R help - Jul 2019 - Conditional Logistic Regression with Multilevel Data [using clogistic() in Epi and lmer() in lme4???]

Wolfgang,

Perhaps I am missing something here, but with clogit(), or Heather's initial
attempt at using clogistic(), as both have 'strata' arguments,
wouldn't the 'strata' argument need to be able to accept a nested
structure, along the lines of strata(id / site)?

On the page that you link to below, I don't see an example of that type of
nesting structure.

The mixed effects models would generally support a nested random effects
specification, if that makes sense here.

Regards,

Marc Schwartz

> On Jul 3, 2019, at 4:41 AM, Viechtbauer, Wolfgang (SP)
<wolfgang.viechtbauer at maastrichtuniversity.nl> wrote:
> 
> Dear Heather,
> 
> You can do this with clogit() from the 'survival' package, using
strata(). See:
> 
> http://www.metafor-project.org/doku.php/tips:clogit_paired_binary_data
> 
> You can also use lme4::glmer(), glmmML::glmmML(), or metafor::rma.glmm()
for that. [1]
> 
> [1] Rerunning this example, I noticed a slight issue with the structuring
of the output from rma.glmm() -- if you want to run that example, install the
'devel' version of metafor as described here:
> 
> https://wviechtb.github.io/metafor/#installation
> 
> Best,
> Wolfgang
> 
> -----Original Message-----
> From: R-help [mailto:r-help-bounces at r-project.org] On Behalf Of Heather
H Kettrey
> Sent: Tuesday, 02 July, 2019 14:59
> To: r-help at r-project.org
> Subject: [R] Conditional Logistic Regression with Multilevel Data [using
clogistic() in Epi and lmer() in lme4???]
> 
> I need to run some conditional logistic regression models on a multilevel
matched dataset (propensity score matched data from multiple research sites).
> 
> I can pretty easily use the clogistic() function in the Epi package to run
conditional logistic regression models on each separate/nested subset of the
data (e.g., data obtained from each research site separately). However, I cannot
figure out how to run an aggregate analysis (of all aggregated sites) in a way
that accounts for clustering.
> 
> When using unconditional logistic regression I can run the aggregate
analysis lmer() function in the lme4 package. However, I cannot figure out how
to run conditional logistic regression with the lmer() function (I am also open
to any other packages/functions that might work).
> 
> Thanks in advance for help with this.
> 
> Heather Hensman Kettrey, PhD
> Assistant Professor of Sociology
> Clemson University

Ouch, that was a brain fart on my part!

But indeed, since we can fit a 'single-site model' using glmer() (as
shown), we should be able to just add 'site' as another random effect
with (1 | site / id).

Best,
Wolfgang

On July 3, 2019 1:26:54 PM GMT+02:00, Marc Schwartz <marc_schwartz at
me.com> wrote:
>Wolfgang,
>
>Perhaps I am missing something here, but with clogit(), or Heather's
>initial attempt at using clogistic(), as both have 'strata'
arguments,
>wouldn't the 'strata' argument need to be able to accept a
nested
>structure, along the lines of strata(id / site)?
>
>On the page that you link to below, I don't see an example of that type
>of nesting structure.
>
>The mixed effects models would generally support a nested random
>effects specification, if that makes sense here.
>
>Regards,
>
>Marc Schwartz
>
>
>> On Jul 3, 2019, at 4:41 AM, Viechtbauer, Wolfgang (SP)
><wolfgang.viechtbauer at maastrichtuniversity.nl> wrote:
>> 
>> Dear Heather,
>> 
>> You can do this with clogit() from the 'survival' package,
using
>strata(). See:
>> 
>>
>http://www.metafor-project.org/doku.php/tips:clogit_paired_binary_data
>> 
>> You can also use lme4::glmer(), glmmML::glmmML(), or
>metafor::rma.glmm() for that. [1]
>> 
>> [1] Rerunning this example, I noticed a slight issue with the
>structuring of the output from rma.glmm() -- if you want to run that
>example, install the 'devel' version of metafor as described here:
>> 
>> https://wviechtb.github.io/metafor/#installation
>> 
>> Best,
>> Wolfgang

Hi Marc and Wolfgang,


Thanks for your help. Is the following interpretation of your exchange correct?


I need to use the glmer() function in the lme4 package (first installing the
devel version of metaphor). I can specify both site and matched pairing as
random effects using (1 | site / id).


Is this understanding correct?


Thanks,


Heather


Heather Hensman Kettrey, PhD

Assistant Professor of Sociology

Clemson University

________________________________
From: Wolfgang Viechtbauer <wolfgang.viechtbauer at
maastrichtuniversity.nl>
Sent: Wednesday, July 3, 2019 9:15:50 AM
To: Marc Schwartz
Cc: Heather H Kettrey; R-help
Subject: Re: [R] Conditional Logistic Regression with Multilevel Data [using
clogistic() in Epi and lmer() in lme4???]

Ouch, that was a brain fart on my part!

But indeed, since we can fit a 'single-site model' using glmer() (as
shown), we should be able to just add 'site' as another random effect
with (1 | site / id).

Best,
Wolfgang

On July 3, 2019 1:26:54 PM GMT+02:00, Marc Schwartz <marc_schwartz at
me.com> wrote:
>Wolfgang,
>
>Perhaps I am missing something here, but with clogit(), or Heather's
>initial attempt at using clogistic(), as both have 'strata'
arguments,
>wouldn't the 'strata' argument need to be able to accept a
nested
>structure, along the lines of strata(id / site)?
>
>On the page that you link to below, I don't see an example of that type
>of nesting structure.
>
>The mixed effects models would generally support a nested random
>effects specification, if that makes sense here.
>
>Regards,
>
>Marc Schwartz
>
>
>> On Jul 3, 2019, at 4:41 AM, Viechtbauer, Wolfgang (SP)
><wolfgang.viechtbauer at maastrichtuniversity.nl> wrote:
>>
>> Dear Heather,
>>
>> You can do this with clogit() from the 'survival' package,
using
>strata(). See:
>>
>>
>https://urldefense.proofpoint.com/v2/url?u=http-3A__www.metafor-2Dproject.org_doku.php_tips-3Aclogit-5Fpaired-5Fbinary-5Fdata&d=DwIFaQ&c=Ngd-ta5yRYsqeUsEDgxhcqsYYY1Xs5ogLxWPA_2Wlc4&r=BBiFu06xLWCETvNuhD6uw-_crP2zH6EmpC772CQach0&m=PP2LCJ_pE4EwJysp-nsvWt8oIJqJ0nhFvSMBuh7T-ww&s=dvbBssAbENU3KBtLTcWb64QHTComM-RmP8my38CxhW0&e>>
>> You can also use lme4::glmer(), glmmML::glmmML(), or
>metafor::rma.glmm() for that. [1]
>>
>> [1] Rerunning this example, I noticed a slight issue with the
>structuring of the output from rma.glmm() -- if you want to run that
>example, install the 'devel' version of metafor as described here:
>>
>>
https://urldefense.proofpoint.com/v2/url?u=https-3A__wviechtb.github.io_metafor_-23installation&d=DwIFaQ&c=Ngd-ta5yRYsqeUsEDgxhcqsYYY1Xs5ogLxWPA_2Wlc4&r=BBiFu06xLWCETvNuhD6uw-_crP2zH6EmpC772CQach0&m=PP2LCJ_pE4EwJysp-nsvWt8oIJqJ0nhFvSMBuh7T-ww&s=POvfvp_zMlzeLnfGS2EavvZp1-P_0utedAIqz994QkI&e>>
>> Best,
>> Wolfgang
	[[alternative HTML version deleted]]

(Back ON-LIST).

(And combining two of Marc's posts).
> Perhaps I am missing something here, but with clogit(), or Heather's
initial attempt at using clogistic(), as both have 'strata' arguments,
wouldn't the 'strata' argument need to be able to accept a nested
structure, along the lines of strata(id / site)?
> I would defer to the mixed models experts, but I think that the basic
approach using nested random effects seems to make
sense.
> The mixed effects models would generally support a nested random effects
specification, if that makes sense here.

I found the coxme() function from the coxme package.

According to it's documentation:
"Fit a Cox model containing mixed (random and fixed) effects."

I found some sources stating that a Cox model, is equivalent to a
Conditional Logistic Model, in some cases.
And another source that said that you need to set time equal to one, for
every data point.
However, I haven't verified any of this.

I don't know if it's possible to use the coxme() function in conjunction
with Marc's idea for nested random effects...?

	[[alternative HTML version deleted]]