We illustrate the impact of different conditional dependence structures on estimates of sensitivity and specificity using data from a study that utilised LCM to estimate the sensitivity and specificity of five different diagnostic tests used in the diagnosis of Melioidosis [16 (link)]. Melioidosis is an infectious disease caused by the bacterium Burkholderia pseudomallei. The data are from a cohort of 320 febrile adult patients recruited over a 6 month period from a hospital in the northeast of Thailand in 2004 [18 (link)]. The five tests included four serological tests (indirect hemagglutination test (IHA), IgM immunochromogenic cassette test (ICT), IgG ICT, and ELISA) and culture test which was assumed 100% specific throughout all their analyses. For comparability we made the same assumption.
In the original analysis, Limmathurotsakul et al. implemented four different LCM with various conditional dependence structures as well as an analysis which assumed culture was a perfect gold standard. The LCM models varied from a model assuming conditional independence between all tests (Model 0) to those considering conditional dependence between a single pair of serological tests using fixed effects (Models 1 and 2) and finally those that use random effects to represent dependence between all serological tests within a disease class (Models 3 and 4) but they did not consider a model that simultaneously accounted for conditional dependence within both true positive and true negative individuals. See Table1 for a summary of the models considered in the original paper. We extend their analysis to consider a ‘Model 5’ which allows dependence between all four serological tests among those individuals truly infected and those individuals truly not infected using random effects. Before reporting the results of this analysis we describe a simulation study used to explore the impact on estimates of sensitivity and specificity of using the wrong conditional dependence structure.
In the original analysis, Limmathurotsakul et al. implemented four different LCM with various conditional dependence structures as well as an analysis which assumed culture was a perfect gold standard. The LCM models varied from a model assuming conditional independence between all tests (Model 0) to those considering conditional dependence between a single pair of serological tests using fixed effects (Models 1 and 2) and finally those that use random effects to represent dependence between all serological tests within a disease class (Models 3 and 4) but they did not consider a model that simultaneously accounted for conditional dependence within both true positive and true negative individuals. See Table
Models and conditional dependence structures compared
| Model | Dependence Structure | Effect Type Used | Included in this paper’s simulation |
|---|---|---|---|
| Conditional Independence between all tests | NA | Yes | |
| Dependence between IHA and IgM ICT in disease positive individuals | Fixed | No | |
| Dependence between IHA and IgG ICT in disease positive individuals | Fixed | No | |
| Dependence between all serological tests in disease positive individuals | Random | Yes | |
| Dependence between all serological tests in disease negative individuals | Random | Yes | |
| Dependence between all serological tests in disease positive and disease negative individuals | Random | Yes |
Models 0–4 considered in Limmathurotsakul et al. [14 (link)]. Model 5 an extension not considered in the previous analyses. The last column highlights the scenarios that are considered in the simulation in this paper
Free full text: Click here
