Example Complex Analysis Function: Modelling Cox Regression

Emily de la Rua and Gabriel Becker

2024-04-15

Introduction

In this vignette we will demonstrate how a complex analysis function can be constructed in order to build highly-customized tables with rtables. This example will detail the steps in creating an analysis function to calculate a basic univariable Cox regression summary table to analyze the treatment effect of the ARM variable and any covariate/interaction effects for a survival analysis. For a Cox regression analysis function with more customization options and the capability of fitting multivariable Cox regression models, see the summarize_coxreg() function from the tern package, which builds upon the concepts used in the construction of this example.

The packages used in this vignette are:

library(rtables)
library(dplyr)

Data Pre-Processing

First, we prepare the data that will be used to generate a table in this example. We will use the example ADTTE (Time-To-Event Analysis) dataset ex_adtte from the formatters package, which contains our treatment variable ARM, several variables that can be chosen as covariates, and censor variable CNSR from which we will derive the event variable EVENT required for our model. For the purpose of this example, we will use age (AGE) and race (RACE) as our covariates.

We prepare the data as needed to observe the desired effects in our summary table. PARAMCD is filtered so that only records of overall survival (OS) are included, and we filter and mutate to include only the levels of interest in our covariates. The ARM variable is mutated to indicate that "B: Placebo" should be used as the reference level of our treatment variable, and the EVENT variable is derived from CNSR.

adtte <- ex_adtte

anl <- adtte %>%
  dplyr::filter(PARAMCD == "OS") %>%
  dplyr::filter(ARM %in% c("A: Drug X", "B: Placebo")) %>%
  dplyr::filter(RACE %in% c("ASIAN", "BLACK OR AFRICAN AMERICAN", "WHITE")) %>%
  dplyr::mutate(RACE = droplevels(RACE)) %>%
  dplyr::mutate(ARM = droplevels(stats::relevel(ARM, "B: Placebo"))) %>%
  dplyr::mutate(EVENT = 1 - CNSR)

Creating Helper Functions: Cox Regression Model Calculations

tidy Method for summary.coxph Objects: tidy.summary.coxph

This method allows the tidy function from the broom package to operate on summary.coxph output, extracting the values of interest to this analysis and returning a tidied tibble::tibble() object.

tidy.summary.coxph <- function(x, ...) {
  is(x, "summary.coxph")
  pval <- x$coefficients
  confint <- x$conf.int
  levels <- rownames(pval)
  pval <- tibble::as_tibble(pval)
  confint <- tibble::as_tibble(confint)

  ret <- cbind(pval[, grepl("Pr", names(pval))], confint)
  ret$level <- levels
  ret$n <- x[["n"]]
  ret
}

Function to Estimate Interaction Effects: h_coxreg_inter_effect

The h_coxreg_inter_effect helper function is used within the following helper function, h_coxreg_extract_interaction, to estimate interaction effects from a given model for a given covariate. The function calculates the desired statistics from the given model and returns a data.frame with label information for each row as well as the statistics n, hr (hazard ratio), lcl (CI lower bound), ucl (CI upper bound), pval (effect p-value), and pval_inter (interaction p-value). If a numeric covariate is selected, the median value is used as the sole “level” for which an interaction effect is calculated. For non-numeric covariates, an interaction effect is calculated for each level of the covariate, with each result returned on a separate row.

h_coxreg_inter_effect <- function(x,
                                  effect,
                                  covar,
                                  mod,
                                  label,
                                  control,
                                  data) {
  if (is.numeric(x)) {
    betas <- stats::coef(mod)
    attrs <- attr(stats::terms(mod), "term.labels")
    term_indices <- grep(pattern = effect, x = attrs[!grepl("strata\\(", attrs)])
    betas <- betas[term_indices]
    betas_var <- diag(stats::vcov(mod))[term_indices]
    betas_cov <- stats::vcov(mod)[term_indices[1], term_indices[2]]
    xval <- stats::median(x)
    effect_index <- !grepl(covar, names(betas))
    coef_hat <- betas[effect_index] + xval * betas[!effect_index]
    coef_se <- sqrt(betas_var[effect_index] + xval^2 * betas_var[!effect_index] + 2 * xval * betas_cov)
    q_norm <- stats::qnorm((1 + control$conf_level) / 2)
  } else {
    var_lvl <- paste0(effect, levels(data[[effect]])[-1]) # [-1]: reference level
    giv_lvl <- paste0(covar, levels(data[[covar]]))
    design_mat <- expand.grid(effect = var_lvl, covar = giv_lvl)
    design_mat <- design_mat[order(design_mat$effect, design_mat$covar), ]
    design_mat <- within(data = design_mat, expr = {
      inter <- paste0(effect, ":", covar)
      rev_inter <- paste0(covar, ":", effect)
    })
    split_by_variable <- design_mat$effect
    interaction_names <- paste(design_mat$effect, design_mat$covar, sep = "/")
    mmat <- stats::model.matrix(mod)[1, ]
    mmat[!mmat == 0] <- 0
    design_mat <- apply(X = design_mat, MARGIN = 1, FUN = function(x) {
      mmat[names(mmat) %in% x[-which(names(x) == "covar")]] <- 1
      mmat
    })
    colnames(design_mat) <- interaction_names
    coef <- stats::coef(mod)
    vcov <- stats::vcov(mod)
    betas <- as.matrix(coef)
    coef_hat <- t(design_mat) %*% betas
    dimnames(coef_hat)[2] <- "coef"
    coef_se <- apply(design_mat, 2, function(x) {
      vcov_el <- as.logical(x)
      y <- vcov[vcov_el, vcov_el]
      y <- sum(y)
      y <- sqrt(y)
      y
    })
    q_norm <- stats::qnorm((1 + control$conf_level) / 2)
    y <- cbind(coef_hat, `se(coef)` = coef_se)
    y <- apply(y, 1, function(x) {
      x["hr"] <- exp(x["coef"])
      x["lcl"] <- exp(x["coef"] - q_norm * x["se(coef)"])
      x["ucl"] <- exp(x["coef"] + q_norm * x["se(coef)"])
      x
    })
    y <- t(y)
    y <- by(y, split_by_variable, identity)
    y <- lapply(y, as.matrix)
    attr(y, "details") <- paste0(
      "Estimations of ", effect, " hazard ratio given the level of ", covar, " compared to ",
      effect, " level ", levels(data[[effect]])[1], "."
    )
    xval <- levels(data[[covar]])
  }
  data.frame(
    effect = "Covariate:",
    term = rep(covar, length(xval)),
    term_label = as.character(paste0("  ", xval)),
    level = as.character(xval),
    n = NA,
    hr = if (is.numeric(x)) exp(coef_hat) else y[[1]][, "hr"],
    lcl = if (is.numeric(x)) exp(coef_hat - q_norm * coef_se) else y[[1]][, "lcl"],
    ucl = if (is.numeric(x)) exp(coef_hat + q_norm * coef_se) else y[[1]][, "ucl"],
    pval = NA,
    pval_inter = NA,
    stringsAsFactors = FALSE
  )
}

Function to Extract Effect Information: h_coxreg_extract_interaction

Using the previous two helper functions, h_coxreg_extract_interaction uses ANOVA to extract information from the given model about the given covariate. This function will extract different information depending on whether the effect of interest is a treatment/main effect or an interaction effect, and returns a data.frame with label information for each row (corresponding to each effect) as well as the statistics n, hr, lcl, ucl, pval, and pval_inter (for interaction effects only). This helper function is used directly within our analysis function to analyze the Cox regression model and extract relevant information to be processed and displayed within our output table.

h_coxreg_extract_interaction <- function(effect, covar, mod, data) {
  control <- list(pval_method = "wald", ties = "exact", conf_level = 0.95, interaction = FALSE)
  test_statistic <- c(wald = "Wald", likelihood = "LR")[control$pval_method]
  mod_aov <- withCallingHandlers(
    expr = car::Anova(mod, test.statistic = test_statistic, type = "III"),
    message = function(m) invokeRestart("muffleMessage")
  )
  msum <- if (!any(attr(stats::terms(mod), "order") == 2)) summary(mod, conf.int = control$conf_level) else mod_aov
  sum_anova <- broom::tidy(msum)
  if (!any(attr(stats::terms(mod), "order") == 2)) {
    effect_aov <- mod_aov[effect, , drop = TRUE]
    pval <- effect_aov[[grep(pattern = "Pr", x = names(effect_aov)), drop = TRUE]]
    sum_main <- sum_anova[grepl(effect, sum_anova$level), ]
    term_label <- if (effect == covar) {
      paste0(levels(data[[covar]])[2], " vs control (", levels(data[[covar]])[1], ")")
    } else {
      unname(formatters::var_labels(data, fill = TRUE)[[covar]])
    }
    y <- data.frame(
      effect = ifelse(covar == effect, "Treatment:", "Covariate:"),
      term = covar, term_label = term_label,
      level = levels(data[[effect]])[2],
      n = mod[["n"]], hr = unname(sum_main["exp(coef)"]), lcl = unname(sum_main[grep("lower", names(sum_main))]),
      ucl = unname(sum_main[grep("upper", names(sum_main))]), pval = pval,
      stringsAsFactors = FALSE
    )
    y$pval_inter <- NA
    y
  } else {
    pval <- sum_anova[sum_anova$term == effect, ][["p.value"]]

    ## Test the interaction effect
    pval_inter <- sum_anova[grep(":", sum_anova$term), ][["p.value"]]
    covar_test <- data.frame(
      effect = "Covariate:",
      term = covar, term_label = unname(formatters::var_labels(data, fill = TRUE)[[covar]]),
      level = "",
      n = mod$n, hr = NA, lcl = NA, ucl = NA, pval = pval,
      pval_inter = pval_inter,
      stringsAsFactors = FALSE
    )
    ## Estimate the interaction
    y <- h_coxreg_inter_effect(
      data[[covar]],
      covar = covar,
      effect = effect,
      mod = mod,
      label = unname(formatters::var_labels(data, fill = TRUE)[[covar]]),
      control = control,
      data = data
    )
    rbind(covar_test, y)
  }
}

Creating a Helper Function: cached_model

Next, we will create a helper function, cached_model, which will be used within our analysis function to cache and return the fitted Cox regression model for the current covariate. The df argument will be directly inherited from the df argument passed to the analysis function, which contains the full dataset being analyzed. The cov argument will be the covariate that is being analyzed depending on the current row context. If the treatment effect is currently being analyzed, this value will be an empty string. The cache_env parameter will be an environment object which is used to store the model for the current covariate, also passed down from the analysis function. Of course, this function can also be run outside of the analysis function and will still cache and return a Cox regression model.

Using these arguments, the cached_model function first checks if a model for the given covariate cov is already stored in the caching environment cache_env. If so, then this model is retrieved and returned by cached_model. If not, the model must be constructed. This is done by first constructing the model formula, model_form, starting with only the treatment effect (ARM) and adding a covariate effect if one is currently being analyzed. Then a Cox regression model is fit using df and the model formula, and this model is both returned and stored in the caching environment object as cache_env[[cov]].

cached_model <- function(df, cov, cache_env) {
  ## Check if a model already exists for
  ## `cov` in the caching environment
  if (!is.null(cache_env[[cov]])) {
    ## If model already exists, retrieve it from cache_env
    model <- cache_env[[cov]]
  } else {
    ## Build model formula
    model_form <- paste0("survival::Surv(AVAL, EVENT) ~ ARM")
    if (length(cov) > 0) {
      model_form <- paste(c(model_form, cov), collapse = " * ")
    } else {
      cov <- "ARM"
    }
    ## Calculate Cox regression model
    model <- survival::coxph(
      formula = stats::as.formula(model_form),
      data = df,
      ties = "exact"
    )
    ## Store model in the caching environment
    cache_env[[cov]] <- model
  }
  model
}

Creating the Analysis Function: a_cox_summary

With our data prepared and helper function created, we can proceed to construct our analysis function a_cox_summary, which will be used to populate all of the rows in our table. In order to be used to generate both data rows (for interaction effects) and content rows (for main effects), we must create a function that can be used as both afun in analyze and cfun in summarize_row_groups. Therefore, our function must accept the labelstr parameter.

The arguments of our analysis function will be as follows:

The analysis function works within a given row/column split context by using the current covariate (cov) and the cached_model function to obtain the desired Cox regression model. From this model, the h_coxreg_extract_interaction function is able to extract information/statistics relevant to the analysis and store it in a data.frame. The rows in this data.frame that are of interest in the current row/column split context are then extracted and the statistic to be printed in the current column is retrieved from these rows. Finally, the formatted cells with this statistic are returned as a VerticalRowsSection object. For more detail see the commented function code below, where the purpose of each line within a_cox_summary is described.

a_cox_summary <- function(df,
                          labelstr = "",
                          .spl_context,
                          stat,
                          format,
                          cache_env,
                          cov_main = FALSE) {
  ## Get current covariate (variable used in latest row split)
  cov <- tail(.spl_context$value, 1)

  ## If currently analyzing treatment effect (ARM) replace empty
  ## value of cov with "ARM" so the correct model row is analyzed
  if (length(cov) == 0) cov <- "ARM"

  ## Use cached_model to get the fitted Cox regression
  ## model for the current covariate
  model <- cached_model(df = df, cov = cov, cache_env = cache_env)

  ## Extract levels of cov to be used as row labels for interaction effects.
  ## If cov is numeric, the median value of cov is used as a row label instead
  cov_lvls <- if (is.factor(df[[cov]])) levels(df[[cov]]) else as.character(median(df[[cov]]))

  ## Use function to calculate and extract information relevant to cov from the model
  cov_rows <- h_coxreg_extract_interaction(effect = "ARM", covar = cov, mod = model, data = df)
  ## Effect p-value is only printed for treatment effect row
  if (!cov == "ARM") cov_rows[, "pval"] <- NA_real_
  ## Extract rows containing statistics for cov from model information
  if (!cov_main) {
    ## Extract rows for main effect
    cov_rows <- cov_rows[cov_rows$level %in% cov_lvls, ]
  } else {
    ## Extract all non-main effect rows
    cov_rows <- cov_rows[nchar(cov_rows$level) == 0, ]
  }
  ## Extract value(s) of statistic for current column and variable/levels
  stat_vals <- as.list(apply(cov_rows[stat], 1, function(x) x, simplify = FALSE))
  ## Assign labels: covariate name for main effect (content) rows, ARM comparison description
  ## for treatment effect (content) row, cov_lvls for interaction effect (data) rows
  nms <- if (cov_main) labelstr else if (cov == "ARM") cov_rows$term_label else cov_lvls
  ## Return formatted/labelled row
  in_rows(
    .list = stat_vals,
    .names = nms,
    .labels = nms,
    .formats = setNames(rep(format, length(nms)), nms),
    .format_na_strs = setNames(rep("", length(nms)), nms)
  )
}

Selecting Parameters

We are able to customize our Cox regression summary using this analysis function by selecting covariates (and their labels), statistics (and their labels), and statistic formats to use when generating the output table. We also initialize a new environment object to be used by the analysis function as the caching environment to store our models in. For the purpose of this example, we will choose all 5 of the possible statistics to include in the table: n, hazard ratio, confidence interval, effect p-value, and interaction p-value.

my_covs <- c("AGE", "RACE") ## Covariates
my_cov_labs <- c("Age", "Race") ## Covariate labels
my_stats <- list("n", "hr", c("lcl", "ucl"), "pval", "pval_inter") ## Statistics
my_stat_labs <- c("n", "Hazard Ratio", "95% CI", "p-value\n(effect)", "p-value\n(interaction)") ## Statistic labels
my_formats <- c(
  n = "xx", hr = "xx.xx", lcl = "(xx.xx, xx.xx)", pval = "xx.xxxx", pval_inter = "xx.xxxx" ## Statistic formats
)
my_env <- new.env()
ny_cache_env <- replicate(length(my_stats), list(my_env)) ## Caching environment

Constructing the Table

Finally, the table layout can be constructed and used to build the desired table.

We first split our basic_table using split_cols_by_multivar to ensure that each statistic exists in its own column. To do so, we choose a variable (in this case STUDYID) which shares the same value in every row, and use it as the split variable for every column so that the full dataset is used to compute the model for every column. We use the extra_args argument for which each list element’s element positions correspond to the children of (columns generated by) this split. These arguments are inherited by all following layout elements operating within this split, which use these elements as argument inputs. To elaborate on this, we have three elements in extra_args: stat, format, and cache_env - each of which are arguments of a_cox_summary and have length equal to the number of columns (as defined above). For each use of our analysis function following this column split, depending on the current column context, the corresponding element of each of these three list elements will be inherited from extra_args and used as input. For example, if analyze_colvars is called with a_cox_summary as afun and is performing calculations for column 1, my_stats[1] ("n") will be given as argument stat, my_formats[1] ("xx") as argument format, and my_cache_env[1] (my_env) as cache_env. This is useful for our table since we want each column to print out values for a different statistic and apply its corresponding format.

Next, we can use summarize_row_groups to generate the content row for treatment effect. This is the first instance where extra_args from the column split will be inherited and used as argument input in cfun.

After generating the treatment effect row, we want to add rows for covariates. We use split_rows_by_multivar to split rows by covariate and apply appropriate labels.

Following this row split, we use summarize_row_groups with a_cox_summary as cfun to generate one content row for each covariate main effect. Once again the contents of extra_args from the column split are inherited as input. Here we specify cov_main = TRUE in the extra_args argument so that main effects rather than interactions are considered. Since this is not a split, this instance of extra_args is not inherited by any following layout elements. As cov_main is a singular value, cov_main = TRUE will be used within every column context.

The last part of our table is the covariate interaction effects. We use analyze_colvars with a_cox_summary as afun, and again inherit extra_args from the column split. Using an rtables “analyze” function generates data rows, with one row corresponding to each covariate level (or median value, for numeric covariates), nested under the content row (main effect) for that same covariate.

lyt <- basic_table() %>%
  ## Column split: one column for each statistic
  split_cols_by_multivar(
    vars = rep("STUDYID", length(my_stats)),
    varlabels = my_stat_labs,
    extra_args = list(
      stat = my_stats,
      format = my_formats,
      cache_env = ny_cache_env
    )
  ) %>%
  ## Create content row for treatment effect
  summarize_row_groups(cfun = a_cox_summary) %>%
  ## Row split: one content row for each covariate
  split_rows_by_multivar(
    vars = my_covs,
    varlabels = my_cov_labs,
    split_label = "Covariate:",
    indent_mod = -1 ## Align split label left
  ) %>%
  ## Create content rows for covariate main effects
  summarize_row_groups(
    cfun = a_cox_summary,
    extra_args = list(cov_main = TRUE)
  ) %>%
  ## Create data rows for covariate interaction effects
  analyze_colvars(afun = a_cox_summary)

Using our pre-processed anl dataset, we can now build and output our final Cox regression summary table.

cox_tbl <- build_table(lyt, anl)
cox_tbl
##                                                                         p-value       p-value   
##                                      n    Hazard Ratio      95% CI      (effect)   (interaction)
## ————————————————————————————————————————————————————————————————————————————————————————————————
## A: Drug X vs control (B: Placebo)   247       0.97       (0.71, 1.32)    0.8243                 
## Covariate:                                                                                      
##   Age                               247                                               0.7832    
##     34                                        0.92       (0.68, 1.26)                           
##   Race                              247                                               0.7441    
##     ASIAN                                     1.03       (0.68, 1.57)                           
##     BLACK OR AFRICAN AMERICAN                 0.78       (0.41, 1.49)                           
##     WHITE                                     1.06       (0.55, 2.04)