Linear Mixed Models - Contrast Estimation - Please help!
Locked
 
|
Hi there - I am trying to run a LMM and am trying to estimate a rather simple contrast option (test linear trend over 4 semesters/time point). But keep getting the below error, could someone please let me know what I should be doing to run this sucessfully.
I am running this on SPSSv21 error: Custom hypothesis test 1 (LINEAR TIME) will not be performed because the L matrix is not estimable. Syntax: MIXED GPA BY Cluster Semester Sex WITH Age1stround Parental_SES SATMath SATVerbal SATWriting /CRITERIA=CIN(95) MXITER(100) MXSTEP(10) SCORING(1) SINGULAR(0.000000000001) HCONVERGE(0, ABSOLUTE) LCONVERGE(0, ABSOLUTE) PCONVERGE(0.000001, ABSOLUTE) /FIXED=Cluster Semester Sex Age1stround Cluster*Semester*Sex Cluster*Semester Cluster*Sex Semester*Sex SATMath SATVerbal SATWriting Parental_SES | SSTYPE(3) /METHOD=REML /PRINT=SOLUTION /REPEATED=Semester | SUBJECT(BARCS_ID) COVTYPE(TPH) /TEST = 'LINEAR TIME' Semester -3 -1 1 3. |
Re: Linear Mixed Models - Contrast Estimation - Please help!
Administrator
|
I don't know why you're getting that error message. I assume your Semester variable has values 1-4. Have you considered treating Semester as a quantitative (scaled) variable rather than as categorical? (I'd probably recode 1-4 into 0-3 if doing this so that the constant gives the fitted value for semester 1.) The one df test for Semester would be testing the linear fit. (You could also include the quadratic & cubic terms if you wanted.) I'd be curious to know if that model would run, or if you'd run into the same problem.
Another thing I might try is simplifying the model a bit (e.g., omit all interactions) to see if the problem still happens. If the simpler model runs okay, gradually add terms back in to see if you can isolate the term or terms that are causing the problem. HTH.
--
Bruce Weaver bweaver@lakeheadu.ca http://sites.google.com/a/lakeheadu.ca/bweaver/ "When all else fails, RTFM." PLEASE NOTE THE FOLLOWING: 1. My Hotmail account is not monitored regularly. To send me an e-mail, please use the address shown above. 2. The SPSSX Discussion forum on Nabble is no longer linked to the SPSSX-L listserv administered by UGA (https://listserv.uga.edu/). |
Re: Linear Mixed Models - Contrast Estimation - Please help!
|
|
In reply to this post by ARIA
You don't mention what "cluster" is and how many d.f. it requires (times 8),
=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD
or whether your N is large enough to support the analysis. -- Rich Ulrich > Date: Fri, 19 Dec 2014 11:06:28 -0700 > From: [hidden email] > Subject: Linear Mixed Models - Contrast Estimation - Please help! > To: [hidden email] > > Hi there - I am trying to run a LMM and am trying to estimate a rather simple > contrast option (test linear trend over 4 semesters/time point). But keep > getting the below error, could someone please let me know what I should be > doing to run this sucessfully. > > I am running this on SPSSv21 > > error: > Custom hypothesis test 1 (LINEAR TIME) will not be performed because the L > matrix is not estimable. > > Syntax: > MIXED GPA BY Cluster Semester Sex WITH Age1stround Parental_SES SATMath > SATVerbal SATWriting > /CRITERIA=CIN(95) MXITER(100) MXSTEP(10) SCORING(1) > SINGULAR(0.000000000001) HCONVERGE(0, ABSOLUTE) LCONVERGE(0, ABSOLUTE) > PCONVERGE(0.000001, ABSOLUTE) > /FIXED=Cluster Semester Sex Age1stround Cluster*Semester*Sex > Cluster*Semester Cluster*Sex Semester*Sex SATMath SATVerbal SATWriting > Parental_SES | SSTYPE(3) > /METHOD=REML > /PRINT=SOLUTION > /REPEATED=Semester | SUBJECT(BARCS_ID) COVTYPE(TPH) > /TEST = 'LINEAR TIME' Semester -3 -1 1 3. > > > > |
|
|
In reply to this post by Bruce Weaver
SPSS is correct to state it is non-estimable. The OP needs to write out the coefficient (L) matrix.
Ryan Sent from my iPhone > On Dec 19, 2014, at 2:54 PM, Bruce Weaver <[hidden email]> wrote: > > I don't know why you're getting that error message. I assume your Semester > variable has values 1-4. Have you considered treating Semester as a > quantitative (scaled) variable rather than as categorical? (I'd probably > recode 1-4 into 0-3 if doing this so that the constant gives the fitted > value for semester 1.) The one df test for Semester would be testing the > linear fit. (You could also include the quadratic & cubic terms if you > wanted.) I'd be curious to know if that model would run, or if you'd run > into the same problem. > > Another thing I might try is simplifying the model a bit (e.g., omit all > interactions) to see if the problem still happens. If the simpler model > runs okay, gradually add terms back in to see if you can isolate the term or > terms that are causing the problem. > > HTH. > > > > > ARIA wrote >> Hi there - I am trying to run a LMM and am trying to estimate a rather >> simple contrast option (test linear trend over 4 semesters/time point). >> But keep getting the below error, could someone please let me know what I >> should be doing to run this sucessfully. >> >> I am running this on SPSSv21 >> >> error: >> Custom hypothesis test 1 (LINEAR TIME) will not be performed because the L >> matrix is not estimable. >> >> Syntax: >> MIXED GPA BY Cluster Semester Sex WITH Age1stround Parental_SES SATMath >> SATVerbal SATWriting >> /CRITERIA=CIN(95) MXITER(100) MXSTEP(10) SCORING(1) >> SINGULAR(0.000000000001) HCONVERGE(0, ABSOLUTE) LCONVERGE(0, ABSOLUTE) >> PCONVERGE(0.000001, ABSOLUTE) >> /FIXED=Cluster Semester Sex Age1stround Cluster*Semester*Sex >> Cluster*Semester Cluster*Sex Semester*Sex SATMath SATVerbal SATWriting >> Parental_SES | SSTYPE(3) >> /METHOD=REML >> /PRINT=SOLUTION >> /REPEATED=Semester | SUBJECT(BARCS_ID) COVTYPE(TPH) >> /TEST = 'LINEAR TIME' Semester -3 -1 1 3. > > > > > > ----- > -- > Bruce Weaver > [hidden email] > http://sites.google.com/a/lakeheadu.ca/bweaver/ > > "When all else fails, RTFM." > > NOTE: My Hotmail account is not monitored regularly. > To send me an e-mail, please use the address shown above. > > -- > View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Linear-Mixed-Models-Contrast-Estimation-Please-help-tp5728227p5728228.html > Sent from the SPSSX Discussion mailing list archive at Nabble.com. > > ===================== > To manage your subscription to SPSSX-L, send a message to > [hidden email] (not to SPSSX-L), with no body text except the > command. To leave the list, send the command > SIGNOFF SPSSX-L > For a list of commands to manage subscriptions, send the command > INFO REFCARD ===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD |
Re: Linear Mixed Models - Contrast Estimation - Please help!
|
Administrator
|
Ryan, I think you'll have elaborate a little bit.
Here is an example from the FM: Example Suppose that factor A has three levels and factor B has four levels. MIXED Y BY A B /FIXED = A B /TEST = ’test example’ A 1 -1 0 DIVISOR=3; B 0 0 1 -1 DIVISOR=4. * For effect A, all contrast coefficients will be divided by 3; therefore, the actual coefficients are (1/3,–1/3,0). * For effect B, all contrast coefficients will be divided by 4; therefore, the actual coefficients are (0,0,1/4,–1/4). --- End of Example --- Presumably, this should also work: MIXED Y BY A B /FIXED = A B /TEST = ’test example’ B 0 0 1 -1. This is very similar to the OP's TEST sub-command, the main difference being the absence of DIVISOR=x. (I can't find an explicit statement in the FM, but presumably the default value = 1.) So...what do you mean by "write out the coefficient (L) matrix"? Thanks for clarifying.
--
Bruce Weaver bweaver@lakeheadu.ca http://sites.google.com/a/lakeheadu.ca/bweaver/ "When all else fails, RTFM." PLEASE NOTE THE FOLLOWING: 1. My Hotmail account is not monitored regularly. To send me an e-mail, please use the address shown above. 2. The SPSSX Discussion forum on Nabble is no longer linked to the SPSSX-L listserv administered by UGA (https://listserv.uga.edu/). |
|
|
Bruce, The OP is trying to test for a linear trend within a more complex model than you obtained from the FM. The model proposed by the OP includes interaction terms, which leads me to wonder exactly what the OP wants. Let me try to clarify how one would normally test for a group-specific linear trends using orthogonal polynomials by providing a simulation experiment (BELOW MY NAME). After running the code and reviewing the output, feel free to write back with further comments. Bottom line--The TEST sub-command the OP wrote is not estimable as the interation terms are being ignored, and frankly it isn't clear to me what the objective is of the test. OTOH, the code I provide below is constructed correctly to estimate and test group-specific slopes and provide a formal interaction test treating "A" as a categorical variable with ordinal properties and then treating "A" as a covariate. Ryan -- *Generate data. set seed 4836235. new file. input program. loop ID = 1 to 1000. IF (ID<= 100) A=1. IF (ID> 100 and ID <= 200) A=2. IF (ID> 200 and ID <= 300) A=3. IF (ID> 300) A=4. compute B = rv.bernoulli(0.50). compute y = 10*(B=0) + 20*(B=1) + 2.0*(B=0)*A + 4.0*(B=1)*A + rv.normal(0,1). end case. end loop. end file. end input program. execute. *Treat A as a categorical variable with ordinal properties. MIXED y BY B A /FIXED=B A B*A | SSTYPE(3) /METHOD=REML /PRINT=SOLUTION /TEST 'linear trend B=0' A -3 -1 1 3 B*A -3 -1 1 3 0 0 0 0 DIVISOR=10 /TEST 'linear trend B=1' A -3 -1 1 3 B*A 0 0 0 0 -3 -1 1 3 DIVISOR=10 /TEST 'linear trend interaction' B*A -3 -1 1 3 3 1 -1 -3 DIVISOR=10. *Treat A as a covariate. MIXED y BY B WITH A /FIXED=B A B*A | SSTYPE(3) /METHOD=REML /PRINT=SOLUTION. *Reparameterize to directly obtain group-specific intercepts and slopes from Estimates of Fixed Effects Table. MIXED y BY B WITH A /FIXED=B B*A | NOINT SSTYPE(3) /METHOD=REML /PRINT=SOLUTION. On Sat, Dec 20, 2014 at 8:56 AM, Bruce Weaver <[hidden email]> wrote: Ryan, I think you'll have elaborate a little bit. |
|
|
One small adjustment. Where I stated that one could obtain the group-specific *intercepts* by reparameterizing the model (last MIXED code), the covariate "A" ought to be recoded as 0, 1, 2, 3 in the data generation portion. This a side, relatively minor point that has nothing to do with the thrust of my post. All the MIXED code is correct. Ryan On Sat, Dec 20, 2014 at 9:48 PM, Ryan Black <[hidden email]> wrote:
|
|
|
In reply to this post by Bruce Weaver
Sorry for the delay in response and thanks for the suggestions. I did make the model simpler and tried the contrasts. There is no error message until i start including interactions of two categorical variables for eg if i test for semester*age, it runs fine but when i add semester*sex (both categorical predictors) it gives me the error i mentioned.
Not sure if SPSS treats the categorical interactions in a different way or if I need to recode all my categorical variables as scale (again not sure if that would be appropriate) cause when it comes to cluster (which is basically my groups) i would like to keep them as categorical. Any suggestions would be helpful. |
Re: Linear Mixed Models - Contrast Estimation - Please help!
|
Administrator
|
In reply to this post by Ryan
Good morning Ryan, and thanks for elaborating.
Now that I have access to SPSS again, I tried your syntax, but I also added this variation, which is similar to what the OP was attempting: * Now try a contrast like the OP attempted. MIXED y BY B A /FIXED=B A B*A | SSTYPE(3) /METHOD=REML /PRINT=SOLUTION /TEST 'linear trend A' A -3 -1 1 3 . It produced the same warning the OP posted, i.e.: Warnings Custom hypothesis test 1 (linear trend A) will not be performed because the L matrix is not estimable. I was thinking that perhaps MIXED would (by default) estimate the linear trend for A with all other terms set to zero (or their reference categories). I.e., I thought that the following two lines would do the same thing (using your data): /TEST 'linear trend A' A -3 -1 1 3 /TEST 'linear trend B=0' A -3 -1 1 3 B*A -3 -1 1 3 0 0 0 0 But clearly, that is not so. Also, in the table of fixed effects, the highest value of B, not the lowest, is treated as the reference category. So even if I had been right about there being some plugging in of default values, I probably would have been wrong about what the defaults were for categorical predictors! Thanks again. I think this should clarify things for the OP.
--
Bruce Weaver bweaver@lakeheadu.ca http://sites.google.com/a/lakeheadu.ca/bweaver/ "When all else fails, RTFM." PLEASE NOTE THE FOLLOWING: 1. My Hotmail account is not monitored regularly. To send me an e-mail, please use the address shown above. 2. The SPSSX Discussion forum on Nabble is no longer linked to the SPSSX-L listserv administered by UGA (https://listserv.uga.edu/). |
|
|
Hi Bruce, We may only eliminate the effects which have all zero (0) coefficients. Since, that is not the case for "B*A -3 -1 1 3 0 0 0 0", it must remain. Moreover, the TEST statement you brought up: /TEST 'linear trend B=0' A -3 -1 1 3 B*A -3 -1 1 3 0 0 0 0 DIVISOR=10 ...is estimating the (linear) slope for one level of B. This again leads me to wonder what the OP really wants. The slope for A for each level of B? The difference in slopes for A between levels of B? The average slope for A across all levels of B? To make matters more complicated, from what I recall, there were up to three-way interactions in the model offered up by the OP. Best, Ryan On Mon, Dec 22, 2014 at 9:29 AM, Bruce Weaver <[hidden email]> wrote: Good morning Ryan, and thanks for elaborating. |
Re: Linear Mixed Models - Contrast Estimation - Please help!
|
Administrator
|
Until the OP speaks up again, we'll have to guess at what they really want.
If they have ANOVA goggles on, they may be wanting to average over the levels of the other variables (rather than get simple main effects). For your example, averaging over the two levels of B would be done something like this, I think: *Treat A as a categorical variable with ordinal properties. MIXED y BY B A /FIXED=B A B*A | SSTYPE(3) /METHOD=REML /PRINT=SOLUTION /TEST 'linear trend B=0' A -3 -1 1 3 B*A -3 -1 1 3 0 0 0 0 DIVISOR=10 /TEST 'linear trend B=1' A -3 -1 1 3 B*A 0 0 0 0 -3 -1 1 3 DIVISOR=10 /TEST 'linear trend averaging over B' A -3 -1 1 3 B*A -3/2 -1/2 1/2 3/2 -3/2 -1/2 1/2 3/2 DIVISOR=10 /TEST 'linear trend interaction' B*A -3 -1 1 3 3 1 -1 -3 DIVISOR=10.
--
Bruce Weaver bweaver@lakeheadu.ca http://sites.google.com/a/lakeheadu.ca/bweaver/ "When all else fails, RTFM." PLEASE NOTE THE FOLLOWING: 1. My Hotmail account is not monitored regularly. To send me an e-mail, please use the address shown above. 2. The SPSSX Discussion forum on Nabble is no longer linked to the SPSSX-L listserv administered by UGA (https://listserv.uga.edu/). |
|
|
Hi Bruce and Ryan - I am trying to test two different things i guess a) if the average slope follows a linear trend averaged across clusters and b) if the linear slopes across my temporal variable (semester w/ 4 levels) are different across my groups (cluster - 3 levels). Since this is the first time i'm doing something like this im not really sure if what I had in the model that I posted was accurate.
|
Re: Linear Mixed Models - Contrast Estimation - Please help!
|
Administrator
|
In reply to this post by ARIA
Nabble is still showing the message below as "not yet posted to the mailing list", so some members may not have seen it yet.
Aria, have you seen Ryan's message posted Dec 21, 2014; 3:26pm? He and I have both posted subsequent messages too. We're still trying to determine if you want the linear trend for Semester at some specific set of levels of the other variables it interacts with, or averaging across levels, or what. HTH.
--
Bruce Weaver bweaver@lakeheadu.ca http://sites.google.com/a/lakeheadu.ca/bweaver/ "When all else fails, RTFM." PLEASE NOTE THE FOLLOWING: 1. My Hotmail account is not monitored regularly. To send me an e-mail, please use the address shown above. 2. The SPSSX Discussion forum on Nabble is no longer linked to the SPSSX-L listserv administered by UGA (https://listserv.uga.edu/). |
|
|
Yeah not sure why its showing that, i am new user here so not sure if I am doing something wrong. I did send a followup msg, are you able to see that?
|
Re: Linear Mixed Models - Contrast Estimation - Please help!
|
Administrator
|
Reading via Nabble (http://spssx-discussion.1045642.n5.nabble.com/), I can see all of your posts--but all but the first one are tagged with "This post has NOT been accepted by the mailing list yet."
UGA's archives for SPSSX-L (https://listserv.uga.edu/cgi-bin/wa?A0=SPSSX-L) show only your first message thus far. The fact that your first message made it to the mailing list suggests you are properly registered, so I'm not sure why your subsequent posts are being held up.
--
Bruce Weaver bweaver@lakeheadu.ca http://sites.google.com/a/lakeheadu.ca/bweaver/ "When all else fails, RTFM." PLEASE NOTE THE FOLLOWING: 1. My Hotmail account is not monitored regularly. To send me an e-mail, please use the address shown above. 2. The SPSSX Discussion forum on Nabble is no longer linked to the SPSSX-L listserv administered by UGA (https://listserv.uga.edu/). |
|
|
In reply to this post by Ryan
Hi Ryan and Bruce - For some reason my msgs have been late to appear. However, my objective is to look at the below
The difference in slopes for A between levels of B? (i.e. difference in slopes for semester(time) between levels of cluster(group)) The average slope for A across all levels of B? (i.e. average slope for semester(time) across all levels of cluster(group). Hope this clarifies things in terms of what I am interested in. I would really appreciate advice on how to prescribe the contrasts for the above. Thx |
|
|
Dear ____, I believe I provided you with simulation code and syntax to answer most of your questions. Bruce provided additional code to show you how to obtain an estimate of the slope averaged across levels of another factor. Have you attempted to adopt what we provided to your situation? Did it work? If not, please explain what went wrong and provide code. Ryan On Tue, Dec 23, 2014 at 10:02 AM, ARIA <[hidden email]> wrote: Hi Ryan and Bruce - For some reason my msgs have been late to appear. |
«
Return to SPSSX Discussion
|
1 view|%1 views
| Free forum by Nabble | Edit this page |