SPSS Hessian Matrix Error - Help!
Locked
 
|
This post was updated on .
I am trying to conduct a linear growth model in SPSS using mixed models. We conducted a training study where participants (interviewers) conducted 10 weeks of interviews with children, receiving regular peer-review. Our dependent measures are their questioning techniques, and how they might change over time (e.g. does the proportion of yes-no questions decrease?; defined as a variable of their proportion of yes-no questions each week, for ten weeks).
When I run the model in SPSS, I get an error stating the following: The final Hessian matrix is not positive definite although all convergence criteria are satisfied. The MIXED procedure continues despite this warning. Validity of subsequent results cannot be ascertained. In addition, on the estimates of covariance parameters, I get the following error: This covariance parameter is redundant. The test statistic and confidence interval cannot be computed. I have tentatively conducted a one-way ANOVA to assess whether there is level 2 variation in my outcome, and there is. This was significant at p < .001. I am unsure why I am receiving the current error for the growth model and would like to solve the problem. Here is my syntax: Please help! MIXED YN_Proportion WITH Interview_Number /CRITERIA=CIN(95) MXITER(100) MXSTEP(10) SCORING(1) SINGULAR(0.000000000001) HCONVERGE(0, ABSOLUTE) LCONVERGE(0, ABSOLUTE) PCONVERGE(0.000001, ABSOLUTE) /FIXED=Interview_Number | SSTYPE(1) /METHOD=ML /PRINT=CPS CORB DESCRIPTIVES SOLUTION TESTCOV /RANDOM=INTERCEPT Interview_Number | SUBJECT(Participant) COVTYPE(ARH1). |
Re: SPSS Hessian Matrix Error - Help!
|
|
I may be misreading and misunderstanding your mixed statement but I read it to say that you have pure level 1 model, i.e., no level 2 predictors. Your level 1 model is y(I,j) = B0+B1*IW# +e(I,j).
Where B0 an B1 are both random, as you specify. Have you run a model development sequence beginning with a random intercepts model? And did that solve without issues? Which covariance parameter is identified as redundant? The B1 variance or the B1,B2 covariance? Could on or both of these be not significant? What happens if you remove the random effect? It sounds like you a number of DVs to cycle through this general model. When you look at the within interviewer distributions for this proportion variable, are you satisfied that the data are adequately modeled by a normal distribution rather than a logistic distribution that is available in Genlinmixed? Gene Maguin -----Original Message----- From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Stolz100 Sent: Friday, April 04, 2014 11:58 AM To: [hidden email] Subject: SPSS Hessian Matrix Error - Help! I am trying to conduct a linear growth model in SPSS using mixed models. We conducted a training study where participants (interviewers) conducted 10 weeks of interviews with children, receiving regular peer-review. Our dependent measures are their questioning techniques, and how they might change over time (e.g. does the proportion of yes-no questions decrease?). When I run the model in SPSS, I get an error stating the following: The final Hessian matrix is not positive definite although all convergence criteria are satisfied. The MIXED procedure continues despite this warning. Validity of subsequent results cannot be ascertained. In addition, on the estimates of covariance parameters, I get the following error: This covariance parameter is redundant. The test statistic and confidence interval cannot be computed. I have tentatively conducted a one-way ANOVA to assess whether there is level 2 variation in my predictor, and there is. This was significant at p < .001. I am unsure why I am receiving the current error for the growth model and would like to solve the problem. Here is my syntax: Please help! MIXED YN_Proportion WITH Interview_Number /CRITERIA=CIN(95) MXITER(100) MXSTEP(10) SCORING(1) SINGULAR(0.000000000001) HCONVERGE(0, ABSOLUTE) LCONVERGE(0, ABSOLUTE) PCONVERGE(0.000001, ABSOLUTE) /FIXED=Interview_Number | SSTYPE(1) /METHOD=ML /PRINT=CPS CORB DESCRIPTIVES SOLUTION TESTCOV /RANDOM=INTERCEPT Interview_Number | SUBJECT(Participant) COVTYPE(ARH1). -- View this message in context: http://spssx-discussion.1045642.n5.nabble.com/SPSS-Hessian-Matrix-Error-Help-tp5725300.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 |
|
|
In reply to this post by Stolz100
1. Your response is binomial, so GENLINMIXED is preferred.
2. Please define what Interview_Number means. I was expecting to see "week" as the random slope. Ryan On Fri, Apr 4, 2014 at 11:58 AM, Stolz100 <[hidden email]> wrote: I am trying to conduct a linear growth model in SPSS using mixed models. We |
|
|
This post was updated on .
Hi. Children's responses are proportions ranging from .00 - 1.00. Interview
number is the week number, ranging from the 1st interview at week 1 to the 10th interview at week 10. The ARH1 rho covariance is identified as redundant. If I remove the random effect for interview number (week), then the model runs just fine. However, the model does not give me any measure of the relationship between the intercept and the slope over the 10 weeks. Ideally, I would like to be able to say something about how interviewer's starting point related to their progress across the training. The data seem normally distributed across the three dependent measures that I need to conduct growth curves for (proportion of: yes-no questions, "Wh" questions, open-invitation questions). Thank you two so much for thinking through this with me! I've never actually conducted a GENLINMIXED so if that is what is necessary I would appreciate any tutorials or academic resources you recommend regarding that analysis. Warm regards, Stacia On Fri, Apr 4, 2014 at 2:28 PM, Ryan Black [via SPSSX Discussion] < ml-node+s1045642n5725318h22@n5.nabble.com> wrote: > 1. Your response is binomial, so GENLINMIXED is preferred. > 2. Please define what Interview_Number means. I was expecting to see > "week" as the random slope. > > Ryan > > > On Fri, Apr 4, 2014 at 11:58 AM, Stolz100 <[hidden email]<http://user/SendEmail.jtp?type=node&node=5725318&i=0> > > wrote: > >> I am trying to conduct a linear growth model in SPSS using mixed models. >> We >> conducted a training study where participants (interviewers) conducted 10 >> weeks of interviews with children, receiving regular peer-review. Our >> dependent measures are their questioning techniques, and how they might >> change over time (e.g. does the proportion of yes-no questions decrease?). >> >> When I run the model in SPSS, I get an error stating the following: The >> final Hessian matrix is not positive definite although all convergence >> criteria are satisfied. The MIXED procedure continues despite this >> warning. >> Validity of subsequent results cannot be ascertained. >> >> In addition, on the estimates of covariance parameters, I get the >> following >> error: This covariance parameter is redundant. The test statistic and >> confidence interval cannot be computed. >> >> I have tentatively conducted a one-way ANOVA to assess whether there is >> level 2 variation in my predictor, and there is. This was significant at >> p < >> .001. >> >> I am unsure why I am receiving the current error for the growth model and >> would like to solve the problem. Here is my syntax: >> >> Please help! >> >> >> MIXED YN_Proportion WITH Interview_Number >> /CRITERIA=CIN(95) MXITER(100) MXSTEP(10) SCORING(1) >> SINGULAR(0.000000000001) HCONVERGE(0, >> ABSOLUTE) LCONVERGE(0, ABSOLUTE) PCONVERGE(0.000001, ABSOLUTE) >> /FIXED=Interview_Number | SSTYPE(1) >> /METHOD=ML >> /PRINT=CPS CORB DESCRIPTIVES SOLUTION TESTCOV >> /RANDOM=INTERCEPT Interview_Number | SUBJECT(Participant) COVTYPE(ARH1). >> >> >> >> -- >> View this message in context: >> http://spssx-discussion.1045642.n5.nabble.com/SPSS-Hessian-Matrix-Error-Help-tp5725300.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] <http://user/SendEmail.jtp?type=node&node=5725318&i=1>(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 >> > > > > ------------------------------ > If you reply to this email, your message will be added to the discussion > below: > > http://spssx-discussion.1045642.n5.nabble.com/SPSS-Hessian-Matrix-Error-Help-tp5725300p5725318.html > To unsubscribe from SPSS Hessian Matrix Error - Help!, click here<http://spssx-discussion.1045642.n5.nabble.com/template/NamlServlet.jtp?macro=unsubscribe_by_code&node=5725300&code=U3RhY2lhLlN0b2x6ZW5iZXJnQGdtYWlsLmNvbXw1NzI1MzAwfDExMjQzODc2ODE=> > . > NAML<http://spssx-discussion.1045642.n5.nabble.com/template/NamlServlet.jtp?macro=macro_viewer&id=instant_html%21nabble%3Aemail.naml&base=nabble.naml.namespaces.BasicNamespace-nabble.view.web.template.NabbleNamespace-nabble.view.web.template.NodeNamespace&breadcrumbs=notify_subscribers%21nabble%3Aemail.naml-instant_emails%21nabble%3Aemail.naml-send_instant_email%21nabble%3Aemail.naml> > |
|
|
Stacia, The covariance between the random components should not be redundant. Given the message you are receiving, I suspect that the variable you believe to represent "week" is in fact equal to your subject identification variable. Are you 100% certain this is not the case?
The dataset should be structured in long format as follows: ID week 1 0 1 1 1 2 1 3 . .
. 2 0 2 1 2 2 2 3 . . . Please write back confirming that you have not inadvertently made the mistake I mentioned above and that your dataset is structured as I have illustrated.
Ryan On Fri, Apr 4, 2014 at 5:45 PM, Stolz100 <[hidden email]> wrote:
|
|
|
Hi, My data is structured as illustrated. Thanks for your help. Stacia
|
|
|
All, I had a discussion with Stacia off-list to determine the reason for the error message. In the spirit of sharing solutions to problems posted to the list, I will briefly explain what we discovered along with a few solutions. We discovered that the estimated random slope variance was near zero. And more critically, the estimated association between the random intercept and slope was so small that it could not be estimated, which led to non-convergence. The following three changes to the model resulted in full convergence: 1. Assuming the correlation between the intercept and slope is zero 2. Removal of the random slope term entirely
3. Fitting a marginal model (REPEATED) statement instead of the a subject-specific model (RANDOM statement) Because the DV is a binomial proportion, a [GEE or random effects] binomial logistic regression model should be considered.
HTH. Ryan On Fri, Apr 4, 2014 at 10:21 PM, Stolz100 <[hidden email]> wrote:
|
«
Return to SPSSX Discussion
|
1 view|%1 views
| Free forum by Nabble | Edit this page |