3
$\begingroup$

Investigators doing studies in large databases (e.g., EMR) in which there is often a lot of missing data usually (in my experience) want to exclude all subjects missing the exposure or outcome of interest, and then use a principled missing data method like multiple imputation to handle missing data on other covariates.

I don't understand this practice, as it seems to result in a cohort that is not defined based on clear scientific grounds as well as in biased and inefficient point estimates [e.g., 1-2].

My question is: Why is it standard practice to treat missing exposure data in a less principled manner than we treat missing covariate data?

References

[1] Demissie, S., LaValley, M. P., Horton, N. J., Glynn, R. J., & Cupples, L. A. (2003). Bias due to missing exposure data using complete‐case analysis in the proportional hazards regression model. Statistics in Medicine, 22(4), 545-557.

[2] Moons, Karel GM, et al. "Using the outcome for imputation of missing predictor values was preferred." Journal of Clinical Epidemiology 59.10 (2006): 1092-1101.

Improve this question
$\endgroup$
3
  • 2
    $\begingroup$ It is not my area of expertise, but I guess it would be worthwhile if you somehow backed up the statement that this is a common practice. $\endgroup$
    – Tim
    Commented Sep 18, 2018 at 14:14
  • 1
    $\begingroup$ I think there was for a time the feeling that imputing the outcome was inappropriate but I do not think that is common these days. $\endgroup$
    – mdewey
    Commented Sep 18, 2018 at 14:20
  • 1
    $\begingroup$ @Tim, I can only refer to my personal experiences collaborating primarily with medical investigators. I'd say I've seen this practice firsthand about 5 times in the past few year. $\endgroup$
    – half-pass
    Commented Sep 18, 2018 at 21:36

1 Answer 1

Reset to default
3
$\begingroup$

This quote suggests that it may be due to fears of "circularity" and a failure to appreciate the stochastic elements of multiple imputation.

At first glance, using an outcome variable to impute an incomplete independent variable may seem incorrect and somewhat circular. However, the addition of a random residual term to each imputed value eliminates any bias that might result from doing so (Little & Rubin, 2002). In fact, multiple imputation programs make no distinction between independent and dependent variables and only require you to specify a set of input variables.

-- Craig Enders, Applied Missing Data Analysis (p. 202)

Improve this answer
$\endgroup$
1
  • 1
    $\begingroup$ Interesting. That seems like a plausible explanation. $\endgroup$
    – half-pass
    Commented Sep 26, 2018 at 21:27

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.