Alternative PEM casting to long

[mattsecrest]

In the end, there may be too many new interdependencies that are not worth keeping. We may want to keep your approach and just put some safeguards around the cutpoints chosen. It is also much slower for dynamic borrowing than it is for no borrowing. Anyways, have a peek and let me know what you think @gravesti.

Similar to yours, shows comparable results:

# Non Bayesian piecewise model
library(eha)
pem_eha <- eha::pchreg(Surv(time, status) ~ trt + cov1 + cov2, data = as.data.frame(psborrow2::example_matrix), cuts = c(0, 10, 20, 30, 40, 100))
summary(pem_eha)
pem_eha$convergence
pem_eha$hazards

# piecewise exponential with no borrowing 
pem_psb2 <- create_analysis_obj(
  data_matrix = psborrow2::example_matrix,
  outcome = outcome_surv_pem(time_var = "time", cens_var = "cnsr", cut_points = seq(10, 40, by = 10),
                                       baseline_prior = prior_normal(0, 1000)),
  borrowing = borrowing_full(ext_flag_col = "ext"),
  covariates = add_covariates(c("cov1", "cov2"), priors = prior_normal(0, 1000)),
  treatment = treatment_details("trt", prior_normal(0, 1000)),
)

samples <- psborrow2::mcmc_sample(pem_psb2, parallel_chains = 4, verbose = TRUE, iter_warmup = 5000)

## results seem to match
samples
log(pem_eha$hazards)
summary(pem_eha)
exp(coef(pem_eha))

Still to do if we go with this approach.

• Test cases
• Variable dictionary