Multiple imputation nested in bootstrapping

[Lina Massou]

Hi all,

I want to run multiple imputation nested in bootstrapping for a cost-effectiveness analysis and I have seen previous posts that include some lines of STATA code with which we can program the bootstrap. I tried that code but it is extremely slow for 1,000 bootstrap reps. I have seen also that mi impute chained comes with a bootstrap option and I wanted to ask if the programming approach and the bootstrap option are the same.
And if so, then how I can define the number of reps in the bootstrap option?

Many thanks in advance,
L.

[paulvonhippel]

Hi, this article describes the best way to combine the bootstrap with multiple imputation:
https://projecteuclid.org/journals/s...20-STS793.full

It is not implemented in Stata, but implemented in R and published on CRAN:
https://cran.r-project.org/web/packa...ute/index.html

[Carlo Lazzaro]

Lina:
as an aside to Paul's helpful reply, you may want to take a look at the following articles:
1) Faria R, Gomes M, Epstein D, White IR. A guide to handling missing data in cost-effectiveness analysis conducted within randomised controlled trials. Pharmacoeconomics. 2014 Dec;32(12):1157-70. doi: 10.1007/s40273-014-0193-3.
2) Gomes M, Díaz-Ordaz K, Grieve R, Kenward MG. Multiple imputation methods for handling missing data in cost-effectiveness analyses that use data from hierarchical studies: an application to cluster randomized trials. Med Decis Making. 2013 Nov;33(8):1051-63. doi: 10.1177/0272989X13492203.
3) Manca A, Palmer S. Handling missing data in patient-level cost-effectiveness analysis alongside randomised clinical trials. Appl Health Econ Health Policy. 2005;4(2):65-75. doi: 10.2165/00148365-200504020-00001.
4) Ling X, Gabrio A, Mason A, Baio G. A Scoping Review of Item-Level Missing Data in Within-Trial Cost-Effectiveness Analysis. Value Health. 2022 Sep;25(9):1654-1662. doi: 10.1016/j.jval.2022.02.009.
5) Briggs AH, Lozano-Ortega G, Spencer S, Bale G, Spencer MD, Burge PS. Estimating the cost-effectiveness of fluticasone propionate for treating chronic obstructive pulmonary disease in the presence of missing data. Value Health. 2006 Jul-Aug;9(4):227-35. doi: 10.1111/j.1524-4733.2006.00106.x.

[skolenik]

The bootstrap option of mi impute chained only has to do with estimation of the model imputation parameters. The appropriate mi chained imputations should follow the pattern of predict + noise + parameter uncertainty (section 3.1.3 of van Buuren's book https://stefvanbuuren.name/fimd/how-...tations.html); the bootstrap is a variation of the latter component,

The proper process would be to bootstrap, impute once, estimate with imputed data, store the pseudo value, repeat the bootstrap number of times (i.e. hundreds), combine using the bootstrap rules. It will be quite slow, indeed. See DOI : 10.1080/01621459.1996.10476997. I trust the judgement of Statistical Science editors and reviewers, but I would not rush to use the method cited by Paul.

Why do you think you need MI + bootstrap?

[Felix Bittmann]

If you are interested in applying the process skolenik just described, I have written a guide for Stata here: https://www.preprints.org/manuscript/202401.0813/v1