Hello,

I'm using Mplus for two-level regression analyses with multiply imputed data. 
I'm interested in performing likelihood ratio tests to check if slope variances 
are significantly different from zero (TYPE = TWOLEVEL RANDOM). For 
conventional models, Mplus now reports results of a combined LRT using the 
method described by Meng and Rubin (1992). I wonder whether I can use the same 
formulas for combining the single (say 5) log-likelihoods from my multilevel 
models (or have to use different ones). Any references related to my question 
would be highly appreciated, too.

Thank you!

Jan

-----
Jan Hochweber
German Institute for International Educational
Research (DIPF), Frankfurt/Main, Germany
-------------- next part --------------
An HTML attachment was scrubbed...
URL: 
http://lists.utsouthwestern.edu/pipermail/impute/attachments/20090713/220a6cdb/attachment.htm
From Craig.Enders <@t> asu.edu  Mon Jul 13 11:02:05 2009
From: Craig.Enders <@t> asu.edu (Craig Enders)
Date: Mon Jul 13 11:03:44 2009
Subject: [Impute] likelihood ratio test for multilevel models
References: <b56feb1467768e4db0a6e78c60b283ed02bc18c...@domexc03>
Message-ID: <[email protected]>

Jan,
I have attached a SAS macro program that combines LR tests, as described by 
Meng & Rubin (1992).  You just enter the LR values and the complete-data df and 
the program does the rest.  A couple of things ... The formula for combining LR 
tests appears in several different sources.  If you look at the formulas 
closely, they are not all the same.  There is apparently a typo in one of the 
early articles that has propagated into subsequent manuscripts/books.  I 
emailed Dr. Meng some time ago to get some clarification, so I'm confident that 
the attached SAS program is using the correct formula.  Second, I believe that 
the LR test that Mplus prints uses a different reference distribution than the 
original article.  The webnote on the Mplus website says that the test uses a 
chi-square distribution with the same df as the original test, but the original 
article uses a complex df expression.  Who knows whether this makes any 
practical difference, but its worth mentioning ....
Best,
Craig Enders

*****************************************************************************
Dr. Craig Enders
Associate Professor
Arizona State University
Department of Psychology
Quantitative Psychology Concentration
Box 871104
Tempe, AZ 85287-1104
Office (480) 727-0739
[email protected]
http://www.asu.edu/clas/psych/people/faculty/cenders.htm

*****************************************************************************




-----Original Message-----
From: [email protected] on behalf of Hochweber, Jan
Sent: Mon 7/13/2009 7:01 AM
To: [email protected]
Subject: [Impute] likelihood ratio test for multilevel models
 
Hello,

I'm using Mplus for two-level regression analyses with multiply imputed data. 
I'm interested in performing likelihood ratio tests to check if slope variances 
are significantly different from zero (TYPE = TWOLEVEL RANDOM). For 
conventional models, Mplus now reports results of a combined LRT using the 
method described by Meng and Rubin (1992). I wonder whether I can use the same 
formulas for combining the single (say 5) log-likelihoods from my multilevel 
models (or have to use different ones). Any references related to my question 
would be highly appreciated, too.

Thank you!

Jan

-----
Jan Hochweber
German Institute for International Educational
Research (DIPF), Frankfurt/Main, Germany


-------------- next part --------------
A non-text attachment was scrubbed...
Name: Combining Likelihood Ratio
        Chi-Square Statistics From a MI Analysis.sas
Type: application/octet-stream
Size: 1706 bytes
Desc: Combining Likelihood Ratio Chi-Square Statistics From a
        MI Analysis.sas
Url : 
http://lists.utsouthwestern.edu/pipermail/impute/attachments/20090713/fc70466c/CombiningLikelihoodRatioChi-SquareStatisticsFromaMIAnalysis.obj
From K.J.M.Janssen <@t> umcutrecht.nl  Mon Jul  6 09:23:14 2009
From: K.J.M.Janssen <@t> umcutrecht.nl (Janssen, K.J.M.)
Date: Thu Jul 16 20:23:51 2009
Subject: [Impute] weird question
In-Reply-To: <[email protected]>
References: <[email protected]>
        <[email protected]>
Message-ID: <[email protected]>

Dear all,

I have attachted the manuscript that Frank mentioned, in which several
methods to deal with missing values are compared, once a prediction
model is applied in practice.
Hope this may help you out.

Best regards,

  
Kristel J.M. Janssen PhD
Clinical Epidemiologist
Julius Center for Health Sciences and Primary Care
University Medical Center Utrecht
PO Box 85500
3508 GA Utrecht
The Netherlands
0031-8875-51752

 
 

-----Oorspronkelijk bericht-----
Van: [email protected]
[mailto:[email protected]] Namens Frank E Harrell
Jr
Verzonden: zondag 5 juli 2009 17:46
Aan: Paul Allison
CC: [email protected]
Onderwerp: Re: [Impute] weird question

Paul Allison wrote:
> I completely agree with Frank Harrell that this is not, in general, a 
> good method. I haven't checked out all his references but, for me, the

> definitive refutation was Jones' 1996 paper in the Journal of the 
> American Statistical Association.

And I reference your excellent 2001 book, p. 9-11.
> 
> Nevertheless, I still believe that this method may be useful in two
> situations:
> 
> 1. Data are "missing" because a variable doesn't apply or is undefined

> for some fraction of cases.  For example, suppose you have a measure 
> of marital happiness, dichotomized as high or low, but your sample 
> contains some unmarried people. Then it is entirely appropriate to 
> have a 3-category variable with values high, low, and unmarried.

Nice example Paul.  I've added that to my notes and book, with
attribution.

> 
> 2. The goal is to build a forecasting model, and it is anticipated 
> that a substantial fraction of the new cases to be forecast will have 
> missing data on one or more variables. Here, the goal is not to get 
> unbiased estimates of population parameters but to minimize some 
> function of prediction errors. A workable forecasting model must have 
> some way of dealing with the cases that have missing data. Maybe there

> are better ways, but I've found almost no literature on this topic 
> (with the exception of Warren Sarle's unpublished paper).

My colleagues Janssen, Donders, and Moons in The Netherlands are working
on that.  Averaging predictions over multiple imputations is one
approach but there are others.  There are some logistical problems to
imputing especially with regard to updating the imputation rules.

Cheers,
Frank

> 
> -----------------------------------------------------------------
> Paul D. Allison, Professor
> Department of Sociology
> University of Pennsylvania
> 581 McNeil Building
> 3718 Locust Walk
> Philadelphia, PA  19104-6299
> 215-898-6717
> 215-573-2081 (fax)
> http://www.ssc.upenn.edu/~allison
>  
> 
> -----Original Message-----
> From: [email protected]
> [mailto:[email protected]] On Behalf Of Maria da

> Conceicao-Saraiva
> Sent: Saturday, July 04, 2009 9:19 AM
> To: [email protected]
> Subject: [Impute] weird question
> 
> 
> 
> 
> Sorry about this question,
> 
> I have been discussing with some people I am working about the need of

> imputation with some of our data. What some of analysist are doing is 
> just to creating a category of missing values inside some variables, 
> they argue this is enough. It has been hard to argue with them that 
> this is not the best way to do. Specially in our variable income, we 
> have about 30% of missings.
> Does anybody know about  refereces discussing this approach of just 
> creating a category for missing values inside a variable?
> 
> Maria
> 
> 
> 
> 
> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> ~~~~~~
> ~~
> Maria da Conceicao P. Saraiva DDS, MSc, Ph.D Departamento de Clinica 
> Infantil e Odontologia Social e Preventiva Faculdade de Odontologia de

> Ribeirao Preto-Universidade de Sao Paulo
> 
>   Aviso: Esta mensagem destina-se exclusivamente ao destinatario,
sendo
>   confidencial. Se V. Sa. nao eh o destinatario, fique advertido de 
> que a divulgacao, distribuicao ou copia desta mensagem eh estritamente
proibida.
> Caso tenha recebido esta mensagem por engano, por favor avise 
> imediatamente seu remetente atraves de resposta por e-mail. Obrigado.
>
-- 
Frank E Harrell Jr   Professor and Chair           School of Medicine
                      Department of Biostatistics   Vanderbilt
University

_______________________________________________
Impute mailing list
[email protected]
http://lists.utsouthwestern.edu/mailman/listinfo/impute
-------------- next part --------------
A non-text attachment was scrubbed...
Name: Janssen KJ Dealing with missing predictor values when
        applying clinical prediction models Clin Chem 2009.pdf
Type: application/octet-stream
Size: 95243 bytes
Desc: Janssen KJ Dealing with missing predictor values when
        applying clinical prediction models Clin Chem 2009.pdf
Url : 
http://lists.utsouthwestern.edu/pipermail/impute/attachments/20090706/3d78a56d/JanssenKJDealingwithmissingpredictorvalueswhenapplyingclinicalpredictionmodelsClinChem2009-0001.obj
From Seppo.Laaksonen <@t> stat.fi  Tue Jul 21 03:58:00 2009
From: Seppo.Laaksonen <@t> stat.fi (Laaksonen Seppo)
Date: Tue Jul 21 04:04:21 2009
Subject: Re(2): [Impute] weird question
In-Reply-To: <[email protected]>
References: <[email protected]>
        <[email protected]>
Message-ID: <[email protected]>

Just reading your comments after being on holiday.

I think as Paul that Maria's approach is OK but not best in several cases. I
have even called this solution as a starting imputation method that should be
always done. If possible, several missingess codes are useful to create since
there are often some information behind the missingness (like refusal,
non-contact, does not know, does not want to answer, does not know well but
approximately, missing for unknown reasons). After this kind of coding, it is
of course good to try for getting better 'imputed values.'
In the case of an exploratory analysis with a number of explanatory
(independent) variables, I have not tried to use such an approach for a
dependent variable (Has someone done?), but this approach is not bad if this
happens for explanatory variables. If the mssingness has even been coded
('imputed') well, you can interpret the results quite well (e.g. when I have
been explaining happiness by a number of variables, income being partially
missing with several codes, fortunately, it has not been difficult to see for
example, that some missingess income people are unhappier than the average;
note that this is not necessarily my main explanatory variable but an important
control variable; imputation for these missing categories could be dramatic
since it would be a hard job but not still a main issue, and may lead to a
bias). And as mentioned in my parenthesis example, this kind of a variable can
be a control variable too. Note also that this is very often more advantageous
than the two major alternatives: (i) data deletion when loosing a lot of
observations (too often used in econometrics, for example) and (ii)
unsuccessful imputation (it is also hard to know how well your imputation has
been succeeded and your results can be problematic).

Keep always in mind what are your main estimates. Be realistic and not always
use MI.

Seppo
University of Helsinki

>> Nevertheless, I still believe that this method may be useful in two
>> situations:
>>
>> 1. Data are "missing" because a variable doesn't apply or is undefined
>
>> for some fraction of cases.  For example, suppose you have a measure
>> of marital happiness, dichotomized as high or low, but your sample
>> contains some unmarried people. Then it is entirely appropriate to
>> have a 3-category variable with values high, low, and unmarried.
>
>Nice example Paul.  I've added that to my notes and book, with
>attribution.
>
>>
>> 2. The goal is to build a forecasting model, and it is anticipated
>> that a substantial fraction of the new cases to be forecast will have
>> missing data on one or more variables. Here, the goal is not to get
>> unbiased estimates of population parameters but to minimize some
>> function of prediction errors. A workable forecasting model must have
>> some way of dealing with the cases that have missing data. Maybe there
>
>> are better ways, but I've found almost no literature on this topic
>> (with the exception of Warren Sarle's unpublished paper).
>
>My colleagues Janssen, Donders, and Moons in The Netherlands are working
>on that.  Averaging predictions over multiple imputations is one
>approach but there are others.  There are some logistical problems to
>imputing especially with regard to updating the imputation rules.
>
>Cheers,
>Frank
>
>>
>> -----------------------------------------------------------------
>> Paul D. Allison, Professor
>> Department of Sociology
>> University of Pennsylvania
>> 581 McNeil Building
>> 3718 Locust Walk
>> Philadelphia, PA  19104-6299
>> 215-898-6717
>> 215-573-2081 (fax)
>> http://www.ssc.upenn.edu/~allison
>>
>>
>> -----Original Message-----
>> From: [email protected]
>> [mailto:[email protected]] On Behalf Of Maria da
>
>> Conceicao-Saraiva
>> Sent: Saturday, July 04, 2009 9:19 AM
>> To: [email protected]
>> Subject: [Impute] weird question
>>
>>
>>
>>
>> Sorry about this question,
>>
>> I have been discussing with some people I am working about the need of
>
>> imputation with some of our data. What some of analysist are doing is
>> just to creating a category of missing values inside some variables,
>> they argue this is enough. It has been hard to argue with them that
>> this is not the best way to do. Specially in our variable income, we
>> have about 30% of missings.
>> Does anybody know about  refereces discussing this approach of just
>> creating a category for missing values inside a variable?
>>
>> Maria
>>
>>
>>
>>
>> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>> ~~~~~~
>> ~~
>> Maria da Conceicao P. Saraiva DDS, MSc, Ph.D Departamento de Clinica
>> Infantil e Odontologia Social e Preventiva Faculdade de Odontologia de
>
>> Ribeirao Preto-Universidade de Sao Paulo
>>
>>   Aviso: Esta mensagem destina-se exclusivamente ao destinatario,
>sendo
>>   confidencial. Se V. Sa. nao eh o destinatario, fique advertido de
>> que a divulgacao, distribuicao ou copia desta mensagem eh estritamente
>proibida.
>> Caso tenha recebido esta mensagem por engano, por favor avise
>> imediatamente seu remetente atraves de resposta por e-mail. Obrigado.
>>
>--
>Frank E Harrell Jr   Professor and Chair           School of Medicine
>                      Department of Biostatistics   Vanderbilt
>University
>
>_______________________________________________
>Impute mailing list
>[email protected]
>http://lists.utsouthwestern.edu/mailman/listinfo/impute
>


Reply via email to