[R] propensity scores & imputation

[David Paul]

Hi,

 

Many thanks in advance for whatever advice / input I may receive.

 

I have a propensity score matching / data imputation question.  The purpose
of the propensity

score modeling is to put subjects from two different clinical trials on a
similar footing so that a key

clinical measurement from one study can be attributed / imputed to the other
study.  The goal is

NOT to directly compare the two studies, so this is a very atypical kind of
propensity score usage.

 

I am using lrm( ) to obtain estimated propensity scores, and my question to
this List is rather more 

philosophical than R-syntax.

 

 

Here is the data setup:

 

   a.frame
b.frame

   -----------
------------

   1. Represents  data from clinical trial A                            1.
Represents  data from clinical trial B

  2. Two arms, 'ACTIVE' and 'PLACEBO'                              2. Two
arms, 'ACTIVE' and 'PLACEBO'

   3. The active drug is the same as with Study B              3. The active
drug is the same as with Study A

   4. The trial design is very similar to Study B                    4. The
trial design is very similar to Study A

   5. One measurement is a clinical continuous                 5. Does NOT
have the clinical continuous measure

        measure obtained via laboratory assay                           that
is available in Study A

   6. Number of randomized subjects = 500                       6. Number of
randomized subjects = 5,000

   7. A subset of the baseline covariates (call it                 7. A
subset of the baseline covariates (call it

        a.subset.frame) has 100% commonality
b.subset.frame) has 100% commonality

        with b.subset.frame
with a.subset.frame

 
8. Primary endpoint is time-to-event

 

 

Here is the analysis setup:

 

I have separately split a.frame and b.frame into 'ACTIVE' and 'PLACEBO'
subjects.  

 

For the 'PLACEBO' subjects I have entered the a.subset.frame =
b.subset.frame baseline 

covariates into lrm( ).  The outcome variable is a factor variable
representing Study A = 'Y', 

so the estimated propensity scores are the estimated probabilities that a
'PLACEBO' subject is

from Study A.  I then, finally, used the %GREEDY algorithm (posted on Mayo
Clinic website)

in SAS to match 1-to-many where the Study A subjects are thought of as
'case' subjects and

the Study B subjects are thought of as 'control' subjects. [I know the
matching can be done

in R, I'm working on that now.]  The average number of Study B subjects
matched to a 

single Study A subject is approximately 5.

 

I have done a similar analysis for the 'ACTIVE' subjects.

 

 

 

Here is my question:

 

At the end, I will combine the Study B matched 'PLACEBO' and 'ACTIVE'
subjects and 

perform a Cox PH regression to compare 'PLACEBO' and 'ACTIVE' - there will
be no Study A 

subjects in this analysis.  I want to incorporate the clinical continuous
measurement "borrowed" 

from Study A as a covariate.  When doing this, how should I best take into
account the 

1-to-many matching?  Do I need to weight the Study B subjects, or can I
simply enter the 

matched Study B subjects into a Cox PH regression and ignore the 1-to-many
issue?

 

 

Kind Regards,

 

     David

 

-------------- next part --------------
A non-text attachment was scrubbed...
Name: PGP.sig
Type: application/pgp-signature
Size: 842 bytes
Desc: not available
URL: <https://stat.ethz.ch/pipermail/r-help/attachments/20170316/5e4c6d3e/attachment.bin>