Probabilistic Sensitivity Analysis: Generation

Fernando Alarid-Escudero, Greg Knowlton, Eva Enns, and the DARTH Team

2021-05-30

Overview

This vignette explains how to use dampack to generate your own PSA using only a decision analytic model and information about the distributions that define your model parameters of interest. If both costs and effects are calculated within the decision model, the resulting psa object will be compatible with all of dampack’s PSA analysis functions, which are explained at length in the psa_analysis vignette (type vignette("psa_generation", package = "dampack") in the console to view this vignette).

Decision Model Format

In order to generate a PSA in dampack, the user must input the code for the decision analytic model in a standardized format that is compatible with the run_psa function. This is the same format required for the FUN argument of run_owsa_det and run_twsa_det.

The user-defined model function must accept a single input containing a list of the parameters from the params_basecase argument. In the example model shown below, this list is named l_params, and the variables contained in this list are the only variables that are allowed to be varied through the params_range argument in the DSA. Optionally, additional function inputs for FUN can be supplied through the ... argument of run_owsa_det/run_twsa_det, but these inputs are not allowed to vary in the sensitivity analysis. These additional inputs must be arguments of FUN, like n_wtp in the example of calculate_ce_out() below. FUN and its component functions are also able to incorporate variables stored in the global environment, such as n_age_init or n_age_max in the example.

The user-defined model function must return a data.frame where the first column contains a character vector of the strategy names, and the subsequent columns contain numeric vectors of all relevant model outcomes. Each row of the data.frame will consist of a strategy name followed by the corresponding outcome values for that strategy. These model outcomes must be calculated internally within FUN. The model outcomes are not limited to typical outcomes like cost or effectiveness and can be any numerical outcome that the user chooses to model.

library(dampack)
run_sick_sicker_model <- function(l_params, verbose = FALSE) {
  with(as.list(l_params), {
    # l_params must include:
    # -- disease progression parameters (annual): r_HD, p_S1S2, hr_S1D, hr_S2D, 
    # -- initial cohort distribution: v_s_init
    # -- vector of annual state utilities: v_state_utility = c(u_H, u_S1, u_S2, u_D)
    # -- vector of annual state costs: v_state_cost = c(c_H, c_S1, c_S2, c_D)
    # -- time horizon (in annual cycles): n_cyles
    # -- annual discount rate: r_disc
    
    ####### SET INTERNAL PARAMETERS #########################################
    
    # state names
    v_names_states <- c("H", "S1", "S2", "D")
    n_states <- length(v_names_states)
    
    # vector of discount weights
    v_dw  <- 1 / ((1 + r_disc) ^ (0:n_cycles))
    
    # state rewards
    v_state_cost <- c("H" = c_H, "S1" = c_S1, "S2" = c_S2, "D" = c_D)
    v_state_utility <- c("H" = u_H, "S1" = u_S1, "S2" = u_S2, "D" = u_D)
    
    # transition probability values
    r_S1D <- hr_S1D * r_HD   # rate of death in sick state
    r_S2D <- hr_S2D * r_HD   # rate of death in sicker state
    p_S1D <- 1 - exp(-r_S1D) # probability of dying when sick
    p_S2D <- 1 - exp(-r_S2D) # probability of dying when sicker
    p_HD  <- 1 - exp(-r_HD)   # probability of dying when healthy
    
    ## Initialize transition probability matrix 
    # all transitions to a non-death state are assumed to be conditional on survival 
    m_P <- matrix(0, 
                  nrow = n_states, ncol = n_states, 
                  dimnames = list(v_names_states, v_names_states)) # define row and column names
    ## Fill in matrix
    # From H
    m_P["H", "H"]   <- (1 - p_HD) * (1 - p_HS1)    
    m_P["H", "S1"]  <- (1 - p_HD) * p_HS1 
    m_P["H", "D"]   <- p_HD
    # From S1
    m_P["S1", "H"]  <- (1 - p_S1D) * p_S1H
    m_P["S1", "S1"] <- (1 - p_S1D) * (1 - (p_S1H + p_S1S2))
    m_P["S1", "S2"] <- (1 - p_S1D) * p_S1S2
    m_P["S1", "D"]  <- p_S1D
    # From S2
    m_P["S2", "S2"] <- 1 - p_S2D
    m_P["S2", "D"]  <- p_S2D
    # From D
    m_P["D", "D"]   <- 1
    
    # check that all transition matrix entries are between 0 and 1 
    if(!all(m_P <= 1 & m_P >= 0)){
      stop("This is not a valid transition matrix (entries are not between 0 and 1")
    } else
      # check transition matrix rows add up to 1
      if (!all.equal(as.numeric(rowSums(m_P)),rep(1,n_states))){
        stop("This is not a valid transition matrix (rows do not sum to 1)")
      }
    
    ####### INITIALIZATION #########################################
    # create the cohort trace
    m_Trace <- matrix(NA, nrow = n_cycles + 1 , 
                  ncol = n_states,
                  dimnames = list(0:n_cycles, v_names_states)) # create Markov trace
    
    # create vectors of costs and QALYs
    v_C <- v_Q <- numeric(length = n_cycles + 1)
    
    ############# PROCESS ###########################################
    
    m_Trace[1, ] <- v_s_init # initialize Markov trace
    v_C[1] <- 0 # no upfront costs
    v_Q[1] <- 0 # no upfront QALYs
    
    for (t in 1:n_cycles){ # throughout the number of cycles
      m_Trace[t + 1, ] <- m_Trace[t, ] %*% m_P # calculate trace for cycle (t + 1) based on cycle t
      
      v_C[t + 1] <- m_Trace[t + 1, ] %*% v_state_cost
      
      v_Q[t + 1] <- m_Trace[t + 1, ] %*% v_state_utility
      
    }
    
    #############  PRIMARY ECONOMIC OUTPUTS  #########################
    
    # Total discounted costs
    n_tot_cost <- t(v_C) %*% v_dw
    
    # Total discounted QALYs
    n_tot_qaly <- t(v_Q) %*% v_dw
    
    #############  OTHER OUTPUTS   ###################################
    
    # Total discounted life-years (sometimes used instead of QALYs)
    n_tot_ly <- t(m_Trace %*% c(1, 1, 1, 0)) %*% v_dw
    
    ####### RETURN OUTPUT  ###########################################
    out <- list(m_Trace = m_Trace,
                m_P = m_P,
                l_params,
                n_tot_cost = n_tot_cost,
                n_tot_qaly = n_tot_qaly,
                n_tot_ly = n_tot_ly)
    
    return(out)
  }
  )
}
simulate_strategies <- function(l_params, wtp = 100000){
    # l_params must include:
    # -- *** Model parameters ***
    # -- disease progression parameters (annual): r_HD, p_S1S2, hr_S1D, hr_S2D, 
    # -- initial cohort distribution: v_s_init
    # -- vector of annual state utilities: v_state_utility = c(u_H, u_S1, u_S2, u_D)
    # -- vector of annual state costs: v_state_cost = c(c_H, c_S1, c_S2, c_D)
    # -- time horizon (in annual cycles): n_cyles
    # -- annual discount rate: r_disc
    # -- *** Strategy specific parameters ***
    # -- treartment costs (applied to Sick and Sicker states): c_trtA, c_trtB
    # -- utility with Treatment_A (for Sick state only): u_trtA
    # -- hazard ratio of progression with Treatment_B: hr_S1S1_trtB
  
  with(as.list(l_params), {
    
    ####### SET INTERNAL PARAMETERS #########################################
    # Strategy names
    v_names_strat <- c("No_Treatment", "Treatment_A", "Treatment_B")
    # Number of strategies
    n_strat <- length(v_names_strat)
    
    ## Treatment_A
    # utility impacts
    u_S1_trtA <- u_trtA
    # include treatment costs
    c_S1_trtA <- c_S1 + c_trtA
    c_S2_trtA <- c_S2 + c_trtA
    
    ## Treatment_B
    # progression impacts
    r_S1S2_trtB <- -log(1 - p_S1S2) * hr_S1S2_trtB 
    p_S1S2_trtB <- 1 - exp(-r_S1S2_trtB)
    # include treatment costs
    c_S1_trtB <- c_S1 + c_trtB
    c_S2_trtB <- c_S2 + c_trtB

    
    
    ####### INITIALIZATION #########################################
    # Create cost-effectiveness results data frame
    df_ce <- data.frame(Strategy = v_names_strat,
                        Cost = numeric(n_strat),
                        QALY = numeric(n_strat),
                        LY = numeric(n_strat),
                        stringsAsFactors = FALSE)
    
    
    ######### PROCESS ##############################################
    for (i in 1:n_strat){
      l_params_markov <- list(n_cycles = n_cycles, r_disc = r_disc, v_s_init = v_s_init,
                              c_H =  c_H, c_S1 = c_S2, c_S2 = c_S1, c_D = c_D,
                              u_H =  u_H, u_S1 = u_S2, u_S2 = u_S1, u_D = u_D,
                              r_HD = r_HD, hr_S1D = hr_S1D, hr_S2D = hr_S2D,
                              p_HS1 = p_HS1, p_S1H = p_S1H, p_S1S2 = p_S1S2)
      
      if (v_names_strat[i] == "Treatment_A"){
        l_params_markov$u_S1 <- u_S1_trtA
        l_params_markov$c_S1 <- c_S1_trtA
        l_params_markov$c_S2 <- c_S2_trtA
        
      } else if(v_names_strat[i] == "Treatment_B"){
        l_params_markov$p_S1S2 <- p_S1S2_trtB
        l_params_markov$c_S1   <- c_S1_trtB
        l_params_markov$c_S2   <- c_S2_trtB
      }

      l_result <- run_sick_sicker_model(l_params_markov)
      
      df_ce[i, c("Cost", "QALY", "LY")] <- c(l_result$n_tot_cost,
                                           l_result$n_tot_qaly,
                                           l_result$n_tot_ly)
      df_ce[i, "NMB"] <- l_result$n_tot_qaly * wtp - l_result$n_tot_cost
    }
    
    return(df_ce)
  })
}

Generating Parameter Samples for a PSA

The gen_psa_samp function creates a data.frame of parameter value samples based on the underlying distributions specified by the user. Each row of the returned data.frame is an independently sampled set of the parameters varied in the PSA. To produce a psa object, the run_psa function will take each row of this data.frame and calculate the outcomes for each strategy in the user-defined model. The data.frame returned by gen_psa_samp matches the format required by the parameters argument of the make_psa_obj function.

gen_psa_samp has five arguments: params is a vector containing the names of each parameter to be varied in the PSA; dists is a vector of the same length indicating which type of distribution this parameter will be drawn from; parameterization_types is a vector indicating the format of how these parameter distributions are defined; dists_params is a list of vectors, where each element of the list contains the values necessary to define the distribution for a parameter based upon its corresponding element of dists and parameterization_types; and finally, nsamp is a numeric value indicating the number of PSA samples to be generated.

Details about the allowable distributions, their parameterization types and the corresponding formats for dists_params can be found in the help documentation by typing ?gen_psa_samp in the console. Within the example below, the first parameter in the PSA, "p_HS1", follows a "beta" distribution, which has an "a, b" parameterization type (which stands for alpha, beta), and the two values for alpha and beta are 30 and 170, respectively. If the user does not possess estimates for the alpha and beta parameters for the beta distribution but does have estimates for the mean and standard deviation of "p_HS1", they also could choose to parameterize this distribution using parameterization_types = "mean, sd". In this case, the first element of the dists_params list would need to be a numeric vector of length 2 containing the estimated mean and standard deviation for "p_HS1". dampack would then use a method-of-moments estimator to calculate an alpha and beta parameter for this distribution from which the PSA sample values are drawn.

my_params <- c(#Transition probabilities
               "p_HS1", 
               "p_S1H", 
               "p_S1S2", 
               #Hazard ratios
               "hr_S1", 
               "hr_S2",
               "hr_S1S2_trtB",
               #Costs
               "c_H", 
               "c_S1", 
               "c_S2",
               "c_trtA",
               "c_trtB",
               #Utilities
               "u_H",
               "u_S1",
               "u_S2", 
               "u_TrtA")

my_dists <- c(#Transition probabilities
              "beta", 
              "beta", 
              "beta",
              #Hazard ratios
              "log-normal",
              "log-normal", 
              "log-normal",
              #Costs
              "gamma", 
              "gamma", 
              "gamma",
              "gamma", 
              "gamma", 
              #Utilities
              "truncated-normal", 
              "truncated-normal", 
              "truncated-normal", 
              "truncated-normal")

my_parameterization_types <- c(#Transition Probabilities
                               "a, b", 
                               "a, b", 
                               "a, b",
                               #Hazard ratios
                               "mean, sd", 
                               "mean, sd", 
                               "mean, sd",
                               #Costs
                               "shape, scale", 
                               "shape, scale", 
                               "shape, scale",
                               "shape, scale",
                               "shape, scale",
                               #Utilities
                               "mean, sd, ll, ul", 
                               "mean, sd, ll, ul", 
                               "mean, sd, ll, ul",
                               "mean, sd, ll, ul")

my_dists_params <- list(#Transition Probabilities
                        c(7.5, 42.5), 
                        c(12, 12), 
                        c(15, 133),
                        #Hazard ratios
                        c(3, 0.5), 
                        c(10, 0.5),
                        c(0.6, .01),
                        #Costs
                        c(44.5, 45), 
                        c(178, 22.5),
                        c(900, 16.67), 
                        c(576, 21),
                        c(676, 19),
                        #Utilities
                        c(1, 0.01, NA, 1),
                        c(0.75, 0.02, NA, 1), 
                        c(0.5, 0.03, NA, 1),
                        c(0.95, 0.02, NA, 1))
                       
my_psa_params <- gen_psa_samp(params = my_params,
                              dists = my_dists,
                              parameterization_types = my_parameterization_types,
                              dists_params = my_dists_params,
                              n = 100)

Generating Outcomes for the PSA

The run_psa function is used to calculate outcomes for each strategy for every PSA sample through the user-defined decision model, FUN. In this example, the data.frame of PSA parameters generated by gen_psa_samp should be used as the input for the psa_samp argument. The combination of the parameters in the psa_samp and the params_basecase argument must define every parameter that FUN expects within its l_params input argument. Other parameters for FUN that are not contained within l_params list, like the n_wtp argument of calculate_ce_out can be passed through ... as an additional argument in run_psa. If the decision model in FUN is computationally slow and/or the number of PSA samples is extremely large, run_psa could take a long time to run. Under these circumstances, it is recommended that you set the progress argument to TRUE in order to print a progress bar in the console while the function is running to monitor its progress.

my_params_basecase <- list(p_HS1 = 0.15,
                           p_S1H = 0.5,
                           p_S1S2 = 0.105,
                           r_HD = 0.002, 
                           hr_S1D = 3, 
                           hr_S2D = 10,
                           hr_S1S2_trtB = 0.6,
                           c_H = 2000,
                           c_S1 = 4000,
                           c_S2 = 15000, 
                           c_D = 0,
                           c_trtA = 12000,
                           c_trtB = 13000,
                           u_H = 1,
                           u_S1 = 0.75,
                           u_S2 = 0.5,
                           u_D = 0,
                           u_trtA = 0.95,
                           n_cycles = 75,
                           v_s_init = c(1, 0, 0, 0),
                           r_disc = 0.03)

psa_output <- run_psa(psa_samp = my_psa_params,
                      params_basecase = my_params_basecase,
                      FUN = simulate_strategies,
                      outcomes = c("Cost", "QALY", "LY", "NMB"),
                      strategies = c("No_Treatment", "Treatment_A", "Treatment_B"),
                      progress = FALSE)

Creating a Fully-functional PSA Object

run_psa will return a named list containing a psa object for each outcome specified in the outcomes argument. Each psa object in the list is compatible with owsa(), twsa(), and their associated downstream functions described in the psa_analysis vignette. However, most PSA analysis functions in dampack rely on the clear designation of both a cost and effectiveness outcome. To create a PSA object that is compatible with these functions related to cost-effectiveness you must input the results of run_psa into the make_psa_obj function in the following manner. make_psa_obj() requires data.frames for cost, effect, and parameters, and a character vector for strategies. The data.frames containing each outcome in the list returned by run_psa are stored within other_outcome. In this example, the outcome associated with effect is named "Effect", and so psa_output$Effect$other_outcome is supplied to the corresponding argument of make_psa_obj.

cea_psa <- make_psa_obj(cost = psa_output$Cost$other_outcome, 
                        effect = psa_output$QALY$other_outcome, 
                        parameters = psa_output$Cost$parameters,
                        strategies = psa_output$Cost$strategies,
                        currency = "$")