Journal Club – March 2016

Multiple imputation to account for measurement error in marginal structural models.

The first journal club article is: Edwards JK, Cole SR, Westreich D, Crane H, Eron J, Mathews WC, Moore R, Stephen BL, Lesko CR, Mugavero MJ. Multiple imputation to account for measurement error in marginal structural models. Epidemiology. 2015 Sep;26(5):645-52.

This journal club is moderated by Dr. Jess Edwards, a research assistant professor in the Department of Epidemiology at the University of North Carolina at Chapel Hill. Her research interests include quantitative methods for causal inference in the context of infectious diseases (particularly HIV) and occupational epidemiology. Dr. Edwards currently focuses on issues related to measurement error, retention in care, and missing data in clinical cohort studies of patients with HIV.

Dr. Edwards has kindly put together a few questions to kickstart the conversation:

Marginal structural models are useful tools for causal inference from observational data. However, when fitting marginal structural models, exposure is almost always assumed to be measured without error. This paper illustrates how multiple imputation can be used to relax this assumption when a “gold standard” measure of exposure is available for some study participants.


  1. The “gold standard” measure of exposure used in this paper did not come from a formal validation study. What are some of the strengths and limitations of using this type of information on the “gold standard” exposure measure? What are the assumptions necessary to use this information?
  2. What are some examples of other settings where rich exposure information may be available for a subset of study participants? Could this information be used to account for exposure measurement error? Why or why not?
  3. In this paper, multiple imputation was used to impute the “true” exposure prior to estimating the inverse probability weights used to account for confounding. Under which situations would you want to estimate the inverse probability weights before imputing the true exposure?  


  1. Melissa M Adams
    March 7, 2016 @ 1:47 pm

    Thank you for initiating this! The plan for a journal club is fantastic!!!

    Where can I find information about the date and time of journal club? Will the club be conducted with online visuals or will it be conducted only via audio?
    Melissa M Adams, BSN, MPH, PhD
    Cell: 678 571-6653

  2. Stephanie Shiau
    March 8, 2016 @ 8:44 am

    Hi Melissa,
    The journal club will be conducted via this comment box over the next several weeks. If you have questions or comments about the article you would like to raise, please do so here!
    Stephanie Shiau
    PhD Candidate, Columbia University
    SER-SPC Media Chair

  3. Salma Musaad
    April 5, 2016 @ 2:27 pm

    Thank you Dr. Edwards for the this interesting study on the application of multiple imputation to account for measurement error. I particularly appreciated the Appendix that lists the SAS code that the authors used for conducting the multiple imputation.

    1-Using patient report of smoking as the ‘gold standard’ assumes that smoking status was accurately reported with no bias (or endearment error!), that it is stable over time (follow up) and that the patients for whom the self-report smoking information is available were representative of the target population under study.

    2-Another setting where rich exposure information may be available for a subset of study participants includes studies conducted over multiple phases or waves by design, where the first phase collects a large amount of information and the second and subsequent phases collects more specialized or relatively expensive information on a selected subset of the participants. What do you think?

    3-By the confounders I assume you mean the time-fixed covariates age, race, etc. I guess I would want to estimate the inverse probability weights before imputing the true exposure if the validation sample is questionable or does not exist to minimize imputation bias.

    I have a few questions for clarification please:
    1-The authors refer to the treatment weight as time-varying. This was a bit unclear to me because at the beginning of the Methods section they state that therapy naive patients were randomized to one of the four treatment-smoking combination arms. This gave me the impression that the patients used in the study were treated at the same time. Then later on under ‘Therapy Initiation and Smoking Status’ the authors explain that the timing of initiation since study entry varied for each patient. Is this why they refer to the treatment weight as time-varying?

    2-The exposure weight used in calculating the weighted partial likelihood for mortality is the product of the smoking weight and treatment weight. Each weight is composed of the likelihood of the exposure divided by the likelihood given the confounders (age, race, etc.). It is unclear to me how the exposure weigh is calculated in the SAS code. Is it correct to assume that the exposure weight is denoted as sw in the SAS code? sw is calculated as numerator / denominator. Each numerator and denominator is calculated in the SAS code as the multiple of two parts. The second part is the ‘x=0’. Does this second part indicate the treatment weight (assuming x=0 is the likelihood given no smoking)?

  4. Salma Musaad
    April 6, 2016 @ 7:46 am

    Sorry for the typo in response to question 1: I meant measurement error (not endearment error).

Leave a Reply