous) +
  geom_histogram(aes(x = rhat), bins = 20)

Effective sample size should ideally be high.

ggplot(fit_continuous) +
  geom_histogram(aes(x = ess_bulk), bins = 20)

ggplot(fit_continuous) +
  geom_histogram(aes(x = ess_tail), bins = 20)

Discrete covariate

tar_load(fit_discrete)

Here the analogous results for the discrete covariate simulations.

fit_discrete

mean(fit_discrete$cover_beta)

The posterior median of beta should be reasonably close to the true value.

ggplot(fit_discrete) +
  geom_point(aes(x = beta_true, y = median)) +
  geom_abline(intercept = 0, slope = 1) +
  theme_gray(16)

We should also check convergence diagnostics. rhat should ideally be close to 1.

ggplot(fit_discrete) +
  geom_histogram(aes(x = rhat), bins = 20)

Effective sample size should ideally be high.

ggplot(fit_discrete) +
  geom_histogram(aes(x = ess_bulk), bins = 20)

ggplot(fit_discrete) +
  geom_histogram(aes(x = ess_tail), bins = 20)

wlandau/targets-stan documentation built on April 26, 2021, 11:19 p.m.