fixed vs random
Terms such as (gender|candidate) estimate a gender effect that can vary across candidates. It is presumably meaningless to discuss the gender effect of any individual candidate, so this term should not be used. The form of m2 is preferred to m1 in most cases when you have a legitimate random slope variable, for example if you had a factor "difficulty" referring to the difficulty of the questions, it would be meaningful to estimate the effect of question difficulty separately for each candidate, so a model like m3 <- lmer(score ~ 1 + gender + difficulty + (difficulty|candidate), mg2006_sub) might be sensible, but not one with gender as a random slope over candidate. Dan On Sun, Mar 28, 2010 at 4:21 PM, Iasonas Lamprianou
<lamprianou at yahoo.com> wrote:
Dear colleagues, I am not sure what the difference between those models is: m0<- lmer(score ~ 1+gender+(1|candidate), mg2006_sub) m1<- lmer(score ~ 1+(1+gender|candidate), mg2006_sub) m2 <- lmer(score ~ 1+gender+(1+gender|candidate), mg2006_sub) the first model is modelling the candidate as a random effect in an examination, where two markers mark each response of a candidate (a repeated measure). I assume that the gender of the candidate is a good predictor of performance on the test, so I can use any of the three models. But I do not understand what the difference is. Why would I get different results between m0 and m1? In effect, I am just adding the gender as a fixed effect.And is m2 a valid model? thank you jason Dr. Iasonas Lamprianou Assistant Professor (Educational Research and Evaluation) Department of Education Sciences European University-Cyprus P.O. Box 22006 1516 Nicosia Cyprus Tel.: +357-22-713178 Fax: +357-22-590539 Honorary Research Fellow Department of Education The University of Manchester Oxford Road, Manchester M13 9PL, UK Tel. 0044 ?161 275 3485 iasonas.lamprianou at manchester.ac.uk --- On Sun, 28/3/10, r-sig-mixed-models-request at r-project.org <r-sig-mixed-models-request at r-project.org> wrote:
From: r-sig-mixed-models-request at r-project.org <r-sig-mixed-models-request at r-project.org> Subject: R-sig-mixed-models Digest, Vol 39, Issue 42 To: r-sig-mixed-models at r-project.org Date: Sunday, 28 March, 2010, 11:00 Send R-sig-mixed-models mailing list submissions to ??? r-sig-mixed-models at r-project.org To subscribe or unsubscribe via the World Wide Web, visit ??? https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models or, via email, send a message with subject or body 'help' to ??? r-sig-mixed-models-request at r-project.org You can reach the person managing the list at ??? r-sig-mixed-models-owner at r-project.org When replying, please edit your Subject line so it is more specific than "Re: Contents of R-sig-mixed-models digest..." Today's Topics: ???1. Re: Could not get a lmer or glmer summary (Eric Edeline) ???2. Re: Significance and lmer (Ben Bolker) ???3. Re: Significance and lmer (Adam D. I. Kramer) ???4. Re: Significance and lmer (David Duffy) ???5. Re: Significance and lmer (Adam D. I. Kramer) ???6. Very weird lmer results, compared to SAS proc mix (Yong Wu) ???7. Re: Very weird lmer results, compared to SAS proc mix ? ? ? (hadley wickham) ---------------------------------------------------------------------- Message: 1 Date: Sat, 27 Mar 2010 14:59:48 +0100 From: Eric Edeline <edeline at biologie.ens.fr> To: David Duffy <David.Duffy at qimr.edu.au> Cc: r-sig-mixed-models at r-project.org Subject: Re: [R-sig-ME] Could not get a lmer or glmer summary Message-ID: <4BAE0F54.3000809 at biologie.ens.fr> Content-Type: text/plain; charset=ISO-8859-1; format=flowed I have solved my problem, which was apparently due to a conflict between lme4 and another library (one of these: tree, VGAM, sn, Matrix, mclust...) about an "rcon" object if I remember well. Removing all the libraries including lme4 from /usr/local/lib/R/site-library and re-installing lme4 in /usr/lib/R/site-library made the trick. Sorry for not providing more detailed information, I just do not remember more! Cheers, eric David Duffy wrote:
On Fri, 26 Mar 2010, Eric Edeline wrote:
Dear Ben, thank you for your feed-back. I have now tested
lmer on several
datasets and I always get the same error message
when asking for
model summary. So the problem is with lme4, not
with the data. Then,
I ran the exact same models and data on another
machine and it works
fine! So the lme4 problem is specific to my
machine. Then, I tried
brute force: uninstalling and re-installing R on
my machine, but the
lme4 problem remains.
Therefore, you either need to "just" extract the
results you want from
m11 directly (doing any necessary calculations yourself),
or step through
using a debugger, or send all the files to Douglas Bates ;) Cheers, David Duffy.
-- Eric Edeline Assistant Professor UMR 7618 BIOEMCO Ecole Normale Sup?rieure 46 rue d'Ulm 75230 Paris cedex 05 France Tel: +33 (0)1 44 32 38 84 Fax: +33 (0)1 44 32 38 85 http://www.biologie.ens.fr/bioemco/biodiversite/edeline.html ------------------------------ Message: 2 Date: Sat, 27 Mar 2010 15:04:42 +0000 (UTC) From: Ben Bolker <bolker at ufl.edu> To: r-sig-mixed-models at r-project.org Subject: Re: [R-sig-ME] Significance and lmer Message-ID: <loom.20100327T160050-336 at post.gmane.org> Content-Type: text/plain; charset=us-ascii Adam D. I. Kramer <adik at ...> writes:
Dear colleagues, Please consider this series of commands: a <- lmer(log(stddiff+.1539) ~ pred + m*v + option
+ (option|studyID),
data=r1, subset=option>1, REML=FALSE) b <- update(a, . ~ . - pred) anova(a,b) ...am I mistaken in thinking that the latter command
will produce a test of
whether "pred" is a significant predictor of
log(stddiff+.1539)? I am
concerned because of the results:
? [snip]
...a significant result completely unrelated to the
t-value. My
interpretation of this would be that we have no good
evidence that the
estimate for 'pred' is nonzero, but including pred in
the model improves
prediction.
? It is possible for Wald tests (as provided by summary()) to disagree radically with likelihood ratio tests (look up "Hauck-Donner effects", but my guess is that's not what's going on here (it definitely can apply in binomial models, don't think it should apply to LMMs but ?). ? I have seen some wonky stuff happen with update() [sorry, can't provide any reproducible details], I would definitely try fitting b by spelling out the full model rather than using update() and see if that makes a difference. ? Other than that, nothing springs to mind. ? (Where does the log(x+0.1539) transformation come from???) ------------------------------ Message: 3 Date: Sat, 27 Mar 2010 10:09:41 -0700 (PDT) From: "Adam D. I. Kramer" <adik at ilovebacon.org> To: Ben Bolker <bolker at ufl.edu> Cc: r-sig-mixed-models at r-project.org Subject: Re: [R-sig-ME] Significance and lmer Message-ID: <Pine.LNX.4.64.1003270955500.17783 at ilovebacon.org> Content-Type: TEXT/PLAIN; charset=US-ASCII; format=flowed On Sat, 27 Mar 2010, Ben Bolker wrote:
...a significant result completely unrelated to
the t-value. My
interpretation of this would be that we have no
good evidence that the
estimate for 'pred' is nonzero, but including pred
in the model improves
prediction.
? It is possible for Wald tests (as provided by
summary()) to disagree
radically with likelihood ratio tests (look up
"Hauck-Donner effects", but
my guess is that's not what's going on here (it
definitely can apply in
binomial models, don't think it should apply to LMMs
but ?). There are no Wald tests produced by the summary()...my understanding from reading this list is that the t-values are provided because they are t-like (effect / se), but that it is difficult (and perhaps foolish) to estimate degrees of freedom for t. So my concern is based on the fact that t is very small.
? I have seen some wonky stuff happen with
update() [sorry, can't provide
any reproducible details], I would definitely try
fitting b by spelling
out the full model rather than using update() and see
if that makes a
difference.
This produces no difference in b's estimates or the anova() statistics. (That said, I originally was fitting [implicitly] with REML=TRUE, which did make a difference, but not a big one).
? Other than that, nothing springs to mind.
Well, thanks for the reply. Are you, then, of the opinion that the above interpretation is reasonable?
? (Where does the log(x+0.1539) transformation
come from???) x is power-law distributed with a bunch of zeroes (but not ordinal, or I'd use family=poisson), and .1539 is the 25th percentile. This normalizes is pretty well. Good question, though! And thanks ofr the response! --Adam ------------------------------ Message: 4 Date: Sun, 28 Mar 2010 08:04:03 +1000 (EST) From: David Duffy <David.Duffy at qimr.edu.au> To: "Adam D. I. Kramer" <adik at ilovebacon.org> Cc: r-sig-mixed-models at r-project.org Subject: Re: [R-sig-ME] Significance and lmer Message-ID: <Pine.LNX.4.64.1003280753090.29716 at orpheus.qimr.edu.au> Content-Type: TEXT/PLAIN; charset=US-ASCII; format=flowed On Sat, 27 Mar 2010, Adam D. I. Kramer wrote:
On Sat, 27 Mar 2010, Ben Bolker wrote:
...a significant result completely unrelated
to the t-value. My
interpretation of this would be that we have
no good evidence that the
estimate for 'pred' is nonzero, but including
pred in the model improves
prediction.
? I have seen some wonky stuff happen with
update() [sorry, can't provide
any reproducible details], I would definitely try
fitting b by spelling
out the full model rather than using update() and
see if that makes a
difference.
This produces no difference in b's estimates or the
anova() statistics.
(That said, I originally was fitting [implicitly] with
REML=TRUE, which did
make a difference, but not a big one).
The two models both have the same number of observations, one hopes?? How many observations per studyID and how many studyIDs?
Well, thanks for the reply. Are you, then, of the
opinion that the above
interpretation is reasonable?
I would be a bit nervous.? My interpretation would be that the model is inappropriate for the data (as the Wald and LR tests should roughly agree for a LMM, as Ben pointed out), and would look at diagnostic plots of residuals etc.? The bunch of zeroes you mention may still be stuffing things up ;)? Is a left-censored model plausible? Just my 2c, David Duffy. -- | David Duffy (MBBS PhD) ???,-_|\ | email: davidD at qimr.edu.au ph: INT+61+7+3362-0217 fax: -0101? / ???* | Epidemiology Unit, Queensland Institute of Medical Research???\_,-._/ | 300 Herston Rd, Brisbane, Queensland 4029, Australia? GPG 4D0B994A v ------------------------------ Message: 5 Date: Sat, 27 Mar 2010 16:17:53 -0700 (PDT) From: "Adam D. I. Kramer" <adik at ilovebacon.org> To: David Duffy <David.Duffy at qimr.edu.au> Cc: r-sig-mixed-models at r-project.org Subject: Re: [R-sig-ME] Significance and lmer Message-ID: <Pine.LNX.4.64.1003271609530.17783 at ilovebacon.org> Content-Type: TEXT/PLAIN; charset=US-ASCII; format=flowed The problem turned out to be, indeed, differing numbers of observations. This is likely due to me relying too much on update() to work as I expected...it did not drop the observations previously dropped. The help page for update makes it very clear that it just re-evaluates an altered call, so this is my fault. Ben's comment about update() being wonky should have given me a hint. Preselecting cases using complete.cases() for both models brought the t values and chi-square values much closer together--when t=.51 for the coefficient, the chisq of a likelihood test for removing the variable from the model was chisq=.25, leading to a reasonable p=.62. Thanks very much to you and Ben Bolker! --Adam On Sun, 28 Mar 2010, David Duffy wrote:
On Sat, 27 Mar 2010, Adam D. I. Kramer wrote:
On Sat, 27 Mar 2010, Ben Bolker wrote:
...a significant result completely
unrelated to the t-value. My
interpretation of this would be that we
have no good evidence that the
estimate for 'pred' is nonzero, but
including pred in the model improves
prediction.
? I have seen some wonky stuff happen with
update() [sorry, can't provide
any reproducible details], I would definitely
try fitting b by spelling
out the full model rather than using update()
and see if that makes a
difference.
This produces no difference in b's estimates or
the anova() statistics.
(That said, I originally was fitting [implicitly]
with REML=TRUE, which did
make a difference, but not a big one).
The two models both have the same number of
observations, one hopes?? How
many observations per studyID and how many studyIDs?
Well, thanks for the reply. Are you, then, of the
opinion that the above
interpretation is reasonable?
I would be a bit nervous.? My interpretation
would be that the model is
inappropriate for the data (as the Wald and LR tests
should roughly agree for
a LMM, as Ben pointed out), and would look at
diagnostic plots of residuals
etc.? The bunch of zeroes you mention may still
be stuffing things up ;)? Is
a left-censored model plausible? Just my 2c, David Duffy. -- | David Duffy (MBBS PhD)
???,-_|\
| email: davidD at qimr.edu.au
ph: INT+61+7+3362-0217 fax: -0101? / ???*
| Epidemiology Unit, Queensland Institute of Medical
Research???\_,-._/
| 300 Herston Rd, Brisbane, Queensland 4029,
Australia? GPG 4D0B994A v
------------------------------ Message: 6 Date: Sat, 27 Mar 2010 23:25:21 -0500 From: Yong Wu <wuyong88 at gmail.com> To: r-sig-mixed-models at r-project.org Subject: [R-sig-ME] Very weird lmer results, compared to SAS proc mix Message-ID: ??? <cfa5b89e1003272125r1677f3ddl8004de6f726683cd at mail.gmail.com> Content-Type: text/plain Sorry to bother you. I am struggling in this issue for long time. Wish somebody can help me. I first used lmer to do the following analysis. fullmodel=lmer(BMI~1+exposure+(age|ID),data, REML=FALSE) reducemodel=lmer(BMI~1+(age|ID),data, REML=FALSE) ? ? ? ? ? anova(full,red) The "fullmodel" has AIC of 6874 and "reducemodel" has AIC of 7106, which cause "anova" analysis giving the p-value< 2.2e-16 . This result is definitely wrong I then did the similar study by SAS. The fullmodel is: proc mixed; class exposure; model BMI=exposure; random age /sub=id; run; The AIC is 7099.7, and type 3 test of fixed effect, exposure, got p-value=0.74. The reducemodel is: proc mixed; ?class exposure; ?model BMI=; ?random age /sub=id; ?run; ?The AIC is 7101.2. The SAS result is correct. Could somebody help me to explain why lmer is wrong? I do not even dare to use lmer now, since I can not trust its result. Thanks in advance for any of your answer. Best, Yong , ??? [[alternative HTML version deleted]] ------------------------------ Message: 7 Date: Sat, 27 Mar 2010 23:55:54 -0500 From: hadley wickham <h.wickham at gmail.com> To: Yong Wu <wuyong88 at gmail.com> Cc: r-sig-mixed-models at r-project.org Subject: Re: [R-sig-ME] Very weird lmer results, compared to SAS proc ??? mix Message-ID: ??? <f8e6ff051003272155l4501611dnebf8d57c8cfe9f5e at mail.gmail.com> Content-Type: text/plain; charset=ISO-8859-1 On Sat, Mar 27, 2010 at 11:25 PM, Yong Wu <wuyong88 at gmail.com> wrote:
Sorry to bother you. I am struggling in this issue for
long time. Wish
somebody can help me. I first used lmer to do the following analysis. fullmodel=lmer(BMI~1+exposure+(age|ID),data,
REML=FALSE)
? ? ? ? ?reducemodel=lmer(BMI~1+(age|ID),data,
REML=FALSE)
? ? ? ? ?anova(full,red) The "fullmodel" has AIC of 6874 and "reducemodel" has
AIC of 7106, which
cause "anova" analysis giving the p-value< 2.2e-16
. This result is
definitely wrong
How do you know?? It would be helpful if you provided the evidence you used to judge SAS correct and R incorrect. Hadley -- Assistant Professor / Dobelman Family Junior Chair Department of Statistics / Rice University http://had.co.nz/ ------------------------------
_______________________________________________ R-sig-mixed-models mailing list R-sig-mixed-models at r-project.org https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models End of R-sig-mixed-models Digest, Vol 39, Issue 42 **************************************************
_______________________________________________ R-sig-mixed-models at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models