Some Recent Developments in Incomplete Data Analysis in Biometrical Studies
Presenter: Michael McIsaac
When: Monday, July 11, 2016 Time: 11:00 AM - 12:30 PM
Room: Lecture Theatre (Level 1)
Session Synopsis:
Adaptive sampling for efficient two-phase designs
Response-dependent two-phase designs are used increasingly often in epidemiological studies to ensure sampling strategies offer good statistical efficiency while working within resource constraints. By using a two-phase sampling approach where only a small sub-sample give complete information, it is possible to obtain precise estimates at a greatly reduced cost. Efficiency gains are realized by determining the optimal sub-sample to completely observe. Optimal response-dependent two-phase designs are difficult to implement, however, since they require specification of unknown parameters. We examine adaptive two-phase designs which exploit information from an internal pilot study to approximate the optimal sampling scheme for an analysis based on mean score estimating equations. Application to a motivating biomarker study illustrates how available data can be exploited mid-study to ensure the final sample provides the best value for money. Extensions to efficient tracing studies in cohorts with loss to follow-up will be introduced.
Some Recent Developments in Incomplete Data Analysis in Biometrical Studies
Presenter: James Hanley
When: Monday, July 11, 2016 Time: 11:00 AM - 12:30 PM
Room: Lecture Theatre (Level 1)
Session Synopsis:
Modeling marginal hazard in the presence of unobserved histories: Does interrupting ART increase the risk of liver fibrosis?
Antiretroviral therapy (ART) has reduced morbidity and mortality rates in HIV infected patients, and yet ART is frequently interrupted. When we addressed the question of the impact of ART interruptions on the risk of liver fibrosis using longitudinal data from the Canadian Co-Infection Cohort (CCC), we found a large number of participants had missing treatment and confounder histories, well beyond the scope of a classic "immortal time bias" issue. In this talk, I will briefly describe the information constraints of such studies, and suggest a simple approach to overcoming them, allowing for the inclusion of all participants in a Cox marginal structural model and thereby reducing bias and potentially gaining precision in the effect estimator.
Some Recent Developments in Incomplete Data Analysis in Biometrical Studies
Presenter: Andrea Rotnitzky
When: Monday, July 11, 2016 Time: 11:00 AM - 12:30 PM
Room: Lecture Theatre (Level 1)
Session Synopsis:
Multiple robust fitting of a log-linear model
We consider estimation of a log-linear model for k discrete variables adjusting for high-dimensional covariates. Due to the curse of dimensionality, it is practically impossible to estimate the parameters indexing the model for the dependence of a given interaction on covariates in a way that that is robust to mis-specification of the models for other interactions. In this talk we show that it is possible to partially disentangle the estimation of different interactions. Specifically, for a given interaction ?_{C} we provide an estimator that is consistent and asymptotically normal (CAN) for it under the disjunctive semiparametric model that assumes that for some i?C, p(i | V - {i}) is correctly specified, or equivalently that for some i?C , all the models for interactions ?_{D} such that i?D are correct. Furthermores, our estimator is locally semi-parametric efficient at the intersection of all of the models defining the disjunctive model. We provide a convex estimating function that is guaranteed to have at most one maximum, thus ensuring that fitting algorithms will produce at most one fit (regardless of starting values). Finally, by constructing a hierarchy of estimators we can estimate an entire log-linear model conditional on covariates in a robust manner.
Some Recent Developments in Incomplete Data Analysis in Biometrical Studies
Presenter: Mei-Cheng Wang
When: Monday, July 11, 2016 Time: 11:00 AM - 12:30 PM
Room: Lecture Theatre (Level 1)
Session Synopsis:
Analyzing Recurrent Marker Data by Forward, Backward or Time-adjusted Models in the Presence of a Terminal Event
Recurrent events and marker data arise in many follow-up and surveillance studies where the observation is ended by a terminal event or censoring. In the study the three different types of outcome (recurrent events, marker and time to terminal event) are possibly correlated. We consider modeling and estimation for such data by forward, backward or time-adjusted models: i) Forward recurrent marker process starts at a time origin, 0, and moves forward along time in the conventional way. ii) Backward recurrent marker process considers the terminal event as the time origin and counts time backward. iii) Time-adjusted models characterize the association of recurrent events, markers and time to terminal event by rescaling the time units with proper adjustments. In this talk we will compare the three different types of models, discuss statistical challenges for each model, and present some methods and data applications.
