Message/Author |
|
EvavdW posted on Wednesday, February 11, 2015 - 8:02 am
|
|
|
Hi, I ran a fairly simple model, in which I want to examine the correlation between four variables, while controlling for age (Grade). The sample consists of 140 participants from 17 classes. When I run the model below, I get a warning: MODEL: Monkey3 WITH Lion3 OOO3 WRB3; Lion3 WITH OOO3 WRB3; OOO3 WITH WRB3; Monkey3 ON Grade; Lion3 ON Grade; OOO3 ON Grade; WRB3 ON Grade; The warning is: THE MODEL ESTIMATION TERMINATED NORMALLY THE STANDARD ERRORS OF THE MODEL PARAMETER ESTIMATES MAY NOT BE TRUSTWORTHY FOR SOME PARAMETERS DUE TO A NON-POSITIVE DEFINITE FIRST-ORDER DERIVATIVE PRODUCT MATRIX. THIS MAY BE DUE TO THE STARTING VALUES BUT MAY ALSO BE AN INDICATION OF MODEL NONIDENTIFICATION. THE CONDITION NUMBER IS -0.101D-16. PROBLEM INVOLVING THE FOLLOWING PARAMETER: Parameter 17, LION3 WITH MONKEY3 THIS IS MOST LIKELY DUE TO HAVING MORE PARAMETERS THAN THE NUMBER OF CLUSTERS MINUS THE NUMBER OF STRATA WITH MORE THAN ONE CLUSTER. However, the standardized estimate between Lion3 and Monkey3 is within the normal range: .335 Can I trust the results from this analysis, despite this warning? With kind regards, Eva |
|
|
Probably, but you don't say what your cluster variable is or if you use Type=Complex. |
|
EvavdW posted on Thursday, February 12, 2015 - 1:10 am
|
|
|
I apologize for leaving out that information :-) My message was to large for the discussion board and so I removed that part. This is the part of the input instructions, indicating the clusters: VARIABLE: NAMES ARE Eprime Grade ClassID Monkey1 Lion1 Monkey2 Lion2 OOO3 WRB3 Monkey3 Lion3; USEV ARE Grade OOO3 WRB3 Monkey3 Lion3; Cluster = ClassID; MISSING ARE ALL (999) ANALYSIS: TYPE = COMPLEX; ESTIMATOR = MLR; So if I understand it correctly, it is probably safe to report the results from this model? Kind regards, Eva |
|
|
Yes, probably ok. But 17 clusters (your classes) is on the low side for getting good SEs and chi-2. |
|
James Algina posted on Saturday, February 14, 2015 - 9:08 am
|
|
|
Hi, Two code fragments follow. Both are intended to allow tests of covariate x treatment interactions at L1 and L2. The first codes a single-group model and brings product terms into the model. The second avoids doing so by coding a two-group model with cross-group equality constraints. The first results in a npd 1ST-order derivative product matrix warning. Estimates and SEs for the two programs were similar but not identical. 1. Do the two approaches specify the same model? 2. Regardless of the answer, is it advisable to avoid bringing product terms into a model? Thanks, Jamie Single Group: InterL1=Cov*Treat; InterL2=Cov_Mean*Treat; analysis: estimator=mlr; type=twolevel; model: %Within% Dv on Cov InterL1 ; Cov with InterL1; %Between% Dv on Cov_Mean InterL2 Treat ; InterL2 on Treat; Cov_Mean on Treat ; Cov_Mean with InterL2; Two Group: %Within% Dv on Cov ; Dv (L1V2); Cov (L1V1); [Cov] (ML1); %Between% Dv on Cov_Mean; Dv (L2V2); Cov_Mean (L2V1); model PYM: %Within% Dv on Cov ; Dv (L1V2); Cov (L1V1); [Cov] (ML1);; %Between% Dv on Cov_Mean ; Dv (L2V2); Cov_Mean (L2V1); |
|
|
Is Treat a between-level binary variable? Is Treat the grouping variable in the 2-grp setup? |
|
James Algina posted on Wednesday, February 18, 2015 - 6:51 pm
|
|
|
1. Treat a between-level binary variable 2. Treat the grouping variable in the 2-grp setup Thanks, Jamie |
|
|
I think either approach is fine. Your 2-group setup is a bit more general in that your residual variances are different in the groups. The npd 1ST-order derivative product matrix warning is most likely due to the binary covariate (mean-variance relationship), so ignorable. |
|
|
I am running a 3-level ML-CFA in an attempt to estimate factorial ICCs. I have applied four constraints to deal with four negative residual variances at the 2nd and 3rd levels. Despite going through steps outlined here, https://www.statmodel.com/download/SaddlePoints2.pdf, I still get the same warning in the output: THE MODEL ESTIMATION HAS REACHED A SADDLE POINT OR A POINT WHERE THE OBSERVED AND THE EXPECTED INFORMATION MATRICES DO NOT MATCH... and, PROBLEM INVOLVING THE FOLLOWING PARAMETER: Parameter 28, %WITHIN%: EM6 Although, estimates for %within% EM6 seem interpretable. Can interpret the output to calculate ICCs? |
|
|
If you get standard errors in your results, you can ignore the message. Otherwise, send the output and your license number to support@statmodel.com. |
|
|
Great Linda, Thanks for that :-) |
|
|
Dear Linda/ Bengt, I'm running a fairly straightforward multilevel cross lagged model: VARIABLE: NAMES = Team IDENT1-IDENT3 I_Bel1-I_Bel3; MISSING = ALL (-999); USEVAR = IDENT1-IDENT3 I_bel1-I_bel3; cluster is team; analysis: type is complex; MODEL: IDENT2 ON IDENT1 (1); IDENT3 ON IDENT2 (1); I_bel2 ON I_bel1 (2); I_bel3 ON I_bel2 (2); IDENT2 ON I_bel1 (3); IDENT3 ON I_bel2 (3); I_bel2 ON IDENT1 (4); I_bel3 ON IDENT2 (4); IDENT2 WITH I_bel2 (5); IDENT3 WITH I_bel3 (5); However, I get the following warning message: THE STANDARD ERRORS OF THE MODEL PARAMETER ESTIMATES MAY NOT BE TRUSTWORTHY FOR SOME PARAMETERS DUE TO A NON-POSITIVE DEFINITE FIRST-ORDER DERIVATIVE PRODUCT MATRIX. THIS MAY BE DUE TO THE STARTING VALUES BUT MAY ALSO BE AN INDICATION OF MODEL NONIDENTIFICATION. THE CONDITION NUMBER IS -0.375D-16. PROBLEM INVOLVING PARAMETER 8. THIS IS MOST LIKELY DUE TO HAVING MORE PARAMETERS THAN THE NUMBER OF CLUSTERS MINUS THE NUMBER OF STRATA WITH MORE THAN ONE CLUSTER. The output appears normal. Am I able to trust the results? |
|
|
With clustered data, independence of observations is at the cluster level. How having more parameters than clusters impacts the results is not well studied. This is a warning. It sounds like you have only 8 clusters. This is not enough for TYPE=COMPLEX or TWOLEVEL. A minimum of 20 or 30-50 is recommended by different sources. You can create a set of dummy variables and use them as covariates to control for non-independence of observations. |
|
|
Thank you Linda! |
|
Back to top |