Cox regression with repeated measures

[David Jaques]

Dear all,

I am rather new to Cox regression models.

I have a database with longitudinal observations on patients who had bone mineral density assessment on multiple occasions and would like to assess the influence of this variable on mortality.

Variable are as follows:
- NECK_BMD: bone mineral density
- Time: Time to follow-up
- Dead: Death status
- MRN: Patient id


An exemple of database is as follows:

* Example generated by -dataex-. To install: ssc install dataex
clear
input long MRN int DATEscan float visit double NECK_BMD
10157 16098 1 .623352083387724
10157 17135 2 .621485280296265
10157 17442 3 .612226300606237
12577 16385 1 .692529753751148
12577 16971 2 .639240592815906
end
format %td DATEscan

I specified the following basic model:

stset time, failure(dead==1) id(MRN)

And then:

stcox NECK_BMD

This gave the following result:

No. of subjects = 593 Number of obs = 749
No. of failures = 405
Time at risk = 1427329
LR chi2(1) = 44.77
Log likelihood = -2340.8138 Prob > chi2 = 0.0000

------------------------------------------------------------------------------
_t | Haz. Ratio Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
NECK_BMD | .1182109 .0377685 -6.68 0.000 .0631972 .2211146



It seems fairly simple. However, I have doubts regarding two things:

1) This model yields very different results compared to using a shared frailty model with the -shared option. There is probably some fundamental differences between these two models but which one would fit my question best?

2) How is number of subjects and observations handled? It seems strange as I would have expected that number of observations used in the model would match number of observation in database without missing values which seems not to be the case based on the following:

tab dead if MRN!=. & NECK_BMD !=. & time!=.

dead | Freq. Percent Cum.
------------+-----------------------------------
0 | 341 33.24 33.24
1 | 685 66.76 100.00
------------+-----------------------------------
Total | 1,026 100.00



OK it is probably a silly and basic question, but still I think I am lost.



Thank you, best regards


David

[David Jaques]

Sorry, results did not come up as expected. Please see below:

stcox NECK_BMD

No. of subjects =          593                  Number of obs    =         749
No. of failures =          405
Time at risk    =      1427329
                                                LR chi2(1)       =       44.77
Log likelihood  =   -2340.8138                  Prob > chi2      =      0.0000

------------------------------------------------------------------------------
          _t | Haz. Ratio   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
    NECK_BMD |   .1182109   .0377685    -6.68   0.000     .0631972    .2211146

and:

tab dead if MRN!=. & NECK_BMD !=. & time!=.

       dead |      Freq.     Percent        Cum.
------------+-----------------------------------
          0 |        341       33.24       33.24
          1 |        685       66.76      100.00
------------+-----------------------------------
      Total |      1,026      100.00

[Paul Dickman]

Thanks for using code tags. Showing how time and dead are coded in your data example would be informative. Also the output from stset.