On 30th July, the head of the Statistical Support Unit (SSU) at MLW, Dr. Marc Henrion, presented at a major international statistics conference, the Joint Statistical Meetings (JSM) in Vancouver, Canada.
The talk focused on extending a particular class of models, latent Markov models (LMMs) to be able to estimate sensitivities and specificities of four new molecular diagnostic tests for Salmonella that have been validated by Angeziwa Chirambo and Dr. Tonney Nyirenda from the Salmonella and Enteric Disease group. The reference diagnostic tests for Salmonella, stool culture, was also assessed by the LMMs.
LMMs are commonly used to analyse longitudinal data from multiple diagnostic tests. LMMs consist of a structural model for the latent infection states, defining probabilities for initial state and transmission between states, and a measurement model for the observed test results, defining the item response probabilities and thus test sensitivities and specificities. LMMs typically assume that tests are independent conditional on the latent infection state. This is likely to be violated for tests using similar technologies (as is the case here: four tests using the same PCR technology and two-by-two the molecular tests use the same primers).
Marc presented how the conditional independence assumption can be relaxed by introducing random subject-level. Using simulation data, Marc illustrated the benefits of the mixed LMMs. The talk concluded with an analysis of longitudinal data from four molecular PCR tests and a stool culture test from patients in Blantyre, Malawi. To assess the tests’ performances, both basic and mixed LMMs were fitted. A PCR assay using primers from the TTR gene achieves the best sensitivity / specificity trade-off.