Saskia le Cessie (LUMC, Leiden):
DAGS in clinical research: how to explain an apparent direct effect after conditioning on an intermediate variable.

Assessing whether the effect of exposure on an outcome is completely mediated by a third variable is often done by conditioning on the intermediate variable. However, when an association remains, it is not always clear how this should be interpreted. It may be explained by a true direct effect of the exposure on the disease, or the adjustment may have been incomplete due to various reasons, such as error in the measured intermediate or unknown confounding of association between the mediator and the outcome.
We will draw DAGS for different situations where despite absence of a direct effect, still a conditional relation between the exposure and the outcome remains. By quantifying the effect on the association parameter in each of the situations we develop tools for sensitivity analysis of the completeness of adjusting for a mediator. We will apply the tools on data of the LETS study. In this case-control study the question was whether the relation between blood group and venous thrombosis was mediated through coagulation factor VIII.

 

 

We estimate how the effect of antiretroviral treatment depends on the time from HIV infection to initiation of treatment, using observational data. A major challenge in making inferences from such observational data is that treatment is not randomly assigned; e.g., if time of initiation depends on disease status, this dependence will induce bias in the estimation of the effect of interest. Our identifying assumption is that there are no unmeasured confounders. We develop a new class of Structural Nested Mean Models (SNMMs) to estimate the impact of time of initiation of treatment after infection on an outcome measured one year after initiation, compared to the effect of not initiating treatment. We illustrate our methods using the AIEDRP Core01 database on HIV. The current standard of care in HIV-infected patients is Highly Active Anti-Retroviral Treatment (HAART). However, the optimal time to start therapy has not yet been identified, and guidelines about when to start therapy have recently undergone several revisions. This is joint work with Victor DeGruttola (Harvard School of Public Health).

 

 

The risk of death in HIV positive individuals is associated with progression to AIDS, specific AIDS defining conditions (ADCs), use of combination antiretroviral therapy (cART) and calendar time. Data from the CASCADE collaboration was used to further explore this relationship.
Two of many possible approaches are: 1) Estimating the causal effect of both initial ADCs and cART use on mortality, including the interaction, while allowing for effect-modifcation by calendar time, using a marginal structural model (MSM). 2) Estimating the causal effect of ADCs on mortality, allowing for effect-modification by calendar time and cART, using a history-adjusted marginal structural model (HA-MSM),
These approaches are distinctly different. It will be discussed which effects are exactly estimated using the two methods, how these models can be fitted using observational data, and which advantages and disadvantages they possess. We conclude that the use of MSMs is often preferable to the use of HA-MSMs, when conditioning on a completely exogenous time-varying covariate such as calendar time.

Els Goetghebeur (Ghent University):

 

Electronic health records are being gathered on a massive scale. Hard and software manufacturers are merging and promise data mining tools that will help predict and intervene to achieve better outcomes. The hope is well alive, and enormous methodological progress is being made that can help deliver the causal effects. Reliable work is not a matter of quick and dirty however. In observational studies, the challenge of adjusting appropriately for the necessary confounders - especially in the time-varying setting- can be formidable. This adds greatly to the appeal of the alternative instrumental variables approach which mimics randomization based inference. In practice, that seemingly simple approach must meet serious challenges of its own. First, justification of the instrumental variable properties is rarely obvious. Second, there is a plausible or at least workable causal model to propose and third unbiased inference to draw. We develop such an approach when the goal is to model risks and discover interventions that might change them. We thus study the effect of Cox-2 inhibitors on the risk of gastro intestinal bleeding and discuss the evaluation of quality of care in cancer centers. Lessons learned carry through when Mendelian randomization provides the `ideal' instrument.
The talk will refer to joint work with Manoochehr Babanezhad and Stijn Vansteelandt.
 

Aad van der Vaart (Vrije Universiteit Amsterdam) joint work with James Robins, Eric Tchetgen, Ling-ling Li (Harvard University):

 

When responses are missing or when estimating a causal effect from observational data, it may be necessary to include a large number of covariates in the analysis. Lacking information on how exactly these covariates influence the effect of the treatment
or the pattern of missingness in the data, one may opt for semiparametric or nonparametric models. A univariate treatment effect or other numerical parameter may still be estimable at the usual level of statistical precision, provided the nonparametric
parts of the model can be estimated sufficiently well. We shall review classical semiparametric estimation methods for this purpose, based on influence functions for the
parameter of interest. We next discuss the limitation of these methods in situations with really many covariates, and propose a solution using higher-order estimating equations.
We end up with estimators with lower (but optimal) order of statistical precision, which however are guaranteed to work under fewer unverifiable a-priori assumptions.