[R] coxph strange result

[David Winsemius]

On Nov 25, 2010, at 5:16 PM, Bond, Stephen wrote:

> The following fit does not make sense to me, please, correct me if I  
> have a logical error.
>
>> moddowsn
> Call:
> coxph(formula = Surv(start, stop, resp) ~ sn + matfac2, data = coxsn1,
>    method = "efron")
>
>
>            coef exp(coef) se(coef)       z       p
> sn2       0.0497     1.051  0.02030   2.450 1.4e-02
> sn3      -0.0532     0.948  0.02038  -2.610 9.0e-03
> sn4      -0.0410     0.960  0.01979  -2.073 3.8e-02
> sn5      -0.0776     0.925  0.01954  -3.973 7.1e-05
> sn6      -0.1133     0.893  0.01839  -6.161 7.2e-10
> sn7      -0.1252     0.882  0.01846  -6.781 1.2e-11
> sn8      -0.1222     0.885  0.01994  -6.130 8.8e-10
> sn9      -0.0507     0.951  0.02047  -2.478 1.3e-02
> sn10     -0.0444     0.957  0.02056  -2.159 3.1e-02
> sn11     -0.0433     0.958  0.02157  -2.008 4.5e-02
> sn12     -0.0114     0.989  0.02037  -0.557 5.8e-01
> matfac22 -0.2599     0.771  0.01727 -15.048 0.0e+00
> matfac25 -0.1804     0.835  0.00924 -19.512 0.0e+00
>
> Likelihood ratio test=651  on 13 df, p=0  n= 253802
>
> This would indicate that in sn6 to sn8 there is less of a chance of  
> an event. ?? do the relative frequencies implied by the following  
> table make any sense??
>
>> table(coxsn1$matfac2,coxsn1$sn,coxsn1$resp)
> , ,  = 0
>
>
>        1     2     3     4     5     6     7     8     9    10     
> 11    12
>  1  3407  3177  3425  3348  3975  3564  3181  3077  2894  2610   
> 3441  3443
>  2   920  1005  1142  1327  1645  1530  1330  1184   964   864    
> 888   860
>  5  9036  9507 10258 11888 16826 15575 13394 12346  9938  9001   
> 8970  8599
>
> , ,  = 1
>
>
>        1     2     3     4     5     6     7     8     9    10     
> 11    12
>  1  1453  1459  1186  1496  1295  1754  1429  1153  1106  1234    
> 965  1532
>  2   312   290   330   390   454   539   479   367   295   276    
> 256   267
>  5  2994  3207  3371  3629  4095  5581  5837  3844  3400  3199   
> 2705  3084
>
> Apparently the frequency of an event is higher in the summer months.

Without knowing how the data was prepared, agreement would seem highly  
speculative on our parts. We haven't been told very much about this  
dataset apart from the cryptic variable names apparently referring to  
calendar months regarding some aspect of the cases.


> I apologize for not being able to disclose the dataset, but think  
> that the table provides enough to address the question.

I disagee. Suppose sn2 cases began observation in February and then  
had events in July or September. Or even supposing that observations  
are confined within a single month. If on average, the durations are  
longer within those months than in others, then you could get lower  
rates without having lower even frequencies.

> Thanks everybody.
>
>
> Stephen B

David Winsemius, MD
West Hartford, CT