two factor hierachical model
Locked
two factor hierachical model
 
|
Hi all, this is my first post.
I am currently attempting to fit a two factor hierachical model with a random subject effect (for nested design). However, when I run my model I get the following message "The final hessian matrix is not positive definite though all convergence criteria are satisfied". When I look at my estimates of covariance parameters table, I find that my Intercept (subject) variance is zero, with the message that my covariance parameter is redundant. If anyone can explain what is occuring and give me some advice on how to fix it, I would be very grateful. Thanks. |
|
|
Benjamin,
Please provide an illustration of your dataset and post your syntax. Ryan On Wed, Jun 1, 2011 at 9:46 AM, Benjamin Spivak (Med) <[hidden email]> wrote: > Hi all, this is my first post. > > I am currently attempting to fit a two factor hierachical model with a > random subject effect (for nested design). However, when I run my model I > get the following message "The final hessian matrix is not positive definite > though all convergence criteria are satisfied". When I look at my estimates > of covariance parameters table, I find that my Intercept (subject) variance > is zero, with the message that my covariance parameter is redundant. If > anyone can explain what is occuring and give me some advice on how to fix > it, I would be very grateful. > > Thanks. ===================== 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 |
|
|
In reply to this post by Benjamin Spivak (Med)
My first guess would be that you have mis-specified the model, with the consequence that the "covariance parameter is redundant". What is it calling a "covariate"? My second guess would be that the data, as it is being used by the problem, is not exactly what you expect. Missings? Did it seem to use all the cases? -- Rich Ulrich |
Fwd: two factor hierachical model
|
|
---------- Forwarded message ---------- From: Benjamin Spivak (Med) <[hidden email]> Date: 2 June 2011 09:34 Subject: Re: two factor hierachical model To: Rich Ulrich <[hidden email]> Hello Rich and Ryan,
I have provided the informaton below
Ryan: I am performing a 2x3 design experiment looking at juries and their understanding of the law. I have 63 juries with roughly 10-12 jurors in each group. Both my IV's are categorical and are based on juror level data. I am also attempting to use age, education and gender as predictors in the model. The DV that I am using appears to be normally distributed and has satisfied the assumption of homogeneity of variance. My Syntax is as follows:
MIXED Standards BY EduCon Jurycharge WITH Age Education Gender
/CRITERIA=CIN(95) MXITER(100) MXSTEP(10) SCORING(1) SINGULAR(0.000000000001) HCONVERGE(0, ABSOLUTE) LCONVERGE(0, ABSOLUTE) PCONVERGE(0.000001, ABSOLUTE) /FIXED=EduCon Jurycharge Age Education Gender | SSTYPE(3) /METHOD=ML /PRINT=SOLUTION TESTCOV /RANDOM=INTERCEPT | SUBJECT(Jury) COVTYPE(VC) /EMMEANS=TABLES(EduCon) COMPARE ADJ(BONFERRONI) /EMMEANS=TABLES(Jurycharge) COMPARE ADJ(BONFERRONI). Rich: It is calling my subject grouping a covariate. As for missings, yes there was a proportion of missing data in my DV (no missing data anywhere else). However, I tried replacing missing values with group means and still encountered the same problem. Thanks to both of you,
Ben.
On 2 June 2011 03:56, Rich Ulrich <[hidden email]> wrote:
|
Re: two factor hierachical model
|
|
Ben, Two ideas. Have you tried the unstructured covariance structure. Or what about looking at the frequencies of scores on your dv and covariates by your categorical ivs. Maybe some of those cells have gotten too small? A quick google search also leads me to think the variance of your intercept across jury’s may not be varying. Although it’s hard for me to imagine what that means for your data. After adjusting for your categorical ivs and your covariates there is no difference in means across jurys? Matt From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Benjamin Spivak (Med) ---------- Forwarded message ---------- Hello Rich and Ryan, I have provided the informaton below Ryan: I am performing a 2x3 design experiment looking at juries and their understanding of the law. I have 63 juries with roughly 10-12 jurors in each group. Both my IV's are categorical and are based on juror level data. I am also attempting to use age, education and gender as predictors in the model. The DV that I am using appears to be normally distributed and has satisfied the assumption of homogeneity of variance. My Syntax is as follows: MIXED Standards BY EduCon Jurycharge WITH Age Education Gender Thanks to both of you, Ben.
On 2 June 2011 03:56, Rich Ulrich <[hidden email]> wrote:
|
Re: two factor hierachical model
|
|
Hi Matt, Thanks for the response.
Yes, I have tried the unstructured covariance structure, unfortunately when I attempt to do this SPSS gives me an error message and proceeds by defaulting to scaled identity covariance structure. As for your second point, I am not sure how to compute this within spss.
In regards to the last point, I am worried that this might be the case, as mean differences from jury to jury are quite small. I'm not sure what to do, any ideas?
Thanks,
Ben.
On 2 June 2011 10:49, Matthew Pirritano <[hidden email]> wrote:
|
Re: two factor hierachical model
|
|
Ben, I just recently had the same experience. A variance that was essentially zero caused spss to raise an error and set to the identity covariance structure. I found out that other software can sometimes estimate models even with some estimated parameters near zero. So it looks like including intercept on your random line is what’s causing the problem. Why not just run the model without intercept as a random variable? That’s what it sounds like anyway. All juries are starting at about the same intercept. Matt From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Benjamin Spivak (Med)
Thanks for the response. Yes, I have tried the unstructured covariance structure, unfortunately when I attempt to do this SPSS gives me an error message and proceeds by defaulting to scaled identity covariance structure. As for your second point, I am not sure how to compute this within spss. In regards to the last point, I am worried that this might be the case, as mean differences from jury to jury are quite small. I'm not sure what to do, any ideas? Thanks, Ben. On 2 June 2011 10:49, Matthew Pirritano <[hidden email]> wrote:
|
|
|
In reply to this post by Benjamin Spivak (Med)
Benjamin,
First off, it does not make sense to specify an UNstructured covariance matrix on the RANDOM statement, since you are only estimating a random intercept. The random intercept has nothing with which to covary. To your original question...A non-positive hessian matrix may suggest a small intraclass correlation coefficient. That is, the variance between the jury means is smaller than would be expected given the within-jury variance. This by itself is not necssarily a problem. You may be just fine reporting fixed effects estimates from such a model. Having stated that, there may be related underlying issues that need to be addressed. HTH, Ryan On Wed, Jun 1, 2011 at 9:14 PM, Benjamin Spivak (Med) <[hidden email]> wrote: > > Hi Matt, > > Thanks for the response. > > Yes, I have tried the unstructured covariance structure, unfortunately when > I attempt to do this SPSS gives me an error message and proceeds by > defaulting to scaled identity covariance structure. As for your second > point, I am not sure how to compute this within spss. > > In regards to the last point, I am worried that this might be the case, as > mean differences from jury to jury are quite small. I'm not sure what to do, > any ideas? > > Thanks, > > Ben. > > > > On 2 June 2011 10:49, Matthew Pirritano <[hidden email]> > wrote: >> >> Ben, >> >> >> >> Two ideas. Have you tried the unstructured covariance structure. Or what >> about looking at the frequencies of scores on your dv and covariates by your >> categorical ivs. Maybe some of those cells have gotten too small? >> >> >> >> A quick google search also leads me to think the variance of your >> intercept across jury’s may not be varying. >> >> >> >> >> http://groups.google.com/group/comp.soft-sys.sas/browse_thread/thread/da82a20cc8aba8d1/aa1d0c37f4d8a0e5?hl=en&lnk=gst&q=hessian+matrix+positive+definite+stringplayer_2#aa1d0c37f4d8a0e5 >> >> >> >> Although it’s hard for me to imagine what that means for your data. After >> adjusting for your categorical ivs and your covariates there is no >> difference in means across jurys? >> >> >> >> Matt >> >> >> >> From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of >> Benjamin Spivak (Med) >> Sent: Wednesday, June 01, 2011 4:35 PM >> To: [hidden email] >> Subject: Fwd: two factor hierachical model >> >> >> >> >> >> ---------- Forwarded message ---------- >> From: Benjamin Spivak (Med) <[hidden email]> >> Date: 2 June 2011 09:34 >> Subject: Re: two factor hierachical model >> To: Rich Ulrich <[hidden email]> >> >> Hello Rich and Ryan, >> >> >> >> I have provided the informaton below >> >> >> >> Ryan: I am performing a 2x3 design experiment looking at juries and their >> understanding of the law. I have 63 juries with roughly 10-12 jurors in each >> group. Both my IV's are categorical and are based on juror level data. I am >> also attempting to use age, education and gender as predictors in the model. >> The DV that I am using appears to be normally distributed and has satisfied >> the assumption of homogeneity of variance. My Syntax is as follows: >> >> >> >> MIXED Standards BY EduCon Jurycharge WITH Age Education Gender >> /CRITERIA=CIN(95) MXITER(100) MXSTEP(10) SCORING(1) >> SINGULAR(0.000000000001) HCONVERGE(0, >> ABSOLUTE) LCONVERGE(0, ABSOLUTE) PCONVERGE(0.000001, ABSOLUTE) >> /FIXED=EduCon Jurycharge Age Education Gender | SSTYPE(3) >> /METHOD=ML >> /PRINT=SOLUTION TESTCOV >> /RANDOM=INTERCEPT | SUBJECT(Jury) COVTYPE(VC) >> /EMMEANS=TABLES(EduCon) COMPARE ADJ(BONFERRONI) >> /EMMEANS=TABLES(Jurycharge) COMPARE ADJ(BONFERRONI). >> >> Rich: It is calling my subject grouping a covariate. As for missings, yes >> there was a proportion of missing data in my DV (no missing data anywhere >> else). However, I tried replacing missing values with group means and still >> encountered the same problem. >> >> >> >> Thanks to both of you, >> >> >> >> Ben. >> >> >> >> On 2 June 2011 03:56, Rich Ulrich <[hidden email]> wrote: >> >> My first guess would be that you have mis-specified the model, >> with the consequence that the "covariance parameter is redundant". >> What is it calling a "covariate"? >> >> My second guess would be that the data, as it is being used by >> the problem, is not exactly what you expect. Missings? Did it >> seem to use all the cases? >> >> -- >> Rich Ulrich >> >> >> >> > ===================== 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: two factor hierachical model
|
|
In reply to this post by Matthew Pirritano
Matt,
Thanks for that advice,
If i remove the intercept will my resultant output still control for the fact that my data is nested.
Regards,
Ben.
On 2 June 2011 11:42, Matthew Pirritano <[hidden email]> wrote:
|
Re: two factor hierachical model
|
|
Maybe you don’t need to use a multilevel model. The idea of the mixed model is that there is not an independence of observations. The mixed model lets you allow for the possibility that there are relationships within groups that may differ by group. If you’re not going to include any random components of your groups, then you’re not allowing for the variability in these relationships by group. Maybe I’m wrong but in that case I think you’re basically running a regular OLS regression. Or you could add the categorical factors and maybe interactions to the random line. It seems possible that your slopes may vary between groups although your intercepts do not. If your slopes turn out not to vary then your back to OLS. From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Benjamin Spivak (Med) Matt, Thanks for that advice, If i remove the intercept will my resultant output still control for the fact that my data is nested. Regards, Ben. On 2 June 2011 11:42, Matthew Pirritano <[hidden email]> wrote: Ben, I just recently had the same experience. A variance that was essentially zero caused spss to raise an error and set to the identity covariance structure. I found out that other software can sometimes estimate models even with some estimated parameters near zero. So it looks like including intercept on your random line is what’s causing the problem. Why not just run the model without intercept as a random variable? That’s what it sounds like anyway. All juries are starting at about the same intercept. Matt From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Benjamin Spivak (Med)
Subject: Re: two factor hierachical model
Thanks for the response. Yes, I have tried the unstructured covariance structure, unfortunately when I attempt to do this SPSS gives me an error message and proceeds by defaulting to scaled identity covariance structure. As for your second point, I am not sure how to compute this within spss. In regards to the last point, I am worried that this might be the case, as mean differences from jury to jury are quite small. I'm not sure what to do, any ideas? Thanks, Ben. On 2 June 2011 10:49, Matthew Pirritano <[hidden email]> wrote:
|
Re: two factor hierachical model
|
|
Hi Matt,
Theoretically, I have violated the assumption of independence. As jurors deliberated in groups before completing questionnaires. However, if there is no clear evidence of relationships within groups, do I still need to perform multilevel modelling, or can I ignore my jury groups and proceed to analyze everything at the level of the juror (perhaps using a simpler ANOVA/ANCOVA design)?
Cheers,
Ben.
On 2 June 2011 12:14, Matthew Pirritano <[hidden email]> wrote:
|
Re: two factor hierachical model
|
|
In reply to this post by Rich Ulrich
Hi Rich,
I get the same message when I run the model without any covariates, so I don't think this is the problem. Thanks anyway, Ben.
On 2 June 2011 16:52, Rich Ulrich <[hidden email]> wrote:
|
| Free forum by Nabble | Edit this page |