bayesian credible intervals on veris data

bayesian_credible_intervals.R

# pick an enumeration
enum <- "action.*.variety"
# establish filter criteria (easier than a complex standard-eval  filter_ line)
df <- vcdb %>%
  dplyr::filter(plus.dbir_year == 2016, subset.2017dbir) %>%
  dplyr::filter(attribute.confidentiality.data_disclosure.Yes) %>%
  dplyr::filter(victim.industry2.92)

# establish priors from previous year
priors <- df %>%
  dbirR::getenumCI(enum, short.names=FALSE) %>%
  dplyr::select(enum, p1 = freq) %>%
  dplyr::mutate(p2 = 1-p1)

# calculate this year's data.  Add wilson CI's for reference.  
chunk <- df %>%
  dbirR::getenumCI(enum, ci.level=0.95, ci.method="wilson", short.names=FALSE) %>%
  dplyr::left_join(priors, by="enum")


# remove those without sample sizes
chunk <- chunk[!is.na(chunk$n), ]

# calculate bayesian credible interval and bind the rows to the chunk with wilson CIs.
chunk <- rbind(
  cbind(
    chunk %>%
      dplyr::select(enum),
    binom::binom.bayes(chunk$x, chunk$n, conf.level=0.95, type="highest", prior.shape1=chunk$p1, prior.shape2=chunk$p2)
  ) %>%
    dplyr::select(enum,    x,    n,    freq=mean, method,   lower,   upper),
  chunk %>%
    dplyr::select(enum,    x,    n,    freq, method,   lower,   upper)
)

# plot the comparison of the results
chunk %>%
  dplyr::filter(x!=0) %>%
  dplyr::mutate(method = as.character(method)) %>%
  dplyr::mutate(method = ifelse(method == "bayes*", "bayes", method)) %>%
  tidyr::gather("loc", "value", lower, upper) %>%
  dplyr::mutate(group = paste(enum, method)) %>%
  ggplot(aes(x=value, y=enum, group=group, color=method)) +
  geom_point(aes(shape=method), alpha=0.5) +
  geom_line(alpha=0.5) +
  ggthemes::scale_color_tableau()