Attributable fraction using a latent class model

By John Aponte and Orvalho Augusto.

Introduction

In malaria endemic areas, asymptomatic carriage of malaria parasites occurs frequently and the detection of malaria parasites in blood films from a febrile individual does not necessarily indicate clinical malaria.

A case definition for symptomatic malaria that is used widely in endemic areas requires the presence of fever or history of fever together with a parasite density above a specific cutoff. If the parasite density is equal or higher than the cutoff point, the fever is considered due to malaria.

How to estimate what is the sensitivity and specificity of the cutoff point in the classification of the fever, without knowing what is the true value of the fevers due to malaria?

Using the attributable fraction, one can estimate the expected number of true cases due to malaria and with the positive predictive value associated with a given cutoff point, we can estimate the expected number of true cases among the fever cases that have a parasite density higher or equal than the selected cutoff point.

In order to estimate the attributable fraction and the positive predictive values, we follow the method proposed by Vounatsou et al (1) fitting a bayesian latent class model.

The latent class model have several advantages over the logistic exponential model: do not make any assumption on the shape of the risk of the fever, only that at higher categories, higher the risk of fever so there is little risk of bias, and provide direct estimation of the attributable fraction and the other quantities like specificity and sensitivity, allowing to estimate directly confidence intervals, however the attributable fraction can be underestimated if the number of categories is too high.

Here it is presented how to do it with the afdx package for the R-software.

Example using synthetic data

The data used in this example (malaria_df1) is a simulated data set as seen frequently in malaria cross-sectionals where two main outcomes are measured, the presence of fever or history of fever (fever column) and the measured parasite density in parasites per \(\mu l\) (density column).

library(afdx)
head(malaria_df1) first 6 observations
fever density
1 475896
1 12008
0 1392
0 1664
0 0
1 0

In this simulation, there are 2000 observations, from which 785 have fever or history of fever and 744 have a density of malaria greater than 0. A total of 437 have both fever and a malaria density higher than 0.

Distribution of fevers by density categories
k (category lower limit) m (no fever) n (fever)
0 908 348
1 12 6
100 8 4
200 19 7
400 37 8
800 47 11
1600 43 28
3200 41 33
6400 42 51
12800 23 59
25600 24 60
51200 10 54
102400 0 50
204800 1 66

The latent class model

* TODO: Include more Details on the model

* TODO: Include Detail on the calculation of sensitivity, specificity and predictive values

Estimating the bayesian latent class model

The adfx provide functions that facilitate the fitting of the bayesian latent class model using the rjags package, but the user is responsible to setup the appropriate bayesian workflow and confirm the convergence of the model. Here we present one way to do it but there are many other possibilities. We use a burn-in of 1000 iterations and monitor samples from 4 chains 10000 iterations.

The function get_latent_model() provides a string with a model that can be use by rjags,

model <- get_latent_model()
cat(model)
#> 
#> data {
#>   # Number of categories
#>   K = length(n)
#>   # Total events by group
#>   Sn <- sum(n[])
#>   Sm <- sum(m[]) 
#> }
#> 
#> model {
#>   for (i in 1:2){   
#>     z0[i] <- (i-1)*0.0001
#>     phi0[i] <- theta[i]*z0[i] 
#>   }
#>   theta[1] <- 1-St
#>   eltheta[1] ~ dgamma(1.0,1.0) 
#>   theta[2] <- eltheta[2]/(1+Sr) 
#>   eltheta[2] ~ dgamma(1.0E-3,1.0E-3) 
#>   for (i in 3:K){ 
#>     phi0[i] <- theta[i]*z0[i] 
#>    eltheta[i] ~ dgamma(1.0E-3,1.0E-3) 
#>     theta[i] <- eltheta[i]/(1+Sr) 
#>     z0[i] <- z0[i-1]/q[i]
#>   }
#>   
#>   Sr <- sum(eltheta[2:K]) 
#>   St <- sum(theta[2:K]) 
#>   Sp <-sum(p0[]) 
#>   Sphi0 <- sum(phi0[]) 
#>   for (i in 1:K){
#>     phi[i] <- phi0[i]/Sphi0 
#>     z[i] <- z0[i]/Sphi0 
#>     q[i] ~ dunif(0.001,0.999)
#>     p0[i] <- theta[i]*(1-lambda)+lambda*phi[i] 
#>     p[i] <- p0[i]/Sp 
#>     lami[i] <- lambda*phi[i]/p0[i]
#>     
#>     # Sens and specificity
#>     sens[i] <- sum(phi[i:K])
#>     P[i] <- (1-lami[i])*p0[i]
#>     spec[i] <- sum(P[1:i])/sum(P)
#>     Q[i] <- lami[i]*p0[i]
#>     ppv[i] <- sum(Q[i:K])/sum(p0[i:K])
#>     npv[i] = spec[i] * (1 - lambda) /(spec[i] * (1 - lambda) + ((1 - sens[i]) * lambda))
#>   }
#> 
#>   # m is afebrile (training sample)
#>   m[1:K] ~ dmulti(theta[1:K], Sm) 
#>   
#>   # n is febrile (mixtured sample)
#>   n[1:K] ~ dmulti(p[1:K], Sn) 
#>   
#>   # lambda is the fraction from the mixture that belong to the g2
#>   lambda~ dunif(0.00001,0.99999)
#> }

Data in the model must be provided as a list with two vectors:

The model calculate as data the number of categories (K) and the total number of subjects in each group (Sn, Sm)

library(rjags)
library(coda)

# compile the model
af_latent <-
  jags.model(
    textConnection(get_latent_model()),
    data = list(n = data$`n (fever)`,
                m = data$`m (no fever)`),
    n.chains = 4,
    n.adapt = 1000,
    inits = list(
      list(.RNG.name = "base::Wichmann-Hill", .RNG.seed = 1111),
      list(.RNG.name = "base::Wichmann-Hill", .RNG.seed = 2222),
      list(.RNG.name = "base::Wichmann-Hill", .RNG.seed = 3333),
      list(.RNG.name = "base::Wichmann-Hill", .RNG.seed = 4444)
    )
  )

# Simulate the posterior
latent_sim <-
  coda.samples(
    model = af_latent,
    variable.names = c('lambda','sens','spec','ppv','npv'),
    n.thinning = 5,
    n.iter =  10000 )

# Extract and Analyze the posterior
latent_sum <-  summary(latent_sim)
latent_eff <-  effectiveSize(latent_sim)
# reformat to present the results
summary_table <-
  data.frame(latent_sum[[1]]) %>%
  bind_cols(data.frame(latent_sum[[2]])) %>%
  mutate(varname = row.names(latent_sum[[1]])) %>%
  mutate(cutoff  = c(NA, rep(cutoffs,4))) %>%
  select(varname, cutoff,Mean, X2.5., X50., X97.5.,Naive.SE ) %>%
  mutate(eff_size = floor(latent_eff)) %>%
  filter(is.na(cutoff) | cutoff != 0) 


mean_table <- summary_table %>%
  rename(point = Mean) %>%
  rename(lci = `X2.5.`) %>%
  rename(uci = `X97.5.`) %>%
  mutate(varname = gsub("\\[.*\\]","",varname)) %>%
  filter(varname != "lambda") %>%
  select(cutoff,varname, lci, uci,point) %>%
  pivot_longer(-c("cutoff","varname"),names_to = "xxv", values_to = "value") %>%
  unite("varx",varname,xxv ) %>%
  pivot_wider(names_from = "varx", values_from = "value") %>%
  select(cutoff,
         sens_point, 
         sens_lci, 
         sens_uci, 
         spec_point,
         spec_lci, 
         spec_uci,
         ppv_point, 
         ppv_lci, 
         ppv_uci,
         npv_point,
         npv_lci,
         npv_uci) %>%
  rename(sensitivity = sens_point) %>%
  rename(specificity = spec_point) %>%
  rename(ppv = ppv_point) %>%
  rename(npv = npv_point) %>%
  mutate_if(is.numeric, round,3)

# Lambda corresponds to the attributable fraction
afrow <- 
  summary_table %>%
  filter(varname == "lambda") %>% 
  mutate_if(is.numeric, round,3)
Summary of diagnostic characteristics at selected cutoff points
cutoff sensitivity sens_lci sens_uci specificity spec_lci spec_uci ppv ppv_lci ppv_uci npv npv_lci npv_uci
1 1.000 1.000 1.000 0.762 0.739 0.784 0.743 0.698 0.784 1.000 1.000 1.000
100 1.000 0.999 1.000 0.769 0.747 0.791 0.751 0.706 0.792 1.000 0.999 1.000
200 1.000 0.999 1.000 0.785 0.762 0.806 0.757 0.712 0.797 1.000 0.999 1.000
400 0.999 0.996 1.000 0.811 0.790 0.832 0.769 0.725 0.809 0.999 0.997 1.000
800 0.998 0.990 1.000 0.845 0.825 0.865 0.792 0.749 0.829 0.998 0.991 1.000
1600 0.993 0.976 1.000 0.883 0.864 0.900 0.821 0.782 0.857 0.994 0.980 1.000
3200 0.968 0.934 0.996 0.917 0.901 0.931 0.855 0.820 0.885 0.975 0.947 0.997
6400 0.916 0.867 0.964 0.951 0.938 0.963 0.887 0.857 0.913 0.940 0.901 0.976
12800 0.808 0.747 0.869 0.972 0.963 0.980 0.922 0.897 0.944 0.875 0.828 0.920
25600 0.666 0.604 0.729 0.990 0.984 0.995 0.945 0.924 0.962 0.804 0.755 0.852
51200 0.502 0.440 0.566 0.998 0.996 1.000 0.973 0.957 0.986 0.735 0.686 0.784
102400 0.348 0.293 0.406 0.999 0.998 1.000 0.994 0.982 0.999 0.680 0.631 0.729
204800 0.201 0.158 0.248 1.000 1.000 1.000 0.995 0.986 1.000 0.634 0.588 0.681

Attributable fraction: 0.419 95%CI(0.373, 0.463 )

Bibliography

  1. Vounatsou P, Smith T, Smith AFM. Bayesian analysis of two-component mixture distributions applied to estimating malaria attributable fractions. Journal of the Royal Statistical Society: Series C (Applied Statistics). 1998;47(4):575–87.

  2. Müller I, Genton B, Rare L, Kiniboro B, Kastens W, Zimmerman P, et al. Three different Plasmodium species show similar patterns of clinical tolerance of malaria infection. Malar J. 2009;8(1):158.

  3. Plucinski MM, Rogier E, Dimbu PR, Fortes F, Halsey ES, Aidoo M, et al. Performance of Antigen Concentration Thresholds for Attributing Fever to Malaria among Outpatients in Angola. J Clin Microbiol [Internet]. 2019 Feb 27 [cited 2020 May 4];57(3). Available from: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6425161/