Length of Stay as covariates
Shola Adeyemi
SR Millis
I'm not certain why you could not use LOS as a covariate. As quick search reveals:
Now, there may be some debate regarding what model is chosen when LOS is the response variable, eg, Poisson model or zero-truncate negative binomial model or zero-truncated Poisson or...
Scott
~~~~~~~~~~~ Scott R Millis, PhD, ABPP, CStat, PStat® Professor Wayne State University School of Medicine Email: aa3...@wayne.edu Email: srmi...@yahoo.com Tel: 313-993-8085 --- On Tue, 7/5/11, Shola Adeyemi <shola....@gmail.com> wrote: |
|
Mitchell Maltenfort
For example, what about LOS vs complications. Someone in hospital longer is more likely to have hospital-related complications (or so they told me last year when they chased me out of my room day after surgery) but then if someone does have complications, they'll probably be in hospital longer receiving treatment for the problem.
.
Marc Schwartz
The problem with LOS as a covariate, is that it is typically a surrogate for other process and outcome related measures. Patients with rather complicated inpatient admissions or post-surgical adverse events will tend to have a longer LOS. On the other hand, events such as post operative mortalities typically occur early and thus will tend to have shorter LOS.
More often than not, one is trying to understand the factors that influence LOS, rather than the other way around. In my mind, that would be kind of like the tail wagging the dog. That is why I suspect you don't see too many examples.
One would need to be careful about the question being posed to have some insight into why and when you would use LOS as a covariate. I am having trouble coming up with a scenario at the moment to be honest.
Regards,
Marc Schwartz
On Jul 5, 2011, at 2:35 PM, SR Millis wrote:
> I'm not certain why you could not use LOS as a covariate. As quick search reveals:
>
> http://jama.ama-assn.org/content/278/23/2063.1.short
>
> Now, there may be some debate regarding what model is chosen when LOS is the response variable, eg, Poisson model or zero-truncate negative binomial model or zero-truncated Poisson or...
>
> Scott
>
Shola Adeyemi
Shola ADEYEMI
School of Informatics
Department of Information Systems and Computing
Health and Social Care Modelling Group
University of Westminster
115 New Cavendish Street
London W1W 6UW.
+447984826384
Quote:
Things may come to those who wait, but only the things
left by those who hustle."
-- Abraham Lincoln, Sixteenth president of the United States
Greg Snow
Here is an article that I was involved with:
Trauma patient hospital-associated infections: risks and outcomes
Lazarus, Harrison M ; Fox, Jolene ; Burke, John P ; Lloyd, James F ; Snow, Gregory L ; Mehta, Rajesh R ; Evans, R Scott ; Abouzelof, Rouett ; Taylor, Carrie ; Stevens, Mark H
The Journal of trauma, Jul, 2005, Vol.59(1), p.188-94
We were interested in Hospital Associated Infections, but realized that the infections would increase length of stay, but also that increased length of stay would increase the risk of infection. So we used LoS as both predictor and outcomes variables. It is probably not exactly what you are looking for, but could be relevant.
--
Gregory (Greg) L. Snow Ph.D.
Statistical Data Center
Intermountain Healthcare
From: meds...@googlegroups.com [mailto:meds...@googlegroups.com] On Behalf Of Shola Adeyemi
Sent: Tuesday, July 05, 2011 1:30 PM
To: meds...@googlegroups.com
Subject: {MEDSTATS} Length of Stay as covariates
Hi All,
--
Peter Flom
One problem I can see is collinearity. LOS is likely to be very tied in and collinear with nature of illness.
But even if you were doing a study of a very specific illness – let’s say, cancer of the doohickey – then LOS is likely to be collinear with severity and other measures. It could also interact with severity. This gets tricky! I think a lot of people are not prepared for this sort of methodological trickiness, but it could get very interesting.
Peter Flom
Peter Flom Consulting
Ted Harding
is perhaps some overlap -- is length of prison sentence
imposed per given case of a given crime, and the probability
of re-offending after release.
There is the "observational" type of study, where judges
may vary their sentences according to their impression of
the capacity of the offender to be rehabilitated, etc.,
so there will be a relationship between the outcome and
the pre-sentence estimation of the offender's propensities.
However, this question could feasibly be the subject of
an RCT, and this has certainly been suggested (though I think
not implemented): see Sheila Bird's comments under the
headline "Electronic tagging of offenders" at URL:
http://news.bbc.co.uk/1/hi/programmes/law_in_action/6054816.stm
The topic was whether early release, with electronic tagging,
could reduce re-offending rates.
I know that Sheila has done a lot of work in this area,
but I can not locate detailed references just now.
In such a case, "Length of Stay" would definitely be a
covariate, since it is the "Treatment".
One might inagine a similar situation for patients in hospital
with psychiatric disorder, where the question could be whether
release in order to benefit psychologically from a "healthy
normal life" might have better outcome than detention for
treatment under clinical observation, with the concomitant
"institutionalisation" that might cause prologation and/or
deterioration of the condition (even if major manifestations
can be suppressed pharmacologically).
In either case, of course, one has to bear in mind that
"early release" can be a surrogate for "let's save money"!
Best wishes to allo,
Ted.
--------------------------------------------------------------------
E-Mail: (Ted Harding) <ted.h...@wlandres.net>
Fax-to-email: +44 (0)870 094 0861
Date: 05-Jul-11 Time: 22:22:20
------------------------------ XFMail ------------------------------
Marc Schwartz
These can be insurance coverage, delivery system and socio-economics among others.
For example, take two patients with similar clinical risk profiles and the same admitting diagnosis.
For each patient, their differing insurance coverage (or lack of it) may dictate to some extent, when the patient would be discharged. The success of their physician in arguing with the insurance company for a longer LOS, when required, may also be a factor.
There may be standardized protocols for one or both patients that may result in differences in when discharge occurs. That may be due to differing physicians and/or different hospitals.
If they are at different hospitals, one facility may have a same day admission program for certain elective situations, where another does not and requires that the patient come in at least a day before for testing, etc.
If one hospital is associated with a skilled nursing facility (SNF) or home nursing care services, the patient at that hospital may be discharged sooner, whereas the patient in the other hospital may need to stay longer for PT/OT or other transitional support services.
There may be socio-economic differences, which may be specific to the patient or even prevalent at one hospital due to location (think suburban versus 'inner-city'). Those issues may impact the patient's post-discharge support environment (eg. family, etc.), which may lead to differences in LOS.
Thus, there are both clinical and non-clinical factors that influence LOS, again raising the need for caution and a deeper understanding of what LOS really means, given the circumstances at hand.
Regards,
Marc Schwartz
Steve Simon, P.Mean Consulting
I can only speculate, but here are a couple of reasons why not to use LOS.
First, LOS is not measured at baseline. Most covariates represent
baseline values.
Second, LOS may be an intermediate variable in the causal pathway. A
treatment in the hospital might lead to better health outcomes. The
patients with better health outcomes get out of the hospital quicker. If
you adjust for this, you are adjusting away much of the explanatory
power of your treatment.
Steve Simon, n...@pmean.com, Standard Disclaimer.
Sign up for the Monthly Mean, the newsletter that
dares to call itself average at www.pmean.com/news
MacLennan, Graeme
Dear all my tuppence worth:
I wasn't quite sure of what the OP's research question or study was, but to echo what Steve points out below, in a randomised controlled trial, LOS will most often be a measurement taken after randomisation, and most probably after (or during intervention). In this sense if LOS is not an outcome, but is believed to be on the casual pathway between intervention and outcome then this is an example I reckon of a post-randomisation potential effect modifier. The problem is then establishing is LOS is a mediator or moderator of intervention effect. An excellent overview of these concepts (in the context of complex interventions) can be read in the Emsley reference I have pasted in below.
Regards, Graeme.
Emsley RA, Dunn G, White IR. (2010). Mediation and moderation of treatment effects in randomised controlled trials of complex interventions. Statistical Methods in Medical Research, 19(3), 237-270.
--
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
The University of Aberdeen is a charity registered in Scotland, No SC013683.
Frank Harrell
complications. I think it would be an unusual situation in which it
makes sense to have LOS as a predictor. Another reason for this is
that usually LOS is right-censored at the point of an unsatisfactory
discharge. If you define LOS to be time to successful discharge
(e.g., don't reward a hospital for having quick deaths), you would
need to treat LOS as right censored at the time of an unsuccessful
discharge (e.g., death).
Frank
Marc Schwartz
Might it make sense, in the setting of treating LOS as a time to event analysis, to consider a competing risks approach?
Right censoring presumes that the patient is still at risk for the event, which in this case is discharge. Since a death would preclude a future discharge under normal circumstances, wouldn't censoring at death would lead to a biased estimate?
As I noted in a prior reply, there is also a third possibility relative to the cause of discharge, which is a transfer to a skilled nursing facility rather than to home. Hospitals likely have financial motivations to expedite such transfers, also biasing an assessment of LOS, where this type of transfer is possible.
Thanks and regards,
Marc
BXC (Bendix Carstensen)
You can model the discharge rates (I assume that this is what people refer to when they term LOS as response) as a function of time since admission. And in a competing risk situation you can of course also take a look at the mortality rates as a function of time since admission. And you can even split the discharge into different discharge causes.
Once you have access to the rates for each of these causes you can glue them back together to answer questions like: What is the probability (at any given time) to be "Admitted", Discharged to....", "Dead", for some set of covariates recorded.
In this case LOS seems to be the explanatory variable, namely the main predictor of the rates of interest. If you subscribe to Cox-modelling the technical implementation is to name the LOS the response variable. But that is a technicality.
A note on how to get your hands on the rates and transform them to probabilities in a competing risk setting can be found as:
http://staff.pubhealth.ku.dk/~bxc/AdvCoh/papers/RatesAndRisks.pdf
Best regards,
Bendix Carstensen
_________________________________________
Bendix Carstensen
Senior Statistician
Steno Diabetes Center A/S
Niels Steensens Vej 2-4
DK-2820 Gentofte
Denmark
+45 44 43 87 38 (direct)
+45 30 75 87 38 (mobile)
b...@steno.dk www.biostat.ku.dk/~bxc
www.steno.dk
> -----Original Message-----
> From: meds...@googlegroups.com
> [mailto:meds...@googlegroups.com] On Behalf Of Marc Schwartz
> Sent: 6. juli 2011 15:12
> To: meds...@googlegroups.com
> Subject: Re: {MEDSTATS} Length of Stay as covariates
>
> --
> To post a new thread to MedStats, send email to
> MedS...@googlegroups.com .
Frank Harrell
would preclude future successful discharge, it can be more correct to
right censor than to treat the LOS as complete, i.e., to reward a
hospital for having a high mortality.
Frank
Marc Schwartz
I fully agree that censoring at death is better than considering death as a discharge event for the reasons cited. Coming from a background in cardiac surgery and cardiology, where in-hospital mortality can be a frequent and early event, this is something that has to be taken into account. There are of course, other low risk in-patient samples where this scenario may be a non-issue.
Thanks again,
Marc
On Jul 8, 2011, at 7:39 AM, Frank Harrell wrote:
> Just to comment on the issue of right censoring, even though a death
> would preclude future successful discharge, it can be more correct to
> right censor than to treat the LOS as complete, i.e., to reward a
> hospital for having a high mortality.
> Frank
>
> On Jul 6, 10:06 am, "BXC (Bendix Carstensen)" <b...@steno.dk> wrote:
>> It seems tat it all depends on what you are after.
>>
>> You can model the discharge rates (I assume that this is what people refer to when they term LOS as response) as a function of time since admission. And in a competing risk situation you can of course also take a look at the mortality rates as a function of time since admission. And you can even split the discharge into different discharge causes.
>>
>> Once you have access to the rates for each of these causes you can glue them back together to answer questions like: What is the probability (at any given time) to be "Admitted", Discharged to....", "Dead", for some set of covariates recorded.
>>
>> In this case LOS seems to be the explanatory variable, namely the main predictor of the rates of interest. If you subscribe to Cox-modelling the technical implementation is to name the LOS the response variable. But that is a technicality.
>>
>> A note on how to get your hands on the rates and transform them to probabilities in a competing risk setting can be found as:http://staff.pubhealth.ku.dk/~bxc/AdvCoh/papers/RatesAndRisks.pdf
>>
>> Best regards,
>> Bendix Carstensen
>>
>>