Celeste Vallejo got us off to a great start to the IDI EEPH seminar this term.
It was a fascinating mix of estimation methods (fitting a continuous-time stochastic process to data collected at fixed, discrete intervals) and application (specifically, to malaria in Nigeria, using the publicly available Garki dataset).
On the estimation side, a key issue is that collecting data at fixed intervals may miss multiple transitions during the interval. For example, an individual who is infected at successive time points might have stayed infectious throughout the time interval. Alternatively, they might have recovered and subsequently became re-infected. Niave methods may lead to biased estimates of transition rates. Celeste presented an approach based upon embedding the observed transition probabilities between states (e.g. infected / susceptible) into a continuous time Markov process. Depending upon the data, this may or may not be feasible (i.e. it may or may not be possible to find rate parameters for the continuous time process that match the observed data), and if it is feasible, the rate parameters may or may not be uniquely determined (i.e. the model may or may not be “identifiable”). Celeste showed sample data where it is impossible for a continuous time process
to precisely match the data, and then resolved this difficulty by adjusting the model to include high and low-immunity segments of the population.
In addition to a learning about the estimation method, Celeste’s talk was also a nice opportunity to learn a bit about malaria. I personally am hoping to learn more about vector-borne disease dynamics through the spring seminar talks.
The Garki data that Celeste applied her methods to look extremely interesting. (And also, as Celeste pointed out, not a study that would easily pass IRB in the present day….) The data looked simultaneously extremely rich, and also difficult to disentangle different effects. I hadn’t heard about this data set before. Residents of several vilages were followed over the course of 3 years, with blood draws taken every 10 weeks to gauge malaria infection status. Several different control measures were applied: no interventions (“control”), indoor residual spraying (IRS), IRS + low-frequency chemotherapy, and IRS + high-frequency chemotherapy. The treatments differed by village, and villages with the same treatments were close to one another geographically, complicating interpretation of the results.
Lots to think about.
Thanks again, Celeste!