multivariate analysis from a score

[Gabriel Reis Ferreira]

Dear friends,

I am trying to analyze the effect of this three variables (ml, mms, mic) in the outcome score. For that, I have tried cut the variable score (which is score10) and apply the command logistic. However, the variable mic is already include when I have to genarate the variable score. Therefore, my questions are:

1) Make sense use the variable mic that is already included in the previous model.

2) If I use the log values of the variables and apply regress without cut the variable the effect is different, what is the best in this case.



[CODE]

* Example generated by -dataex-. To install: ssc install dataex
clear
input float(score score10 ml mms mic)
  .3  0      25.6  6122.143 .
  .3  0      19.6  5509.286 0
 3.6  0  28.01327  5147.143 0
  .7  0  22.11574      5025 0
  .7  0      28.2      6150 0
 3.6  0 37.449318  3859.286 0
 6.5  0  28.97162      5925 0
 7.6  0  24.54847  4317.857 0
  .1  0      20.7  6537.857 .
 7.6  0  73.35053  3544.286 0
 1.1  0 29.856247      4635 1
  .1  0  24.62219  6304.286 0
  .3  0      14.2  6812.143 0
 3.6  0      34.1      4920 .
47.4 10      41.4  7088.571 1
28.8 10        39  5631.429 0
 3.6  0  52.04571      4680 0
 1.6  0      10.8  5492.143 .
  .1  0  67.08441  5942.143 0
12.7 10       9.1  6893.571 1
 1.1  0  74.53004      5010 0
 1.1  0      21.2  5723.571 0
  .1  0 11.573903      6465 0
 7.6  0      31.6      5145 0
  .1  0  49.24438  4452.857 0
 1.1  0      26.3 4836.4287 0
 7.6  0  97.23553 4836.4287 .
 6.5  0 127.38666  5792.143 .
 1.6  0 14.596388  5498.571 .
47.4 10  96.27718 4444.2856 0
47.4 10      29.6  6083.571 1

logistic score10 ml mms  mic

Logistic regression                               Number of obs   =         24
                                                  LR chi2(3)      =      12.98
                                                  Prob > chi2     =     0.0047
Log likelihood = -5.7902812                       Pseudo R2       =     0.5285

------------------------------------------------------------------------------
     score10 | Odds Ratio   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
           ml |   1.087406   .0461957     1.97   0.049     1.000531    1.181824
          mmc |   1.001812   .0011957     1.52   0.129     .9994716    1.004158
          mic |   78.14645    165.468     2.06   0.040     1.231878    4957.363
------------------------------------------------------------------------------

. regress score ml mms  mic

      Source |       SS       df       MS              Number of obs =      24
-------------+------------------------------           F(  3,    20) =    5.84
       Model |  2698.21218     3  899.404059           Prob > F      =  0.0049
    Residual |  3077.56648    20  153.878324           R-squared     =  0.4672
-------------+------------------------------           Adj R-squared =  0.3872
       Total |  5775.77865    23  251.120811           Root MSE      =  12.405

------------------------------------------------------------------------------
       score |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
       lml   |    15.0346   5.339857     2.82   0.011     3.895857    26.17335
         mmc |   24.36535   17.51601     1.39   0.179     -12.1724     60.9031
         mic |   21.82136   7.222988     3.02   0.007     6.754472    36.88825
       _cons |  -254.8052   160.3527    -1.59   0.128    -589.2951     79.6846
-----------------------------------------------------------------------------

. regress score ml mms

      Source |       SS       df       MS              Number of obs =      31
-------------+------------------------------           F(  2,    28) =    2.17
       Model |   813.12484     2   406.56242           Prob > F      =  0.1332
    Residual |  5249.63484    28  187.486959           R-squared     =  0.1341
-------------+------------------------------           Adj R-squared =  0.0723
       Total |  6062.75968    30  202.091989           Root MSE      =  13.693

------------------------------------------------------------------------------
       score |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
         ml |   8.534702   4.276687     2.00   0.056    -.2256938     17.2951
        mmc |   23.85516   16.56384     1.44   0.161    -10.07432    57.78464
      _cons |  -226.2016   149.5938    -1.51   0.142    -532.6305    80.22734
------------------------------------------------------------------------------

. logistic score10 ml mms  

Logistic regression                               Number of obs   =         31
                                                  LR chi2(2)      =       3.80
                                                  Prob > chi2     =     0.1494
Log likelihood = -11.794517                       Pseudo R2       =     0.1388

------------------------------------------------------------------------------
     score10 | Odds Ratio   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
          ml |   1.017974   .0178374     1.02   0.309     .9836072    1.053542
         mmc |   1.001284   .0007506     1.71   0.087     .9998135    1.002756
------------------------------------------------------------------------------

Thank you very much

[Carlo Lazzaro]

Gabriel:
some comments on your post:
- if the same variable is part of the regressand and predictor at the same time, the regression model is biased;
- I'm not clear with your statement about cutting the dependent variable. It seems that myou have trasformed score in score10 (10/0), but the underlying criteria of that transformation look obscure (to e, at least);
- moreover, catgorizing (score10) a continuous variabale (score) is a bad idea, as you throw away pieces of information (http://citeseerx.ist.psu.edu/viewdoc...=rep1&type=pdf);
- eventually, what you have in mind, I guess, is not a multivariate regression model (that implies more than one dependent variable and a one or more predictors, like in -mvreg-, but a multiple regression model, that implies a plurality of predictors but one regressand only),

[Gabriel Reis Ferreira]

Dear Carlo Lazzaro, thank you for your kind answer,

The variable score is a model to predict death, once the mortality of this disease is around 10%, I cut the variable to divide individuals who are above the mortality rate.
I want to test if the variable ml and mms could predict score, in other words, if these markers could predict bad outcome for this individual.

best,

Gabriel
(Stata 10.1 SE)

[Carlo Lazzaro]

Gabriel:
I would stick with -score- as a continuous variable and go -regress-:

regress score ml mms

As an amateur of biostatistics, I find strange that sex, age and comorbidities are not mentioned among your predictors.

[Gabriel Reis Ferreira]

Dear Carlo Lazzaro,

You are right, but age, sex, HIV and others confusion factors are included in score. That is my main doubt in this analysis. The distribution of these variables are different between age and sex. I have controlled in previous analysis, now I want to have a overview about the effect of this markers inside all population at first.

My knowledge in biostatistics is limited, specially to multivariate data.

Thank you!

best
Gabriel
(Stata 10.1 SE)