Re: interpretation of the warning message

[Yampolskaya, Svetlana]

Dear List,

I am trying to do Cox regression (bivariate analysis) and I have this
message that I cannot interpret.

Can you please help and tell me what was wrong?



Warnings

Since coefficients did not converge, no further models will be fitted.

Thank you,

Lana

[Hector Maletta]

You probably were doing it stepwise, and the procedure failed to converge at
some step, so it judged it wise to stop at that step and not trying to
introduce further variables. Or perhaps you were doing it in a single step,
and the model simply did not converge. It just happens sometimes, when the
data fail to fit the model.
The reasons may vary. Increasing the number of iterations seldom helps.
Relaxing the convergence criterion is close to cheating and seldom
advisable. Also, it seldom works either (the failure to converge is often by
a much wider margin than any reasonable relaxation you may introduce in the
convergence criterion). Perhaps you have too few cases to arrive at a
significant solution, or perhaps the model is poorly specified. Perhaps the
hazards are not proportional, so Cox does not apply unless some time-related
covariates are introduced. Try to modify the model, e.g. introducing
time-varying covariates, or changing the list of covariates by withdrawing
some of them that seem to have less strong relationship with the event of
interest. Perhaps you may try some simpler models first, to see whether the
covariates fit the data (predict survival) one by one, or in pairs, before
running a more complicated model. Sorry to say I do not have a magic bullet,
but this, as poetry, is more perspiration than inspiration.
Hector

-----Mensaje original-----
De: SPSSX(r) Discussion [mailto:[hidden email]] En nombre de
Yampolskaya, Svetlana
Enviado el: Wednesday, August 23, 2006 11:20 AM
Para: [hidden email]
Asunto: Re: interpretation of the warning message

Dear List,

I am trying to do Cox regression (bivariate analysis) and I have this
message that I cannot interpret.

Can you please help and tell me what was wrong?



Warnings

Since coefficients did not converge, no further models will be fitted.

Thank you,

Lana

[Marta García-Granero]

Hi Hector&Lana

Whenever I got a message like the one Lana reports (in logistic or Cox
regression), it was a problem of small number of cases/events combined
with too many qualitative predictors. This caused an elevated number
of patterns (combination of every level of every predictor) and some
cells had 0 cases/events. I even remember a memorable case where a 3
binary predictors model had been fitted with a total sample size of 36
subjects, with only 10 events (and, yes, it was stepwise - ouch! -,
and I daren't mention the total number of predictors involved in the
process that the authors tested). The statistical analysis was done
with a very old version of SPSS (PC+, before Windows era), that gave
the same warning Lana reports, but also gave a model (with OR of
around 5,000,000 for every predictor - fortunately, new versions of
SPSS just quit fitting the model). Those boars (no typo error here,
I'm insulting them) insisted in publishing the results, against my
opinion (at least, I got my name OUT of that hideous paper, I shudder
when I imagine getting thanked for my statistical assistance in that
horrible contraption).

I would not complicate the model, but simplify it. Run a CROSSTABS of
events against predictors, and check if for some predictor the number
of events in one cell at least is low. Recode your qualitative
predictors to a lower number of categories (if possible), or consider
not including those qualitative predictors with low variability (a
single value has a frequency of 90% or more, while the rest of the
values are poorly represented).

HTH,
Marta

Hector wrote:

HM> You probably were doing it stepwise, and the procedure failed to converge at
HM> some step, so it judged it wise to stop at that step and not trying to
HM> introduce further variables. Or perhaps you were doing it in a single step,
HM> and the model simply did not converge. It just happens sometimes, when the
HM> data fail to fit the model.
HM> The reasons may vary. Increasing the number of iterations seldom helps.
HM> Relaxing the convergence criterion is close to cheating and seldom
HM> advisable. Also, it seldom works either (the failure to converge is often by
HM> a much wider margin than any reasonable relaxation you may introduce in the
HM> convergence criterion). Perhaps you have too few cases to arrive at a
HM> significant solution, or perhaps the model is poorly specified. Perhaps the
HM> hazards are not proportional, so Cox does not apply unless some time-related
HM> covariates are introduced. Try to modify the model, e.g. introducing
HM> time-varying covariates, or changing the list of covariates by withdrawing
HM> some of them that seem to have less strong relationship with the event of
HM> interest. Perhaps you may try some simpler models first, to see whether the
HM> covariates fit the data (predict survival) one by one, or in pairs, before
HM> running a more complicated model. Sorry to say I do not have a magic bullet,
HM> but this, as poetry, is more perspiration than inspiration.

Lana wrote:

HM> I am trying to do Cox regression (bivariate analysis) and I have this
HM> message that I cannot interpret.

HM> Can you please help and tell me what was wrong?

HM> Warnings
HM> Since coefficients did not converge, no further models will be fitted.


---
"It is unwise to use a statistical procedure whose use one does
not understand. SPSS syntax guide cannot supply this knowledge, and it
is certainly no substitute for the basic understanding of statistics
and statistical thinking that is essential for the wise choice of
methods and the correct interpretation of their results".

(Adapted from WinPepi manual - I'm sure Joe Abrahmson will not mind)

[Hector Maletta]

I am puzzled too. That makes two of us, Lana.  Just to speculate about it I
may offer this (speculating is free): the % of the event is about 9%, and
even less than that would have reunification as a reason for discharge, so
perhaps having few cases involved is not really to discard out of hand,
especially if the relationship between the two is not strong. Try a 2x2
table of reunification as a reason (Yes/No) and reentry in foster care
within 12 months (Yes/No) and check the degree of association and the number
of Yes-Yes cases. If the association is weak, and there are relatively few
Yes-Yes cases, we may be into something. I don't bet much on it, but it is a
possibility.
Hector


-----Mensaje original-----
De: Yampolskaya, Svetlana [mailto:[hidden email]]
Enviado el: Wednesday, August 23, 2006 11:45 AM
Para: Hector Maletta
Asunto: RE: interpretation of the warning message

Hector,

Thank you for your reply but I am still puzzled because I have 34,830
cases of children who exited foster care during two years. 3,022
reentered foster care within 12 months. So I am trying to look at time
to reentry. In this model I used only one covariate - reunification as a
reason for discharge. I coded that variable 1 - reunification as a
reason for discharge and 0 - other reasons.  When I do the same analysis
and use a different covariate - placement with relatives as a reason for
discharge - the model runs without any problem. Multivariate model runs
without problems too.

Any advise will be greatly appreciated,

Lana


-----Original Message-----
From: Hector Maletta [mailto:[hidden email]]
Sent: Wednesday, August 23, 2006 10:32 AM
To: Yampolskaya, Svetlana; [hidden email]
Subject: RE: interpretation of the warning message

You probably were doing it stepwise, and the procedure failed to
converge at some step, so it judged it wise to stop at that step and not
trying to introduce further variables. Or perhaps you were doing it in a
single step, and the model simply did not converge. It just happens
sometimes, when the data fail to fit the model.
The reasons may vary. Increasing the number of iterations seldom helps.
Relaxing the convergence criterion is close to cheating and seldom
advisable. Also, it seldom works either (the failure to converge is
often by a much wider margin than any reasonable relaxation you may
introduce in the convergence criterion). Perhaps you have too few cases
to arrive at a significant solution, or perhaps the model is poorly
specified. Perhaps the hazards are not proportional, so Cox does not
apply unless some time-related covariates are introduced. Try to modify
the model, e.g. introducing time-varying covariates, or changing the
list of covariates by withdrawing some of them that seem to have less
strong relationship with the event of interest. Perhaps you may try some
simpler models first, to see whether the covariates fit the data
(predict survival) one by one, or in pairs, before running a more
complicated model. Sorry to say I do not have a magic bullet, but this,
as poetry, is more perspiration than inspiration.
Hector

-----Mensaje original-----
De: SPSSX(r) Discussion [mailto:[hidden email]] En nombre de
Yampolskaya, Svetlana Enviado el: Wednesday, August 23, 2006 11:20 AM
Para: [hidden email]
Asunto: Re: interpretation of the warning message

Dear List,

I am trying to do Cox regression (bivariate analysis) and I have this
message that I cannot interpret.

Can you please help and tell me what was wrong?



Warnings

Since coefficients did not converge, no further models will be fitted.

Thank you,

Lana

[Hector Maletta]

Lana,
Sorry my guess didn't work. I'm afraid that was my one bright idea for
today. Even sorrier not to have another one for the moment...
The Yes-No and No-Yes cells in your table (copied below) have fewer cases
than the Yes-Yes and No-No cells, but they are still in the hundreds, so I
do not think lack of cases is the problem. Marta Garcia-Granero, the
resident Math Stat genius in this list, also suspected zero cells and small
number of cases in some cells, but that's not the case I'm afraid. Neither
are the two variables unrelated, as shown by the highly significant chi
square you found. There are more things in Cox Regression, apparently, than
our philosophy can dream, to paraphrase Hamlet.
Back to square one, for the time being. The non convergence message you
received is yet to be explained.
May perhaps the hazards be non proportional? Perhaps they are so strongly
non proportional that the whole procedure fails to converge? You may do a
Kaplan-Meier of reentry times for children discharged due to reunification
and another for children discharged due to other reasons, and see whether
the reentry curves are parallel, or you may run Cox again with cause of
discharge as a binary covariate as before but also with a time varying
covariate, like time itself to keep it simple (If that fails try with log
time or time squared, just in case). You may also run Cox with just the
time-varying covariate, stratified by the cause of discharge, and inspect
the curves for hazard proportionality.
Hector


-----Mensaje original-----
From: Yampolskaya, Svetlana [mailto:[hidden email]]
Date: Wednesday, August 23, 2006 1:29 PM
To: Hector Maletta
Subject: RE: interpretation of the warning message

Hector,
I did the table. Please see attached. Pearson Chi-Square = 26107.16.
There are 2594 yes-yes cases... Any ideas?

         Crosstabulation
Count
                                REENTRY Total
                                No      Yes
REUNIFICATION   No      31508   428     31936
                        Yes     300     2594    2894
                        Total   31808   3022    34830

Lana

-----Original Message-----
From: Hector Maletta [mailto:[hidden email]]
Sent: Wednesday, August 23, 2006 11:07 AM
To: Yampolskaya, Svetlana
Subject: RE: interpretation of the warning message

I am puzzled too. That makes two of us, Lana.  Just to speculate about
it I may offer this (speculating is free): the % of the event is about
9%, and even less than that would have reunification as a reason for
discharge, so perhaps having few cases involved is not really to discard
out of hand, especially if the relationship between the two is not
strong. Try a 2x2 table of reunification as a reason (Yes/No) and
reentry in foster care within 12 months (Yes/No) and check the degree of
association and the number of Yes-Yes cases. If the association is weak,
and there are relatively few Yes-Yes cases, we may be into something. I
don't bet much on it, but it is a possibility.
Hector


-----Mensaje original-----
De: Yampolskaya, Svetlana [mailto:[hidden email]] Enviado el:
Wednesday, August 23, 2006 11:45 AM
Para: Hector Maletta
Asunto: RE: interpretation of the warning message

Hector,

Thank you for your reply but I am still puzzled because I have 34,830
cases of children who exited foster care during two years. 3,022
reentered foster care within 12 months. So I am trying to look at time
to reentry. In this model I used only one covariate - reunification as a
reason for discharge. I coded that variable 1 - reunification as a
reason for discharge and 0 - other reasons.  When I do the same analysis
and use a different covariate - placement with relatives as a reason for
discharge - the model runs without any problem. Multivariate model runs
without problems too.

Any advise will be greatly appreciated,

Lana


-----Original Message-----
From: Hector Maletta [mailto:[hidden email]]
Sent: Wednesday, August 23, 2006 10:32 AM
To: Yampolskaya, Svetlana; [hidden email]
Subject: RE: interpretation of the warning message

You probably were doing it stepwise, and the procedure failed to
converge at some step, so it judged it wise to stop at that step and not
trying to introduce further variables. Or perhaps you were doing it in a
single step, and the model simply did not converge. It just happens
sometimes, when the data fail to fit the model.
The reasons may vary. Increasing the number of iterations seldom helps.
Relaxing the convergence criterion is close to cheating and seldom
advisable. Also, it seldom works either (the failure to converge is
often by a much wider margin than any reasonable relaxation you may
introduce in the convergence criterion). Perhaps you have too few cases
to arrive at a significant solution, or perhaps the model is poorly
specified. Perhaps the hazards are not proportional, so Cox does not
apply unless some time-related covariates are introduced. Try to modify
the model, e.g. introducing time-varying covariates, or changing the
list of covariates by withdrawing some of them that seem to have less
strong relationship with the event of interest. Perhaps you may try some
simpler models first, to see whether the covariates fit the data
(predict survival) one by one, or in pairs, before running a more
complicated model. Sorry to say I do not have a magic bullet, but this,
as poetry, is more perspiration than inspiration.
Hector

-----Mensaje original-----
De: SPSSX(r) Discussion [mailto:[hidden email]] En nombre de
Yampolskaya, Svetlana Enviado el: Wednesday, August 23, 2006 11:20 AM
Para: [hidden email]
Asunto: Re: interpretation of the warning message

Dear List,

I am trying to do Cox regression (bivariate analysis) and I have this
message that I cannot interpret.

Can you please help and tell me what was wrong?



Warnings

Since coefficients did not converge, no further models will be fitted.

Thank you,

Lana

[Marta García-Granero]

Hi Lana

Can you send to me (and to Hector)?:

1) The Kaplan-Meier graph for the analysis of that particular
variable
2) The "Case Processing Summary" table at the beginning of the output
3) The full iteration history (to see if the model was running wild,
with the coefficients growing almost exponentially or was close to
finding a solution, perhaps at iteration number 25...)

This might help to identify the problem. Right now, I can think of:

- Almost all cases for one group are censored (did you take a look at
the Case Processing Summary table at the beginning of the output? you
might find out that a substantial amount censored cases before the
first event were dropped before any model was fitted)

- There is clearly a pattern of non  proportionality (although this
should render the results non significant, but not "non computable")


Hector wrote:

HM> Back to square one, for the time being. The non convergence message you
HM> received is yet to be explained.

Lana wrote:

HM> Hector,
HM> I did the table. Please see attached. Pearson Chi-Square = 26107.16.
HM> There are 2594 yes-yes cases... Any ideas?

HM>          Crosstabulation
HM> Count
HM>                                 REENTRY Total
HM>                                 No      Yes
HM> REUNIFICATION   No      31508   428     31936
HM>                         Yes     300     2594    2894
HM>                         Total   31808   3022    34830


--
Regards,
Dr. Marta García-Granero,PhD           mailto:[hidden email]
Statistician

---
"It is unwise to use a statistical procedure whose use one does
not understand. SPSS syntax guide cannot supply this knowledge, and it
is certainly no substitute for the basic understanding of statistics
and statistical thinking that is essential for the wise choice of
methods and the correct interpretation of their results".

(Adapted from WinPepi manual - I'm sure Joe Abrahmson will not mind)