Cox model with time dependent covariate

[Munyaradzi Dimairo]

Dear All

 

I am looking at risk factors associated with mortality in HIV+ve patients (among other things) and have been using Cox model for this. During model diagnostics I noticed that cotrimoxazole is time dependent (using schoenfeld global test and stphplot), and this really makes sense as time when a patient was on treatment really matters. The problem that i have now is...in the dataset i don't have the time when patients were started on treatment in order to use it as a time dependent covariate. This variable was measured during study follow-up and i just know whether the patient was on cotrimoxazole therapy or not. It's really wierd how the data was collected for this variable.

Is there any other way to solve this problem?

 

regards

 

Munya 

--
Munyaradzi Dimairo
Medical Statistician
University of Sheffield

[Bjoern]

On 20 Dez., 15:58, Munyaradzi Dimairo <mdima...@gmail.com> wrote:
> Dear All
>
> I am looking at risk factors associated with mortality in HIV+ve patients
> (among other things) and have been using Cox model for this. During model
> diagnostics I noticed that cotrimoxazole is time dependent (using schoenfeld
> global test and stphplot), and this really makes sense as time when a
> patient was on treatment really matters. The problem that i have now is...in
> the dataset i don't have the time when patients were started on treatment in
> order to use it as a time dependent covariate. This variable was measured
> during study follow-up and i just know whether the patient was on
> cotrimoxazole therapy or not. It's really wierd how the data was collected
> for this variable.
> Is there any other way to solve this problem?

For a start, am I right in assuming we are talking about the
combination anti-bacterial cotrimoxazole (as a quick google search
indicated to me)? If that is so, is there not a risk that
cotrimoxazole use is more of an indirect indicator of disease
progression than a explanatory variable? What I mean is that
presumably it would be given when there are infections, the incidence
of which might be strongly correlated with a weakened immune system
which would in turn presumably increase the risk of death. Or would it
be given prophylactically (but I assume even then not to all HIV
positive patients, but perhaps rather to those at the highest risk of
infections, which would in turn then be an indicator of the
physician's judgement on the state of the patient's immmune system?)?
For the above reasons I would be very, very cautious with this
variable. Even if knew when treatment with the drug had been started,
I am not sure what one would conclude if it were to be showing an
effect in a Cox model, it's a classic epidemiological problem
("confounding by indication").

By "being time dependent" I assume you mean not following the
proportional hazards assumption, i.e. there appears to be a follow-up
time by cotrimoxazole use interaction? One issue with any concomitant
therapy which could cause it to appear to have an effect that does not
follow the proportional hazards assumption even that is the case,
would be if its usage changes over time. E.g. if hypothetically (I do
not know enough about the drug and its usage) at the start of the
observational period it had been used widely for prophylaxis in medium
risk patients and by the end of the observational period medical
practice had changed so that it was prescribed to only in case of
severe acute infections (or the other way around or any other similiar
change in how the drug is prescribed). Another thing could be that if
the drug is used for acute infections then patients are at the highest
risk of death immediately during and after an infection, while the
longer they have made it after the infection the better their
prognosis (in terms of hazard of death). That's something that's very
typical after certain medical events (not sure whether it medically
makes sense here and it all gets complicated and harder to explore
because you do not know when the drug was used). However I am sure
that there are plenty of other possible explanations and with this
kind of thing and with limited data it's probably very hard to
untangle it.

[Pedro Emmanuel Alvarenga Americano do Brasil]

Hello all,

Right now I would speak as a physician. HIV infection and AIDS patients are complex due to the amount of bad stuff that may happen to them. A lot of different cancers, metabolic complications, and mainly the so called opportunist infections. Currently, since the ealy 1990's, cotrimoxazole is used almost exclusively as prophilaxys of pneumocitosys, a pneumonia due to a fungus infection.

A 'normal' (not gaussian) person is expected to have CD4+ cel counts something about 800. CD4+ is a immune system cell key to to both cell and antibobody mediated response and for our 'good luck' is also the target of the HIV virus. As long as disease progress CD4+ cell decreases over years.

 In very very short low CD4+ is very bad prognosis and high CD4+ is good prognosis.

Currently HAARV is indicated when patitens comes close to or below 350 CD+ cell count and contrimozazole is indicates as prophilaxis when CD4+ is close or below 200 cells. Cotrimoxazole should be interrupted after three months the patients has CD4+ over 200. This may take months or (usually) years. Not taking the cotrimoxazole in this period really increases morbidity. But cotrimoxazole is a cheap and easy to use and suprises me you have data of patients not using cotrimoxazole!

Perhaps you may take the dates of the CD4 cell count on the 200 as a proxy of dates of cotrimoxazole prophilatical use. However, clinicaly would make more sense if you take the CD4+ , viral load and the HAART as  time dependent and later check if other oportunistic infections (and its therapies and prophilaxys) are also good in the fit.

Kind regards
-
Abraço forte e que a força esteja com você,

Dr. Pedro Emmanuel A. A. do Brasil
Instituto de Pesquisa Clínica Evandro Chagas
Fundação Oswaldo Cruz
Rio de Janeiro - Brasil

[Munyaradzi Dimairo]

Thanks Bjoern and Pedro for bringing an interesting discussion especially from the medical pespective. I strongly agree that the cotrimoxazole status changes along the follow-up time due to interuption whenever CD4+ cell counts became >200.

 

its really strange the way the data was collected, actually it was collected in a poor resource setting (Zimbabwe) where HIV care was not at its best. As i heard, drug supplies were so erratic so that patients couldn't get the drugs even when prescribed by the physician. So I believe at this stage, with all these stuff in mind there is no "sophisticated" statistical technique that can help and salvage something. The study seemed to be in a messy

with many thanks

 

Munya

--
To post a new thread to MedStats, send email to MedS...@googlegroups.com .
MedStats' home page is http://groups.google.com/group/MedStats .
Rules: http://groups.google.com/group/MedStats/web/medstats-rules

--
Munyaradzi Dimairo
Medical Statistician

Clinical Trials Research Unit
ScHARR (School of Health Related Research)
University of Sheffield
  

[Pedro Emmanuel Alvarenga Americano do Brasil]

Dear Munya,

Once I heard from a teacher saying that "the best way to work with missing values is to avoid them" and "There is no multilevel or MCMC sophisticated technique to fix up bad quality data".

However, working with bad quality data is way better than working wiht no data at all. You may not be able to publish your work at New England Journal of Medicine but once you have your work done, it is best to show around then hidding it, even if it is to say "dont do like this because it does not work!". I guess there is very few mortality data about HIV/AIDS in zimbabwe and Im pretty shure that morbidity is very different in zimbabwe from John Hopikns... so what about mortality?

Show what you have got ... whatever it is! If we hide stuff because it is bad quality, neglected diseases would be not only neglected but also forgotten!

kind regards
--

[Peter Flom]

Pedro Emmanuel Alvarenga Americano do Brasil wrote

<<<
However, working with bad quality data is way better than working wiht no data at all
>>>

Not necessarily.

Bad data leads to bad answers.  You then report it in a journal .... with, of course, a lot of warnings in the limitations section.  Someone else cites your results (without the limitations, of course).  Then a journalist gets hold of the result, simplifies it, and PRESTO!  Major finding.  Based on bad data.

Does this happen?  You bet.

For example, for years, it was widely reported (in the mainstream media and in academic journals) that gay men had more sex partners than straight men.  It turns out that this finding was based on one study - and that study took its sample from gay singles bars.

The original study noted the limitations.  But that got lost.

Peter

Peter L. Flom, PhD
Statistical Consultant
Website: http://www DOT statisticalanalysisconsulting DOT com/
Writing; http://www.associatedcontent.com/user/582880/peter_flom.html
Twitter:   @peterflom

[Pedro Emmanuel Alvarenga Americano do Brasil]

I agree... well remembered!

But what is the solution??? Taking decisions on ... "I believe it will work fine like this"? with no data at all?

I think researches will have a lot of work on tracking what other researchers are doing in the same field... or somehow (not always true) attach results to or get involved in policy making. But very seldom this is a technichal issue, rather a political issue!  Also, I believe that is one of the reasons why evidence based medicine became popular! The reader could have some criticism on what is written (and methods sections are no longer written in smaller fonts).

Well ... nothing to do with statistics!!! Sorry!

 Pedro

2009/12/21 Peter Flom <peterflom...@mindspring.com>

--