I am working on a large dataset from a cross-sectional survey. About 15% of cases with missing data for both exposure and outcome are present in this dataset. Is it possible to conduct complete case analysis to address this type of missingness instead of doing a multiple imputation? If no, what could be the possible bias for conducting a complete case analysis when deleting cases with missing both exposure and outcome variables? Also, what could be the best method for handling this kind of missingness?
P.S. recommendation for reads is also very much appreciated!
Thank you!
Deborah Lagura Herrera Complete case analysis (listwise deletion, LD) is bias-free only if the missing data mechanism is "missing completely at random" (MCAR) -- a rather strict assumption. Even if the MCAR assumption is reasonable, you would still loose statistical power when using LD compared to multiple imputation (power may or may not matter in your case -- you write that your sample is large). This is because LD leads to omission of cases that provide valid data.
The advantages of multiple imputation or full information maximum likelihood (FIML) estimation with missing data are that these procedures provide bias-free estimates under both MCAR and the weaker (more realistic) assumption of missing at random (MAR) data and that they allow you to retain more statistical power than LD. Modern software programs such as Mplus allow you to conveniently include auxiliary variables (variables related to missingness) in the analysis that can help establish or at least approximate the MAR condition when using multiple imputation or FIML.
Multiple imputation and FIML are equivalent in terms of the underlying statistical model and often lead to similar results. For continuous outcome variables, FIML is perhaps the more straightforward/convenient approach because you do not have to impute values or pool estimates across multiple imputed data sets. Instead, a single set of parameter estimates is provided as if you had complete data. On the other hand, many software programs nowadays also automate multiple imputations and the pooling of the estimates behind the scenes (e.g., Mplus), making the imputation approach also fairly user friendly.
For details, see
Enders, C. K. (2010). Applied missing data analysis. New York: Guilford Press.
Schafer, J. L., & Graham, J. W. (2002). Missing data: Our view of the state of the art. Psychological Methods, 7(2), 147–177. https://doi.org/10.1037/1082-989X.7.2.147
Please have a look at: Van Ginkel et al. (2019). It contains all the references you need.
Journal of Personality Assessment.
Article Rebutting Existing Misconceptions About Multiple Imputation ...
Hi. Some methodologies that you can use are: Hot Deck, Multivariate Imputation by Chained, Equations (MICE), Amelia II, missForest, Hmisc, Mi, Rf2e.
Some references are:
Van Buuren, S. (2018). Flexible Imputation of Missing Data. Second Edition. Chapman & Hall/CRC. Boca Raton, FL.
Why You Probably Need More Imputations Than You Think. 2012, https://statisticalhorizons.com/more-imputations
Bodner, Todd E. (2008) “What improves with increased missing data imputations?” Structural Equation Modeling: A Multidisciplinary Journal 15: 651-675.
Graham, John W., Allison E. Olchowski and Tamika D. Gilreath (2007) “How many imputations are really needed? Some practical clarifications of multiple imputation theory.” Prevention Science 8: 206–213.
Regards
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Sampling Design
- Sample Size