Purpose

The goal of this demo is to conduct a Bayesian Dynamic Borrowing (BDB) analysis using the hierarchical commensurate prior on a dataset.

# Load packages ----
library(psborrow2)

# Survival analysis
library(survival)
library(ggsurvfit)
library(flexsurv)

# Additional tools for draws objects
library(bayesplot)
library(posterior)

# Comparing populations
library(table1)

Explore example data

head(example_matrix)

##      id ext trt cov4 cov3 cov2 cov1       time status cnsr resp
## [1,]  1   0   0    1    1    1    0  2.4226411      1    0    1
## [2,]  2   0   0    1    1    0    1 50.0000000      0    1    1
## [3,]  3   0   0    0    0    0    1  0.9674372      1    0    1
## [4,]  4   0   0    1    1    0    1 14.5774738      1    0    1
## [5,]  5   0   0    1    1    0    0 50.0000000      0    1    0
## [6,]  6   0   0    1    1    0    1 50.0000000      0    1    0

?example_matrix # for more details

example_dataframe <- as.data.frame(example_matrix)

table(ext = example_matrix[, "ext"], trt = example_matrix[, "trt"])

##    trt
## ext   0   1
##   0  50 100
##   1 350   0

Naive internal comparisons

Kaplan-Meier curves

km_fit <- survfit(Surv(time = time, event = 1 - cnsr) ~ trt + ext,
  data = example_dataframe
)

ggsurvfit(km_fit)

plot of chunk unnamed-chunk-5

The internal and external control arms look quite different…

Cox model

cox_fit <- coxph(Surv(time = time, event = 1 - cnsr) ~ trt,
  data = example_dataframe,
  subset = ext == 0
)

cox_fit

## Call:
## coxph(formula = Surv(time = time, event = 1 - cnsr) ~ trt, data = example_dataframe, 
##     subset = ext == 0)
## 
##        coef exp(coef) se(coef)      z     p
## trt -0.1097    0.8961   0.1976 -0.555 0.579
## 
## Likelihood ratio test=0.3  on 1 df, p=0.5809
## n= 150, number of events= 114

exp(confint(cox_fit)) # The internal HR is 0.90 (95% CI 0.61 - 1.32)

##         2.5 %   97.5 %
## trt 0.6083595 1.319842

Hybrid control analysis

Let’s start by demonstrating the utility of BDB by trying to borrow data from the external control arm which we know experiences worse survival.

The end goal is to create an Analysis object with:

?create_analysis_obj

A note on prior distributions

psborrow2 allows the user to specify priors with the following functions:

?prior_bernoulli
?prior_beta
?prior_cauchy
?prior_exponential
?prior_gamma
?prior_normal
?prior_poisson
?prior_uniform

Prior distributions can be plotted with the plot() method

plot(prior_normal(0, 1), xlim = c(-100, 100), ylim = c(0, 1))

plot of chunk unnamed-chunk-7

plot(prior_normal(0, 10), xlim = c(-100, 100), ylim = c(0, 1))

plot of chunk unnamed-chunk-7

plot(prior_normal(0, 10000), xlim = c(-100, 100), ylim = c(0, 1))

plot of chunk unnamed-chunk-7

Outcome objects

psborrow2 currently supports 4 outcomes:

?outcome_surv_weibull_ph # Weibull survival w/ proportional hazards
?outcome_surv_exponential # Exponential survival
?outcome_bin_logistic # Logistic binary outcome
?outocme_cont_normal # Normal continuous outcome

Create an exponential survival distribution Outcome object

exp_outcome <- outcome_surv_exponential(
  time_var = "time",
  cens_var = "cnsr",
  baseline_prior = prior_normal(0, 10000)
)

Borrowing object

Borrowing objects are created with:

?borrowing_hierarchical_commensurate # Hierarchical commensurate borrowing
?borrowing_none # No borrowing
?borrowing_full # Full borrowing

bdb_borrowing <- borrowing_hierarchical_commensurate(
  ext_flag_col = "ext",
  tau_prior = prior_gamma(0.001, 0.001)
)

Treatment objects

Treatment objects are created with:

?treatment_details

trt_details <- treatment_details(
  trt_flag_col = "trt",
  trt_prior = prior_normal(0, 10000)
)

Analysis objects

Combine everything and create object of class Analysis:

analysis_object <- create_analysis_obj(
  data_matrix = example_matrix,
  outcome = exp_outcome,
  borrowing = bdb_borrowing,
  treatment = trt_details
)

## Inputs look good.

## Stan program compiled successfully!

## Ready to go! Now call `mcmc_sample()`.

analysis_object

## Analysis Object
## 
## Outcome model: OutcomeSurvExponential 
## Outcome variables: time cnsr 
## 
## Borrowing method: Bayesian dynamic borrowing with the hierarchical commensurate prior 
## External flag: ext 
## 
## Treatment variable: trt 
## 
## Data: Matrix with 500 observations 
##     -  50  internal controls
##     -  350  external controls 
##     -  100  internal experimental
## 
## Stan model compiled and ready to sample.
##  Call mcmc_sample() next.

MCMC sampling

Conduct MCMC sampling with:

?mcmc_sample

results <- mcmc_sample(
  x = analysis_object,
  iter_warmup = 1000,
  iter_sampling = 5000,
  chains = 1
)

class(results)

## [1] "CmdStanMCMC" "CmdStanFit"  "R6"

results

##  variable     mean   median   sd  mad       q5      q95 rhat ess_bulk ess_tail
##  lp__     -1617.94 -1617.65 1.46 1.26 -1620.73 -1616.19 1.00     2005     2457
##  beta_trt    -0.15    -0.16 0.20 0.20    -0.47     0.18 1.00     1933     2381
##  alpha[1]    -3.36    -3.36 0.16 0.17    -3.64    -3.11 1.00     1943     2545
##  alpha[2]    -2.40    -2.40 0.06 0.06    -2.49    -2.31 1.00     2486     2637
##  tau          1.16     0.48 1.82 0.66     0.01     4.64 1.00     2408     1617
##  HR_trt       0.88     0.86 0.18 0.17     0.63     1.20 1.00     1933     2381

Interpret results

Dictionary to interpret parameters:

variable_dictionary(analysis_object)

##   Stan_variable                        Description
## 1           tau         commensurability parameter
## 2      alpha[1] baseline log hazard rate, internal
## 3      alpha[2] baseline log hazard rate, external
## 4      beta_trt                   treatment log HR
## 5        HR_trt                       treatment HR

Create a draws object

draws <- results$draws()

Rename draws object parameters

draws <- rename_draws_covariates(draws, analysis_object)

Get 95% credible intervals with posterior package

posterior::summarize_draws(draws, ~ quantile(.x, probs = c(0.025, 0.50, 0.975)))

## # A tibble: 6 × 4
##   variable                                `2.5%`     `50%`   `97.5%`
##   <chr>                                    <dbl>     <dbl>     <dbl>
## 1 lp__                               -1622.      -1618.    -1616.   
## 2 treatment log HR                      -0.531      -0.156     0.254
## 3 baseline log hazard rate, internal    -3.70       -3.36     -3.06 
## 4 baseline log hazard rate, external    -2.51       -2.40     -2.29 
## 5 commensurability parameter             0.00123     0.481     6.02 
## 6 treatment HR                           0.588       0.856     1.29

Look at histogram of draws with bayesplot package

bayesplot::mcmc_hist(draws, c("treatment HR"))

## Registered S3 method overwritten by 'plyr':
##   method    from  
##   [.indexed table1

## `stat_bin()` using `bins = 30`. Pick better value `binwidth`.

plot of chunk unnamed-chunk-18

Our model does not borrow much from the external arm! This is the desired outcome given how different the control arms were.

Control arm imbalances

Check balance between arms

table1(
  ~ cov1 + cov2 + cov3 + cov4 |
    factor(ext, labels = c("Internal RCT", "External data")) +
      factor(trt, labels = c("Not treated", "Treated")),
  data = example_dataframe
)

Because the imbalance may be conditional on observed covariates, let’s adjust for propensity scores in our analysis

Create a propensity score model

ps_model <- glm(ext ~ cov1 + cov2 + cov3 + cov4,
  data = example_dataframe,
  family = binomial
)
ps <- predict(ps_model, type = "response")
example_dataframe$ps <- ps
example_dataframe$ps_cat_ <- cut(
  example_dataframe$ps,
  breaks = 5,
  include.lowest = TRUE
)
levels(example_dataframe$ps_cat_) <- c(
  "ref", "low",
  "low_med", "high_med", "high"
)

Convert the data back to a matrix with dummy variables for ps_cat_ levels

example_matrix_ps <- create_data_matrix(
  example_dataframe,
  outcome = c("time", "cnsr"),
  trt_flag_col = "trt",
  ext_flag_col = "ext",
  covariates = ~ps_cat_
)

## Call `add_covariates()` with `covariates =  c("ps_cat_low", "ps_cat_low_med", "ps_cat_high_med", "ps_cat_high" ) `

Propensity score analysis without borrowing

anls_ps_no_borrow <- create_analysis_obj(
  data_matrix = example_matrix_ps,
  covariates = add_covariates(
    c("ps_cat_low", "ps_cat_low_med", "ps_cat_high_med", "ps_cat_high"),
    prior_normal(0, 10000)
  ),
  outcome = outcome_surv_exponential("time", "cnsr", prior_normal(0, 10000)),
  borrowing = borrowing_none("ext"),
  treatment = treatment_details("trt", prior_normal(0, 10000))
)

res_ps_no_borrow <- mcmc_sample(
  x = anls_ps_no_borrow,
  iter_warmup = 1000,
  iter_sampling = 5000,
  chains = 1
)

draws_ps_no_borrow <- rename_draws_covariates(
  res_ps_no_borrow$draws(),
  anls_ps_no_borrow
)

summarize_draws(draws_ps_no_borrow, ~ quantile(.x, probs = c(0.025, 0.50, 0.975)))

## # A tibble: 8 × 4
##   variable                   `2.5%`    `50%`   `97.5%`
##   <chr>                       <dbl>    <dbl>     <dbl>
## 1 lp__                     -466.    -462.    -460.    
## 2 treatment log HR           -0.736   -0.342    0.0575
## 3 baseline log hazard rate   -4.76    -4.20    -3.71  
## 4 ps_cat_low                 -0.342    0.403    1.16  
## 5 ps_cat_low_med              0.454    1.01     1.60  
## 6 ps_cat_high_med             1.42     2.10     2.78  
## 7 ps_cat_high                 2.41     3.01     3.65  
## 8 treatment HR                0.479    0.710    1.06

Propensity score analysis with BDB

anls_ps_bdb <- create_analysis_obj(
  data_matrix = example_matrix_ps,
  covariates = add_covariates(
    c("ps_cat_low", "ps_cat_low_med", "ps_cat_high_med", "ps_cat_high"),
    prior_normal(0, 10000)
  ),
  outcome = outcome_surv_exponential("time", "cnsr", prior_normal(0, 10000)),
  borrowing = borrowing_hierarchical_commensurate("ext", prior_gamma(0.001, 0.001)),
  treatment = treatment_details("trt", prior_normal(0, 10000))
)

res_ps_bdb <- mcmc_sample(
  x = anls_ps_bdb,
  iter_warmup = 1000,
  iter_sampling = 5000,
  chains = 1
)

draws_ps_bdb <- rename_draws_covariates(
  res_ps_bdb$draws(),
  anls_ps_bdb
)

summarize_draws(draws_ps_bdb, ~ quantile(.x, probs = c(0.025, 0.50, 0.975)))

## # A tibble: 10 × 4
##    variable                               `2.5%`     `50%`    `97.5%`
##    <chr>                                   <dbl>     <dbl>      <dbl>
##  1 lp__                               -1426.     -1421.    -1418.    
##  2 treatment log HR                      -0.687     -0.344    -0.0294
##  3 baseline log hazard rate, internal    -4.62      -4.19     -3.77  
##  4 baseline log hazard rate, external    -4.64      -4.19     -3.80  
##  5 commensurability parameter             0.0759    56.2    1653.    
##  6 ps_cat_low                            -0.305      0.259     0.830 
##  7 ps_cat_low_med                         0.626      1.06      1.52  
##  8 ps_cat_high_med                        1.64       2.09      2.57  
##  9 ps_cat_high                            2.53       2.94      3.39  
## 10 treatment HR                           0.503      0.709     0.971

It looks like PS + BDB allowed us to most accurately recover the true hazard ratio of 0.70.