Discussion: View Thread

  • 1.  converting odds ratio to relative risk - multinomal repeated measures

    Posted 06-11-2013 10:01

    Does anyone know of an approach for converting the OR from a multinomal, repeated measures logistic regression to a RR? 

    BACKGROUND

    Performed an analysis to assess the treatment differences in severity (rated from 0 through to 5 each day) for both the primary time period (Week1, Day1-7) as well as across the trial (Days 1-28).  The analysis was performed on each of the four outcomes using SAS GLIMMIX procedure:

    PROC GLIMMIX DATA=dsn;

    WHERE DAY LE 7;  * LE 28 for the 1-28 analysis;

    CLASS PATIENT TRT DAY;

    MODEL SVRITY(ORDER=INTERNAL) = TRT DAY TRT*DAY

    / DIST=MULTINOMIAL LINK=CUMLOGIT;

    RANDOM INT / SUB=PATIENT TYPE=type;

    ESTIMATE 'Treatment 1 vs. 2' TRT 1 -1 / EXP;

    RUN;

     

    The UN (unstructured) covariance structure, as employed in the primary analysis, was used in the model for the assessment of severity for each of the four outcomes.

    The data were primarily "0s" (none) across the trial; there was one outcome that was associated with statistically significantly lower severity - and the client is happy about this.  The statistical output was, as expected, the Odds Ratio (OR) and the associated probability.  The client wants this converted to a Risk Ratio (RR) because it is easier to understand.

    I know any of the conversions are approximate, and the client is also aware of this.  Any help would be appreciated.


    -------------------------------------------
    Janet McDougall
    President
    McDougall Scientific Ltd
    -------------------------------------------


  • 2.  RE:converting odds ratio to relative risk - multinomal repeated measures

    Posted 06-11-2013 10:08
    Produce the Predicted Probability of the outcomes from the model.
    You'll have the "risk" of each outcome for each TRT and DAY combination. ... I'd also strongly suggest that you care about the CIs on these estimates.

    ... Then calculate the RR for the cases you care about.

    -------------------------------------------
    Al Best
    Associate Professor
    Virginia Commonwealth University
    -------------------------------------------








  • 3.  RE:converting odds ratio to relative risk - multinomal repeated measures

    Posted 06-11-2013 10:25
    Hi Al.  The client wants one summary RR - for the 28 day period - basically indicating that treatment 1 is associated with lower severity across the trial. 

    -------------------------------------------
    Janet McDougall
    President
    McDougall Scientific Ltd
    -------------------------------------------








  • 4.  RE:converting odds ratio to relative risk - multinomal repeated measures

    Posted 06-11-2013 11:13
    Hey Janet,
    Clients ask for silly things all the time.

    Assuming random assignment to TRT, the RR at time zero is (almost by definition) nearly =1.
    Your model allows for the RR to be DIFFERENT across time (good!).
    SO: it is silly to ask for a summary averaged across the whole time period.

    The statistical analysis plan probably said that they were interested in differences at specified time points. Say, at 7 days or say, at 28 days.
    So: estimate the risk at the specified day, and the RR can thus be determined at the time point(s) of interest.




  • 5.  RE:converting odds ratio to relative risk - multinomal repeated measures

    Posted 06-11-2013 13:21
    Thanks for your patience Al.  The across time summary measure of risk is what they are looking for (which quite boogles my mind).  There is a trend for both groups to have the severity decrease across time - but it is much more dramatic in treatment 1.  So a point estimate at one day (day 7 or 28) would not capture this.

    -------------------------------------------
    Janet McDougall
    President
    McDougall Scientific Ltd
    -------------------------------------------