Bootstrap before mi impute in cost-utility analysis?

[Hanna Gyllensten]

Hi,

After all the helpful advice I have been given the last couple of days from members in this forum, I am now hoping that this is my last question "for now":
Do anyone know how to bootstrap a dataset with missing data and to use the result for imputation with, preferably, mi impute? I have a dataset with data from a clinical study where health-related quality of life (EQ-5D questionnaire) has been collected at a number of occations, but many participants have not responded to all questionnaires.

Bootstrap should apparently be implemented before the imputation, according to Schomaker and Heumann. In a previous question I was suggested to try out the programs/codes provided by Glick for bootstrapping and analysis of cost-effectiveness/-utility data. I have also found the code provided by Faria and colleagues (in the supplemental material to an article from 2014), which does multiple imputation first and afterwards bootstrappes to get the results. However, I have not been able to find a method for doing the opposite, bootstrap before mi impute, or to adjust the code from these two sources for my needs. Does anyone have a suggestion on how to do this?

* The main issue at hand is however the application of mi impute to my boostrapped data with missing information? Does anyone know, or have a suggestion on how to get started on getting my bootstrapped results to be useful for imputation, and getting mi impute to bootstrap all of my results from the bootstrap. I have "a feeling" that this is somewhat connected to my difficulties in understanding scalars and matrises, se below.

* One of the problems is that I have difficulties understanding in which scalars, matrises et cetera "things" are put, as the names of these appears to differ between the ones that I have found in the previous codes and compared to the ones I find in my Stata: Maybe there are differences between Stata versions? Does anyone have a suggestion for reading about this that is more accessible for someone with no previous experience of this type of data, than the pdf manual or help-files. For example, in the code by Faria and colleagues they are refering to beta[1,1] and vari[3,3], and other. I cannot find any such in my "list", and I do not know how to read the scalars and matrises well enough to be able to translate to what I find in my list or in the manual and help-file. I'm sorry for the confused question but I don't really know how to explain my problem better.

* A related issue, of course, that the code by Faria and collegues gives a hint on how I could solve it (through mi impute chained), is that many participants have responded to most but not all questionnaires. Thus, I will not be able to calculate the quality adjusted life years beforehand, but will have to first do the bootstrap, thereafter chained imputation of the five responses to the health-related quality of life questionnaire (baseline, 4 weeks, 8 weeks, 18 weeks and 52 weeks: looking at the Faria-code I guess I will have to create a number of new variables indicating the quality adjusted life years during different same-length time intervals of the studied year), and finally to calculate the quality adjusted life years. As I have not managed to get any of my attempted codes to work yet, I cannot say how this will influence other parts of the code.

I am using Stata 14.

I would be happy to get any advise on how to get on with this!

Kind regards,
Hanna

[Carlo Lazzaro]

Hanna:
as per your previous post, I assume that you're dealing with the results (and the covariates) of a cost-effectiveness analysis piggybacked onto a longitudinal clinical study (RCT or non interventistic but anyway empirical) and you're using boostrap to investigate the uncertainty surrounding the base case estimate of the incremental stuff (cost; QALYs; incrmental cost-effectiveness ratio).
It seems that you your dataset is plagued with missing data (it is frequent in this kind of study, especially for EQ-5D-5L domains).
However, you seemingy side-track the mechanism of the missingess you are experiencing: if, say, patiens can't stand the side-effect burden of one of the two treatments and leave the study, you have probably an informative missingness.
That said, I would probably start with dealing with the missing data issue (if the missingness is not informative, you can go -mi-; if it not, things get messier) and then investigating te uncertainty surrounding the incremental stuff via bootstrap.
I do hope that the following references can be useful:
https://www.ncbi.nlm.nih.gov/pubmed/12720255
https://www.crcpress.com/Flexible-Im.../9781439868249

[Hanna Gyllensten]

Thanks for the suggested reading. You are correct in that this is data from such a "piggyback" study, although it is actually not the medical treatment we are investigating but instead the "care" provided to patients, in this case the provision of a person-centred care vs "usual" care. Thus, we do have informative missingness in the form of people being less prone to reply to questionnaires due to ill health during a specific period of the disease (this is patients with head and neck cancer, they often have very low health-related quality of life during a specific period after initiating treatment, much due actually to side-efects of the medical treatment), but this effect should be similar between the intervention groups (although it appears as if the negative symptoms to some extent alleviated by the person-centred care approach).

So, do you think (or anyone else having an opinion/suggestion),it would be ok to do mi impute before the bootstrap instead? According to that paper by Schomaker and Heumann it would/could result in wider confidence intervals, if I remember it correctly.

[Carlo Lazzaro]

Hanna:
Hanna:
as per Karl Claxton's https://www.ncbi.nlm.nih.gov/pubmed/10537899, we know that inference is hardly the main goal of a cost-effectiveness analysis (and, in general, audiences feel more comfortable with a plot of the cost-effectiveness acceptability curve than with the cost-effectiveness plane).
Besides, we know that is the conjoint uncertainty surrounding the base case estimates of incremental cost and QALYs that matters.
Hence, I would not be be concerned about the lenght of the bootstrap 95% confidence intervals (and, by the way, which bootstrap interval do you refer to? [https://www.ncbi.nlm.nih.gov/pubmed/11113956]).
The main issue there is that you have informative missingness: if it persists (that is, it does not translate into missing at random - MAR) even after adjusing for plausible covariates, https://www.crcpress.com/Flexible-Im.../9781439868249 can provide you with some useful hints (including sensitivity analyses).
If you are as lucky as I wish you to be, the MAR situation will allow you to exploit all the capabilities of the -mi- suite.
Anyway, I would deal with the missing data issue first, and only then I would go -bootstrap-.

[Hanna Gyllensten]

Thanks a lot, both for additional suggested reading and advice.