Full-information maximum likelihood. Mplus doesn't have REML.
Anonymous posted on Saturday, November 22, 2003 - 10:31 am
When I check a loglikelihood value of HO in MPLUS this value is almost the same as one calculated by SPSS using REML method. Is the default estimator (MLR : maximum likelihood estimator with robust standard error) in MPLUS is similar to REML in other software ?
*********************** other software MPLUS
FML ML REML MLR ***********************
If not, what is the difference between MLR in MPLUS and REML ?
Thanks for the quick reply
bmuthen posted on Saturday, November 22, 2003 - 10:51 am
The Mplus ML estimator is the same as FML (also called FIML). Mplus does not have REML. REML ("restricted ML") has been proposed with the aim of getting better variance estimates with small samples. You will probably see rather small differences between REML and ML (FML). The Mplus MLR estimator gives ML (FML) parameter estimates and robust SEs. Such SEs are also called sandwich or Huber-White. The robustness is against non-normality and model misspecification primarily.
Helen Dennis posted on Tuesday, December 30, 2003 - 12:47 pm
I ran a null model in mplus using the format: MODEL: %WITHIN% y; %BETWEEN% y;
To be sure that I was running what I thought I was running, I also ran the null model in HLM 5.0 with FIML. I got similar estimates for the intercept, but VERY different estimates for the level 1 and level 2 variances (much larger estimates in Mplus). Should I have expected this? If yes, Why? THANKS
bmuthen posted on Tuesday, December 30, 2003 - 12:54 pm
I don't think that you have the same data for the two runs. HLM has 54 clusters and Mplus has 57.
Sally Czaja posted on Friday, November 05, 2004 - 9:16 am
As in the 11/20/03 posting, I want to run an unconditional multi-level model. The dependent variable is dichotomous. Following the 11/21/03 advice, my commands include: ANALYSIS: TYPE IS twolevel ; ESTIMATOR IS mlr; MODEL: %WITHIN% Resi6o8A; %BETWEEN% Resi6o8A; OUTPUT: SampStat standardized; I get 2 error messages: *** ERROR in Model command Variances for categorical outcomes are not currently defined on the within level. Variance given for: RESI6O8A *** ERROR in Model command Variances for categorical outcomes can only be specified using PARAMETERIZATION=THETA with estimators WLS, WLSM, or WLSMV. Variance given for: RESI6O8A
What should I be doing differently? Thanks.
bmuthen posted on Friday, November 05, 2004 - 11:21 am
In 2-level analysis of a binary variable, there is only a second-level variance to be estimated, not a first-level variance. So, delete Resi6o8A from the Within part of the model.
Vera posted on Saturday, December 17, 2005 - 12:15 pm
I am also trying to run the unconditional means model using the syntax you provided. However, I get the following message: *** ERROR in Model command Unknown variables: %WITHIN% in line: %WITHIN%
What am I doing wrong? Here's what I entered: DATA: FILE IS anx.dat; VARIABLE: NAMES ARE y; MODEL: %WITHIN% y; %BETWEEN% y;
bmuthen posted on Saturday, December 17, 2005 - 1:01 pm
I am comparing variance estimates between my null model and one with fixed individual level predictors (no between). The variance estimate for my null model is smaller compared to the second model mentioned above with predictors.
The response variable is dichotomous. Seems counterintuitive since I am adding predictors, one would expected variances to decrease. Is this something you have seen before? Thank you.
With categorical outcomes there is in general not a free parameter for a residual variance of a factor indicator because it cannot be identified as a separate parameter. Exceptions are multiple-group modeling and longitudinal modeling. The negative residual variance that you see is computed as a function of other parameters. Therefore it is not easy to fix or constrain to be positive.
Typically a negative residual variance is an indication that something is wrong with the model so that it should be modified. For example, is the exe factor really unrelated to gender and peduc?
I am comparing the estimation of an empty or unconditional 2 level model (no covariates - just variance decomposition) between different software packages. HLM and Mlwin give me very similar results whereas Mplus is doing something differently. Overall, Mplus gives me higher estimates for the within and between variances and SE. Can you please explain me why this is happening?
If you have the same data, the same sample size, the same model, and the same estimator, the three programs will give the same results. I would check these three things first. Otherwise I would need to see the three outputs and your license number at email@example.com.
I am trying to discern if it is possible to use Mplus to run a meta-analysis. I am able to run an unconditional random effects model in HLM, using the V-Known feature. The V-Known feature, the level-1 variance of the model is treated as "known", such that:
Level one model Y_ij = b_0j + e_ij
where e_ij = "known" sampling variance derived from each study
I would like to know if it is possible to specify an unconditional twolevel model in which the y is a latent variable (y by x1 and x2). How to deal with the fact that within-level variables cannot by used on the between level?
If you don't put a variable on either the BETWEEN or WITHIN lists, the variable can be used on both levels. See Example 9.1 for a more detailed explanation. Note that a factor with two indicators is not identified.
Utkun Ozdil posted on Thursday, April 28, 2011 - 12:42 pm
Hi, I'm analyzing a two-level structural equation model with categorical outcomes. While testing the unconditional model my syntax is as below in its simplest form: VARIABLE: NAMES ARE class q1-q8 USEVARIABLES ARE class q1-q8; CATEGORICAL ARE q1-q8; CLUSTER IS class; ANALYSIS: TYPE IS TWOLEVEL; ESTIMATOR IS WLSM; MODEL: %WITHIN% fw BY q1 q3 q2 q4; aw BY q5 q6 q8 q7; %BETWEEN% fb BY q4 q2 q1 q3; ab BY q8 q6 q7 q5; OUTPUT: SAMPSTAT STANDARDIZED;
In order to report how much of the total variance is attributable to the within-level and to the between-level partially, are we going to use the "Variances" section for the latent variables fw,aw,fb,and ab? Or are we to request something else in the OUTPUT command? Thanks in advance...
For the factors, you would need to hold the factor loadings equal across between and within. This may not make sense. If it does, the model estimated variances of the factors would show you the within and between parts.
Utkun Ozdil posted on Friday, April 29, 2011 - 10:30 am
I have the estimated variances of the latents for the within and between parts. I calculated the ICCs for these latents by hand.
In some articles the variance attributed to the within-level and between-level are reported in percentages partially. I expected to find these percentages in my output but I think Mplus does not provide these values?
And my last question is about two-level EFA. The Mplus output did not display the variances accounted for by each factor though the eigenvalues are demonstrated. Is there a specific calculation or a command to obtain these percentages?
Joe King posted on Saturday, November 12, 2011 - 4:00 pm
when doing the full unconditional model is OK to just leave the model statement blank and have the outcome variable in the place where i specify the count model (its a ZINB model). also how would one get an estimate of ICC in a two level ZINB model?
kirsten way posted on Monday, September 03, 2012 - 1:06 am
Hi, I've run an unconditional model with the syntax below and the output is stating df=0. Shouldn't the degrees of freedom be related to the number of groups (i.e. for me, 75)?
This is the syntax I have used below
VARIABLE: NAMES ARE ORGID WKGRPID ID TC RC AgYld AgColl AgFce AgAvd anxdep wellbeing harass1;
USEV ARE WKGRPID harass1; CLUSTER IS WKGRPID;
ANALYSIS: TYPE IS TWOLEVEL;
MODEL: %WITHIN% harass1; %BETWEEN% harass1;
*** WARNING in MODEL command Variable is uncorrelated with all other variables: HARASS1 *** WARNING in MODEL command All least one variable is uncorrelated with all other variables in the model. Check that this is what is intended.
I'm running an unconditional means model to get percentage of variance at the within and between level. I wanted to use the B/(B+W) formula for this using the unstandardized variance estimates but I get a between level variance of .000. However the output shows an ICC of .015, I'm wondering if I'm using the wrong information from the output to calculate the variance or if I'm doing something wrong
Thanks for your quick response! So what I'm getting from your response is that the percentage of variance that I want to calculate and the ICCs that Mplus provides are somewhat different things, is this correct? If that's the case, where can I find the variance estimates needed to calculate the percentage of variance explained? Is it the unstandardized variance estimates or should I be looking somewhere else?
I'm not sure I made myself clear. I'm trying to calculate percentage of variance explained of a variable at the between and the within level, not the percentage of variance explained of a dependent variable by an independent variable.
You can do a TYPE=TWOLEVEL BASIC to get the within and between variances that we use to compute ICC's.
anonymous Z posted on Tuesday, February 17, 2015 - 8:44 am
I am working on data of small sample size. I have outcome variables that are dichotomous. I fit several growth models. As you can imagine, the results do not look good. I wondered if you have any general advice about working with small data set using Mplus. For instance, Mplus doesn’t’ have REML. Given REML matters when sample size is small, what will be an alternative way to improve the estimate?
In an earlier post on this page someone asked about just trying to obtain the amount of variance explained in a variable at the within and the between level. Linda responded that the person could "do a TYPE=TWOLEVEL BASIC to get the within and between variances that we use to compute ICC's." I am trying to follow Bliese & Ployhart's 2002 approach to run the null model (i.e., with no random intercept or other random effects added) before I move on to adding time and/or random effects. The approach calls for comparing model fit as you add complexity. If I use the TWOLEVEL BASIC command, I am confused about what I put on the %Between% and %Within% statements. Thank you!
Bliese, P., & Ployhart, R. (2002). Growth Modeling Using Random Coefficient Models: Model Building, Testing, and Illustrations. Organizational Research Methods, 5, 362-387.
With Type=Twolevel Basic you don't specify a %Between% and %Within% model. The Basic analysis uses a model with random intercepts for each outcome. I am not familiar with the Bliese article.
Hongguo Wei posted on Sunday, February 19, 2017 - 8:30 pm
Dear Dr. Muthen,
I am studying leader-follower dyads at work groups. I need to examine the nonindependence of variables and see whether I can analyze them at the individual, dyadic, or group levels.
Can I use the three level unconditional mean model to calculate the within (individual) and between variances (dyad and group), and thus get ICC1s at both the dyad and group levels? May I just use TYPE=Threelevel Random to get the variance at three levels? Thank you very much!
I am trying to run a null model: MODEL: %WITHIN% y; %BETWEEN% y;
and I'm getting the error message [truncated for size]:
THE MODEL ESTIMATION DID NOT TERMINATE NORMALLY DUE TO AN ILL-CONDITIONED FISHER INFORMATION MATRIX..
THE MODEL ESTIMATION DID NOT TERMINATE NORMALLY DUE TO A NON-POSITIVE DEFINITE FISHER INFORMATION MATRIX. THIS MAY BE DUE TO THE STARTING VALUES BUT MAY ALSO BE AN INDICATION OF MODEL NONIDENTIFICATION..
THE STANDARD ERRORS OF THE MODEL PARAMETER ESTIMATES COULD NOT BE COMPUTED...PROBLEM INVOLVING THE FOLLOWING PARAMETER:... %BETWEEN%: y
A MATRIX COULD NOT BE INVERTED DURING THE H1 MODEL ESTIMATION (same with baseline). THE ESTIMATED WITHIN COVARIANCE MATRIX IS NOT POSITIVE DEFINITE AS IT SHOULD BE. COMPUTATION COULD NOT BE COMPLETED. THE VARIANCE OF y APPROACHES 0. FIX THIS VARIANCE AND THE CORRESPONDING COVARIANCES TO 0, DECREASE THE MINIMUM VARIANCE, OR SPECIFY THE VARIABLE AS A BETWEEN VARIABLE.
THE H1 MODEL ESTIMATION DID NOT CONVERGE. CHI-SQUARE TEST AND SAMPLE STATISTICS COULD NOT BE COMPUTED.
THE BASELINE MODEL ESTIMATION DID NOT CONVERGE. THE CHI-SQUARE VALUE COULD NOT BE COMPUTED.
CLUSTER is idclass1 regio1; stratification is stratum1; !corewgt;
ANALYSIS: TYPE IS COMPLEX TWOLEVEL;
idclass1 (primary sampling units - PPS sampling). I am however interested in the second level of school districts (regio1) and just want to control standard errors for the sampling design clusters (idclass1).
The survey weight is sort of a total weight, but: the within-weight is mainly 1 for everybody (complete classrooms were surveyed & almost no non-response existed, hence the survey weight(corewgt) mainly mirrors differences in between classroom selection probabilities due to PPS). However the survey weight has nothing to do with regio1 (second level cluster of analysis) since it mirrors the design cluster (idclass1) selection. What should I do in this situation? I i) run and report the analysis without weights and then ii) I calculated averages of the corewgt across school districts(regio1) to mirror approximately over-and undersampling across school districts and included the survey weight as bweight (between weight) to check if results remain the same, which is the case...
Another question refers to the TLI, which is 1.47 (should I be concerned ?)