Regression with Numeric Rating Scale (NRS) as a dependant variable

[Sean Stevens]

Good Afternoon,

I am looking to investigate whether there is a relationship between a daily drug dose and a patients response to Survey where items are recorded as 0-10 on a pain scale.

Is a 0-10 pain scale considered to be nominal or ordinal? Should I be running a OLS regression or ologit?
If the ologit is the way to go, how would I interpret the results of a output like the following

. ologit QuestionA Drugdose if Visit==1 & initialpbsv1==1 & DailyMed==1, or

Iteration 0:   log likelihood = -227.39935  
Iteration 1:   log likelihood = -224.81938  
Iteration 2:   log likelihood =  -224.8125  
Iteration 3:   log likelihood =  -224.8125  

Ordered logistic regression                     Number of obs     =        100
                                                LR chi2(1)        =       5.17
                                                Prob > chi2       =     0.0229
Log likelihood =  -224.8125                     Pseudo R2         =     0.0114

------------------------------------------------------------------------------
QuestionA    | Odds Ratio   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
Drugdose     |   1.044084   .0203256     2.22   0.027     1.004997    1.084691
-------------+----------------------------------------------------------------
       /cut1 |   -1.29939   .3389115                     -1.963644   -.6351353
       /cut2 |  -.9457047   .3164983                      -1.56603   -.3253795
       /cut3 |   -.545074   .3005536                     -1.134148    .0440002
       /cut4 |  -.1565631   .2937582                     -.7323186    .4191925
       /cut5 |  -.0672968   .2929305                       -.64143    .5068365
       /cut6 |   .2689752   .2911189                     -.3016073    .8395578
       /cut7 |   .4730803   .2921538                     -.0995307    1.045691
       /cut8 |    .935625   .3022131                      .3432981    1.527952
       /cut9 |    1.61021    .330961                      .9615387    2.258882
      /cut10 |    2.11523   .3640026                      1.401798    2.828662
------------------------------------------------------------------------------

Kind Regards
Sean

[Clyde Schechter]

I believe the developers of the NRS scale intended it to be treated as interval, or even ratio-level data, not just ordinal. Of course, it will be problematic if the collected data suggest a floor or ceiling effect. But assuming that you are dealing with a typical situation of patients with mild to moderate pain, the scores are likely to be nicely distributed, mostly away from the ends of the scale. In that situation, using a continuous-data model like linear regression would be fine.

I suggest you do a histogram of the scores and see what you're dealing with to be sure.

[Sean Stevens]

Hi Clyde,
Thanks for the response. There appears to be a mode around zero for most patients, this impacts the distribution away from normality. Any suggestions?
Regards
Sean

[Clyde Schechter]

Normality is not the issue. But if there is a large cluster of scores at zero and nearby, then you are seeing a floor effect and that would argue for treating the variable as ordinal in this case.

Interpreting the results of an ordinal logistic regression is somewhat complicated to explain in words. I would suggest you use the -margins- command following your regression to get the predicted probabilities of each response level at the different drug doses. And then follow that up with -marginsplot- to visualize it graphically. I think that will be more helpful.