6. Comparison of Fixed Weights

Matt Secrest and Isaac Gravestock

Introduction

In a psborrow2 analysis it is possible to specify fixed weights for an observation’s log-likelihood contribution. This is similar to a weighted regression or a fixed power prior parameter.

This vignette will show how weights can be specified and compare regression model results with other packages. We will compare a glm model with weights, a weighted likelihood in Stan with psborrow2, and BayesPPD::glm.fixed.a0 for generalized linear models with fixed a0 (power prior parameter).

Note that we’ll need cmdstanr to run this analysis. Please install cmdstanr if you have not done so already following this guide.

library(psborrow2)
library(cmdstanr)
library(BayesPPD)
# Error in `library()`:
# ! there is no package called 'BayesPPD'
library(ggplot2)

Logistic regression

We fit logistic regression models with the external control arm having weights (or power parameters) equal to 0, 0.25, 0.5, 0.75, 1. The internal treated and control patients have weight = 1. The model has a treatment indicator and two covariates, resp ~ trt + cov1 + cov2.

glm

logistic_glm_reslist <- lapply(c(0, 0.25, 0.5, 0.75, 1), function(w) {
  logistic_glm <- glm(
    resp ~ trt + cov1 + cov2,
    data = as.data.frame(example_matrix),
    family = binomial,
    weights = ifelse(example_matrix[, "ext"] == 1, w, 1)
  )
  glm_summary <- summary(logistic_glm)$coef
  ci <- confint(logistic_glm)
  data.frame(
    fitter = "glm",
    borrowing = w,
    variable = c("(Intercept)", "trt", "cov1", "cov2"),
    estimate = glm_summary[, "Estimate"],
    lower = ci[, 1],
    upper = ci[, 2]
  )
})

BayesPPD

set.seed(123)
logistic_ppd_reslist <- lapply(c(0, 0.25, 0.5, 0.75, 1), function(w) {
  logistic_ppd <- glm.fixed.a0(
    data.type = "Bernoulli",
    data.link = "Logistic",
    y = example_matrix[example_matrix[, "ext"] == 0, ][, "resp"],
    x = example_matrix[example_matrix[, "ext"] == 0, ][, c("trt", "cov1", "cov2")],
    historical = list(list(
      y0 = example_matrix[example_matrix[, "ext"] == 1, ][, "resp"],
      x0 = example_matrix[example_matrix[, "ext"] == 1, ][, c("cov1", "cov2")],
      a0 = w
    )),
    lower.limits = rep(-100, 5),
    upper.limits = rep(100, 5),
    slice.widths = rep(1, 5),
    nMC = 10000,
    nBI = 1000
  )[[1]]
  ci <- apply(logistic_ppd, 2, quantile, probs = c(0.025, 0.975))
  data.frame(
    fitter = "BayesPPD",
    borrowing = w,
    variable = c("(Intercept)", "trt", "cov1", "cov2"),
    estimate = colMeans(logistic_ppd),
    lower = ci[1, ],
    upper = ci[2, ]
  )
})
# Error in `glm.fixed.a0()`:
# ! could not find function "glm.fixed.a0"

psborrow2

logistic_psb_reslist <- lapply(c(0, 0.25, 0.5, 0.75, 1), function(w) {
  logistic_psb2 <- create_analysis_obj(
    data_matrix = as.matrix(cbind(
      example_matrix,
      w = ifelse(example_matrix[, "ext"] == 1, w, 1)
    )),
    covariates = add_covariates(c("cov1", "cov2"),
      priors = prior_normal(0, 100)
    ),
    borrowing = borrowing_full("ext"),
    treatment = treatment_details("trt", prior_normal(0, 100)),
    outcome = outcome_bin_logistic("resp",
      baseline_prior = prior_normal(0, 1000),
      weight_var = "w"
    ),
    quiet = TRUE
  )

  mcmc_logistic_psb2 <- mcmc_sample(logistic_psb2, chains = 1, verbose = FALSE, seed = 123)
  mcmc_summary <- mcmc_logistic_psb2$summary(
    variables = c("alpha", "beta_trt", "beta[1]", "beta[2]"),
    mean,
    ~ quantile(.x, probs = c(0.025, 0.975))
  )

  data.frame(
    fitter = "psborrow2",
    borrowing = w,
    variable = c("(Intercept)", "trt", "cov1", "cov2"),
    estimate = mcmc_summary$mean,
    lower = mcmc_summary$`2.5%`,
    upper = mcmc_summary$`97.5%`
  )
})

Results

logistic_res_df <- do.call(
  rbind,
  c(logistic_glm_reslist, logistic_ppd_reslist, logistic_psb_reslist)
)
# Error:
# ! object 'logistic_ppd_reslist' not found

logistic_res_df$est_ci <- sprintf(
  "%.3f (%.3f, %.3f)",
  logistic_res_df$estimate, logistic_res_df$lower, logistic_res_df$upper
)
# Error:
# ! object 'logistic_res_df' not found

wide <- reshape(
  logistic_res_df[, c("fitter", "borrowing", "variable", "est_ci")],
  direction = "wide",
  timevar = "fitter",
  idvar = c("borrowing", "variable"),
)
# Error:
# ! object 'logistic_res_df' not found
new_order <- order(wide$variable, wide$borrowing)
# Error:
# ! object 'wide' not found
knitr::kable(wide[new_order, ], digits = 3, row.names = FALSE)
# Error:
# ! object 'wide' not found
logistic_res_df$borrowing_x <- logistic_res_df$borrowing +
  (as.numeric(factor(logistic_res_df$fitter)) - 3) / 100
# Error:
# ! object 'logistic_res_df' not found

ggplot(logistic_res_df, aes(x = borrowing_x, y = estimate, group = fitter, colour = fitter)) +
  geom_errorbar(aes(ymin = lower, ymax = upper)) +
  geom_point() +
  facet_wrap(~variable, scales = "free")
# Error:
# ! object 'logistic_res_df' not found

Exponential models

Now we fit models with an exponentially distributed outcome. There is no censoring in this data set. For glm we use family = Gamma(link = "log") and specify fixed dispersion = 1 to fit a exponential model. As before, the external control arm having weights (or power parameters) equal to 0, 0.25, 0.5, 0.75, 1. The internal treated and control patients have weight = 1. The model has a treatment indicator and two covariates, eventtime ~ trt + cov1 + cov2.

set.seed(123)
sim_data_exp <- cbind(
  simsurv::simsurv(
    dist = "exponential",
    x = as.data.frame(example_matrix[, c("trt", "cov1", "cov2", "ext")]),
    betas = c("trt" = 1.3, "cov1" = 1, "cov2" = 0.1, "ext" = -0.4),
    lambdas = 5
  ),
  example_matrix[, c("trt", "cov1", "cov2", "ext")],
  censor = 0
)
head(sim_data_exp)
#   id  eventtime status trt cov1 cov2 ext censor
# 1  1 0.14802638      1   0    0    1   0      0
# 2  2 0.05065174      1   0    1    0   0      0
# 3  3 0.01727805      1   0    1    0   0      0
# 4  4 0.13168620      1   0    1    0   0      0
# 5  5 0.07740706      1   0    0    0   0      0
# 6  6 0.04999778      1   0    1    0   0      0
## glm
glm_reslist <- lapply(c(0, 0.25, 0.5, 0.75, 1), function(w) {
  exp_glm <- glm(
    eventtime ~ trt + cov1 + cov2,
    data = sim_data_exp,
    family = Gamma(link = "log"),
    weights = ifelse(sim_data_exp$ext == 1, w, 1)
  )
  glm_summary <- summary(exp_glm, dispersion = 1)
  est <- -glm_summary$coefficients[, "Estimate"]
  lower <- est - 1.96 * glm_summary$coefficients[, "Std. Error"]
  upper <- est + 1.96 * glm_summary$coefficients[, "Std. Error"]

  data.frame(
    fitter = "glm",
    borrowing = w,
    variable = c("(Intercept)", "trt", "cov1", "cov2"),
    estimate = est,
    lower = lower,
    upper = upper
  )
})
## BayesPPD
set.seed(123)
ppd_reslist <- lapply(c(0, 0.25, 0.5, 0.75, 1), function(w) {
  exp_ppd <- glm.fixed.a0(
    data.type = "Exponential",
    data.link = "Log",
    y = sim_data_exp[sim_data_exp$ext == 0, ]$eventtime,
    x = as.matrix(sim_data_exp[sim_data_exp$ext == 0, c("trt", "cov1", "cov2")]),
    historical = list(list(
      y0 = sim_data_exp[sim_data_exp$ext == 1, ]$eventtime,
      x0 = as.matrix(sim_data_exp[sim_data_exp$ext == 1, c("cov1", "cov2")]),
      a0 = w
    )),
    lower.limits = rep(-100, 5),
    upper.limits = rep(100, 5),
    slice.widths = rep(1, 5),
    nMC = 10000,
    nBI = 1000
  )[[1]]
  ci <- apply(exp_ppd, 2, quantile, probs = c(0.025, 0.975))
  data.frame(
    fitter = "BayesPPD",
    borrowing = w,
    variable = c("(Intercept)", "trt", "cov1", "cov2"),
    estimate = colMeans(exp_ppd),
    lower = ci[1, ],
    upper = ci[2, ]
  )
})
# Error in `glm.fixed.a0()`:
# ! could not find function "glm.fixed.a0"
## psborrow2
psb_reslist <- lapply(c(0, 0.25, 0.5, 0.75, 1), function(w) {
  exp_psb2 <- create_analysis_obj(
    data_matrix = as.matrix(cbind(sim_data_exp, w = ifelse(example_matrix[, "ext"] == 1, w, 1))),
    covariates = add_covariates(c("cov1", "cov2"),
      priors = prior_normal(0, 100)
    ),
    borrowing = borrowing_full("ext"),
    treatment = treatment_details("trt", prior_normal(0, 100)),
    outcome = outcome_surv_exponential("eventtime", "censor",
      baseline_prior = prior_normal(0, 1000),
      weight_var = "w"
    ),
    quiet = TRUE
  )

  mcmc_exp_psb2 <- mcmc_sample(exp_psb2, chains = 1, verbose = FALSE, seed = 123)
  mcmc_summary <- mcmc_exp_psb2$summary(
    variables = c("alpha", "beta_trt", "beta[1]", "beta[2]"),
    mean,
    ~ quantile(.x, probs = c(0.025, 0.975))
  )

  data.frame(
    fitter = "psborrow2",
    borrowing = w,
    variable = c("(Intercept)", "trt", "cov1", "cov2"),
    estimate = mcmc_summary$mean,
    lower = mcmc_summary$`2.5%`,
    upper = mcmc_summary$`97.5%`
  )
})
knitr::knit_hooks$set(output = output_hook)

Results

Note: Wald confidence intervals are displayed here for glm for the exponential models.

res_df <- do.call(rbind, c(glm_reslist, ppd_reslist, psb_reslist))
# Error:
# ! object 'ppd_reslist' not found

res_df$est_ci <- sprintf(
  "%.3f (%.3f, %.3f)",
  res_df$estimate, res_df$lower, res_df$upper
)
# Error:
# ! object 'res_df' not found

wide <- reshape(
  res_df[, c("fitter", "borrowing", "variable", "est_ci")],
  direction = "wide",
  timevar = "fitter",
  idvar = c("borrowing", "variable"),
)
# Error:
# ! object 'res_df' not found
new_order <- order(wide$variable, wide$borrowing)
# Error:
# ! object 'wide' not found
knitr::kable(wide[new_order, ], digits = 3, row.names = FALSE)
# Error:
# ! object 'wide' not found
res_df$borrowing_x <- res_df$borrowing +
  (as.numeric(factor(res_df$fitter)) - 3) / 100
# Error:
# ! object 'res_df' not found

ggplot(res_df, aes(x = borrowing_x, y = estimate, group = fitter, colour = fitter)) +
  geom_errorbar(aes(ymin = lower, ymax = upper)) +
  geom_point() +
  facet_wrap(~variable, scales = "free")
# Error:
# ! object 'res_df' not found