Week "9" - Weighting Methods for Time-Varying Treatment Flashcards
What is the difference between modeling the probability of treatment assignment of time varying treatments as compared with time-invariant treatments?
TVTs differ from TITs in that they depend on the previous treatment history, the previous outcomes of treatment, time-invariant covariates, and time-varying covariates.
How does the inverse probability of treatment weight accomplish reduction of selection bias in the estimation of the treatment effect of time-varying treatments?
By weighting each observation at time t with the inverse of the probability of exposure to the conditions the individual was exposed to by time t. given previous treatment history, the previous outcomes of treatment, time-invariant covariates, and time-varying covariates.
What is the difference between inverse probability of treatment weights and stabilized inverse probability of treatment weights?
IPTW is simply the inverse of the product of treatment given a vector of previous treatment indicators, time-varying covariates (including the previous outcomes), and time invariant covariates. Meanwhile, the numerator of SIPTW is the probability that the individual was exposed to the condition at a time given indicators of previous exposure.
Why is it advantageous to use stabilized inverse probability of treatment weights instead of inverse probability of treatment weights?
It addresses the problem that the IPTW formula is likely to result in extreme weights.
What is the difference between stabilized inverse probability of treatment weights and basic stabilized inverse probability of treatment weights?
BSIPTW does not take into account treatment history in the weighting, while SIPTW does.
What issue with stabilized inverse probability of treatment weights does the basic stabilized inverse probability of treatment weights address?
BIPTW addresses the issue that SIPTW does not control for confounding due to treatment history.
What are the advantages and disadvantages of inverse probability of treatment weights over marginal mean weights through stratification?
The disadvantage is IPTW might produce extreme weights.
The advantage is IPTW allows removal of bias due to time-varying and time-invariant covariates, as well as previous history of treatment.
What are two possible outcome models for estimating the effects of time-varying treatments?
- weighted least squares regression
* generalized estimating equations
What is the role of the choice of correlation structure in generalized estimating equations?
The parameter estimates of the GEE are dependent on the correlation structure of two observations for the same individual.
What is the advantage of using generalized estimating equations instead of multilevel models for estimating the effects of time-varying treatments?
GEEs focus on estimating a nonvarying (or average) coefficient in the presence of clustering, whereas MLMs (HLMs) focus on estimating the aspects of the model that vary by group. Less complex models. GEEs are more flexible and less impacted by distributional assumptions.
