---
title: "Analysis Example of Bayesian MCPMod for Continuous Data"
format: 
  html:
    fig-height: 3.5
    self-contained: true
    toc: true
    number-sections: true
    bibliography: references.bib
    code-summary: setup
    #code-fold: true
    message: false
    warning: false
vignette: >
  %\VignetteIndexEntry{Analysis Example of Bayesian MCPMod for Continuous Data}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---



```{r}
#| code-summary: setup
#| code-fold: true
#| message: false
#| warning: false

library(BayesianMCPMod)
library(RBesT)
library(clinDR)
library(dplyr)
library(tibble)
library(reactable)

set.seed(7015)

#' Display Parameters Table
#'
#' This function generates a markdown table displaying the names and values of parameters
#' from a named list.
#'
#' @param named_list A named list where each name represents a parameter name and the list
#'   element represents the parameter value. Date values in the list are automatically
#'   converted to character strings for display purposes.
#'
#' @return Prints a markdown table with two columns: "Parameter Name" and "Parameter Values".
#'   The function does not return a value but displays the table directly to the output.
#'
#' @importFrom knitr kable
#' @examples
#' params <- list("Start Date" = as.Date("2020-01-01"),
#'                "End Date" = as.Date("2020-12-31"),
#'                "Threshold" = 10)
#' display_params_table(params)
#'
#' @export
display_params_table <- function(named_list) {
  display_table <- data.frame()
  value_names <- data.frame()
  for (i in 1:length(named_list)) {
    # dates will display as numeric by default, so convert to char first
    if (class(named_list[[i]]) == "Date") {
      named_list[[i]] = as.character(named_list[[i]])
    }
    if (!is.null(names(named_list[[i]]))) {
      value_names <- rbind(value_names, paste(names(named_list[[i]]), collapse = ', '))
    }
    values <- data.frame(I(list(named_list[[i]])))
    display_table <- rbind(display_table, values)
  }
  
  round_numeric <- function(x, digits = 3) {
    if (is.numeric(x)) {
      return(round(x, digits))
    } else {
      return(x)
    }
  }
  
  display_table[1] <- lapply(display_table[1], function(sublist) {
    lapply(sublist, round_numeric)
  })
  
  class(display_table[[1]]) <- "list"
  
  if (nrow(value_names) == 0) {
    knitr::kable(
      cbind(names(named_list), display_table),
      col.names = c("Name", "Value")
    )
  } else {
    knitr::kable(
      cbind(names(named_list), value_names, display_table),
      col.names = c("Name", "Value Labels", "Value")
    )
  }
}
```

# Introduction 

This vignette demonstrates the application of the {BayesianMCPMod} package for 
analyzing a phase 2 dose-finding trial using the Bayesian MCPMod approach.

# Calculation of a MAP prior

In a first step, a meta analytic prior will be calculated using historical data from 4 trials with main endpoint Change from baseline in MADRS score after 8 weeks.
Please note that only information from the control group will be integrated leading to an informative mixture prior for the control group, while for the active groups a non-informative prior will be specified.
```{r Historical Data for Control Arm}
data("metaData")
dataset     <- filter(as.data.frame(metaData), bname == "BRINTELLIX")
histcontrol <- filter(
  dataset,
  dose       == 0,
  primtime   == 8,
  indication == "MAJOR DEPRESSIVE DISORDER",
  protid     != 5)

hist_data   <- data.frame(
  trial = histcontrol$nctno,
  est   = histcontrol$rslt,
  se    = histcontrol$se,
  sd    = histcontrol$sd,
  n     = histcontrol$sampsize)
```
Here, we suggest a function to construct a list of prior distributions for the different dose groups.
This function is adapted to the needs of this example.
Other applications may need a different way to construct prior distributions.
```{r Defining MAP prior function}
getPriorList <- function (
  
  hist_data,
  dose_levels,
  dose_names    = NULL,
  robust_weight = 0.5
  
) {
  
  sd_tot <- with(hist_data, sum(sd * n) / sum(n))
  
  gmap <- RBesT::gMAP(
    formula    = cbind(est, se) ~ 1 | trial,
    weights    = hist_data$n,
    data       = hist_data,
    family     = gaussian,
    beta.prior = cbind(0, 100 * sd_tot),
    tau.dist   = "HalfNormal",
    tau.prior  = cbind(0, sd_tot / 4))
  
  prior_ctr <- RBesT::automixfit(gmap)
  
  if (!is.null(robust_weight)) {
    
    prior_ctr <- suppressMessages(RBesT::robustify(
      priormix = prior_ctr,
      weight   = robust_weight,
      sigma    = sd_tot))
    
  }
  
  prior_trt <- RBesT::mixnorm(
    comp1 = c(w = 1, m = summary(prior_ctr)[1], n = 1),
    sigma = sd_tot,
    param = "mn")
  
  prior_list <- c(list(prior_ctr),
                  rep(x     = list(prior_trt),
                      times = length(dose_levels[-1])))
  
  if (is.null(dose_names)) {
    
    dose_names <- c("Ctr", paste0("DG_", seq_along(dose_levels[-1])))
    
  }
  
  names(prior_list) <- dose_names
  
  return (prior_list)
  
}
```
With the dose levels to be investigated, the prior distribution can be constructed.
```{r Getting the MAP prior}
dose_levels <- c(0, 2.5, 5, 10)

prior_list  <- getPriorList(
  hist_data     = hist_data,
  dose_levels   = dose_levels,
  robust_weight = 0.3)

getESS(prior_list)
```

# Dose-Response Model Shapes

Candidate models are specified using the {DoseFinding} package. Models can be 
parameterized using guesstimates or by directly providing distribution parameters. 
Note that the linear candidate model does not require parameterization.

[**Note:** The LinLog model is rarely used and not currently supported by `{BayesianMCPMod}`.]{.aside}

In the code below, the models are "guesstimated" using the `DoseFinding::guesst` function. 
The `d` option usually takes a single value (a dose level), and the corresponding `p` 
for the maximum effect achieved at `d`. 


```{r}
# Guesstimate estimation
exp_guesst  <- DoseFinding::guesst(
  model = "exponential", 
  d = 5, p = 0.2, Maxd = max(dose_levels)
)
emax_guesst <- DoseFinding::guesst(
  model = "emax",
  d = 2.5, p = 0.9
)
sigEmax_guesst <- DoseFinding::guesst(
  model = "sigEmax",
  d = c(2.5, 5), p = c(0.5, 0.95)
)
logistic_guesst <- DoseFinding::guesst(
  model = "logistic",
  d = c(5, 10), p = c(0.1, 0.85)
)
```

In some cases, you need to provide more information. For instance, `sigEmax` 
requires a pair of `d` and `p` values, and `exponential` requires the specification of 
the maximum dose for the trial (`Maxd`). 

[See the help files for model specifications by typing `?DoseFinding::guesst` in your console]{.aside}



Of course, you can also specify the models directly on the parameter scale (without using `DoseFinding::guesst`).

For example, you can get a betaMod model by specifying `delta1` and `delta2` 
parameters (`scale` is assumed to be `1.2` of the maximum dose), or a quadratic model with the `delta2` parameter.

```{r}
betaMod_params <- c(delta1 = 1, delta2 = 1)
quadratic_params <- c(delta2 = -0.1)
```

Now, we can go ahead and create a `Mods` object, which will be used in the remainder
of the vignette.

```{r}
mods <- DoseFinding::Mods(
  linear      = NULL,
  # guesstimate scale
  exponential = exp_guesst,
  emax        = emax_guesst,
  sigEmax     = sigEmax_guesst,
  logistic    = logistic_guesst,
  # parameter scale
  betaMod     = betaMod_params,
  quadratic   = quadratic_params,
  # Options for all models
  doses       = dose_levels,
  maxEff      = -1,
  placEff     = -12.8
)

plot(mods)
```

The `mods` object we just created above contains the full model parameters, which can be helpful for 
understanding how the guesstimates are translated onto the parameter scale. 

```{r}
display_params_table(mods)
```

And we can see the assumed treatment effects for the specified dose groups below:

```{r}
knitr::kable(DoseFinding::getResp(mods, doses = dose_levels))
```

## Trial Data

We will use the trial with ct.gov number NCT00735709 as our phase 2 trial data,
available in the `{clinDR}` package [@nct00735709_2024a].

```{r}
data("metaData")

trial_data <- dplyr::filter(
  dplyr::filter(tibble::tibble(metaData), bname == "BRINTELLIX"),
  primtime == 8,
  indication == "MAJOR DEPRESSIVE DISORDER",
  protid == 5
)

n_patients <- c(128, 124, 129, 122)
```

# Posterior Calculation

In the first step of Bayesian MCPMod, the posterior is calculated by combining 
the prior information with the estimated results of the trial [@fleischer_2022]. 

```{r}
posterior <- getPosterior(
  prior_list = prior_list,
  mu_hat = trial_data$rslt,
  S_hat = trial_data$se,
  calc_ess = TRUE
)

knitr::kable(summary(posterior))
```

# Bayesian MCPMod Test Step

The testing step of Bayesian MCPMod is executed using a critical value on the 
probability scale and a pseudo-optimal contrast matrix. 

The critical value is calculated using (re-estimated) contrasts for frequentist 
MCPMod to ensure error control when using weakly-informative priors.

A pseudo-optimal contrast matrix is generated based on the variability of the 
posterior distribution (see [@fleischer_2022] for more details).

```{r}
crit_pval <- getCritProb(
  mods           = mods,
  dose_levels    = dose_levels,
  se_new_trial   = trial_data$se,
  alpha_crit_val = 0.05
)

contr_mat <- getContr(
  mods         = mods,
  dose_levels  = dose_levels,
  sd_posterior = summary(posterior)[, 2]
)
```

Please note that there are different ways to derive the contrasts.
The following code shows the implementation of some of these ways but it is not 
executed and the contrast specification above is used.

```{r}
#| eval: false

# i) the frequentist contrast
contr_mat_prior <- getContr(
  mods           = mods,
  dose_levels    = dose_levels,
  dose_weights   = n_patients,
  prior_list     = prior_list)
# ii) re-estimated frequentist contrasts
contr_mat_prior <- getContr(
  mods           = mods,
  dose_levels    = dose_levels,
  se_new_trial   = trial_data$se)
# iii)  Bayesian approach using number of patients for new trial and prior distribution
contr_mat_prior <- getContr(
  mods           = mods,
  dose_levels    = dose_levels,
  dose_weights   = n_patients,
  prior_list     = prior_list)
```

The Bayesian MCP testing step is then executed:  

```{r}
BMCP_result <- performBayesianMCP(
  posterior_list = posterior,
  contr          = contr_mat, 
  crit_prob_adj  = crit_pval)
```

Summary information:

```{r}
BMCP_result
```

The testing step is significant, indicating a non-flat dose-response shape.
All models are significant, with the `emax` model indicating the greatest deviation 
from the null hypothesis.

# Model Fitting and Visualization

In the model fitting step the posterior distribution is used as basis.

Both simplified and full fitting are performed. 

For the simplified fit, the multivariate normal distribution of the control group 
is approximated and reduced by a one-dimensional normal distribution.

The actual fit (on this approximated posterior distribution) is then performed 
using generalized least squares criterion. In contrast, for the full fit, the 
non-linear optimization problem is addressed via the Nelder-Mead algorithm 
[@neldermead_2024a] implemented by the `{nloptr}` package.

The output of the fit includes information about the predicted effects for the 
included dose levels, the generalized AIC, and the corresponding weights.

For the considered case, the simplified and the full fit are very similar, so we
present the full fit.

```{r}
# If simple = TRUE, uses approx posterior
# Here we use complete posterior distribution
fit <- getModelFits(
  models      = mods,
  dose_levels = dose_levels,
  posterior   = posterior,
  simple      = FALSE)
```

Estimates for dose levels not included in the trial:

```{r}
display_params_table(stats::predict(fit, doses = c(0, 2.5, 4, 5, 7, 10)))
```

Plots of fitted dose-response models and an AIC-based average model:

```{r}
plot(fit)
```

To assess the uncertainty, one can additionally visualize credible bands 
(orange shaded areas, default levels are 50% and 95%).

These credible bands are calculated with a bootstrap method as follows:

- Samples from the posterior distribution are drawn and for every sample the 
simplified fitting step and a prediction is performed.

- These predictions are then used to identify and visualize the specified quantiles.

```{r}
plot(fit, cr_bands = TRUE)
```

The bootstrap based quantiles can also be directly calculated via the 
`getBootstrapQuantiles()` function.

For this example, only 6 quantiles are bootstrapped for each model fit.

```{r}
bootstrap_quantiles <- getBootstrapQuantiles(
  model_fits = fit,
  quantiles  = c(0.025, 0.5, 0.975),
  doses      = c(0, 2.5, 4, 5, 7, 10),
  n_samples  = 6
)
```

```{r}
#| code-fold: true
reactable::reactable(
  data = bootstrap_quantiles,
  groupBy = "models",
  columns = list(
    doses = colDef(aggregate = "count", format = list(aggregated = colFormat(suffix = " doses"))),
    "2.5%" = colDef(aggregate = "mean", format = list(aggregated = colFormat(prefix = "mean = ", digits = 2), cell = colFormat(digits = 4))),
    "50%" = colDef(aggregate = "mean", format = list(aggregated = colFormat(prefix = "mean = ", digits = 2), cell = colFormat(digits = 4))),
    "97.5%" = colDef(aggregate = "mean", format = list(aggregated = colFormat(prefix = "mean = ", digits = 2), cell = colFormat(digits = 4)))
  )
)
```

**Technical note:** The median quantile of the bootstrap based procedure is not 
necessary similar to the main model fit, as they are derived via different procedures.

The main fit (black line) minimizes residuals for the posterior distribution, 
while the bootstrap median is the median fit of random sampling.

# Additional note

Testing and modeling can also be combined via `performBayesianMCPMod()`, 
but this is not run here.

```{r}
#| eval: false
performBayesianMCPMod(
  posterior_list   = posterior,
  contr            = contr_mat,
  crit_prob_adj    = crit_pval,
  simple           = FALSE)
```