Anonymous posted on Sunday, October 02, 2005 - 4:33 pm
I am estimating several models which involve a single factor or multiple factors predicting various outcome variables (one at a time). In some of these models (i.e., those that contain a substantial amount of missing data) I am using the TYPE = MISSING command. Data missing on variables used to estimate the latent factors have already been listwise deleted, so technically, missing data on the outcome variable can only be handeled using this command. All analyses use the WLSMV estimator. I understand that MPLUS models missingness differently when WLSMV is used depending on whether covariates are included in the model.
1. Would it be correct that when one of my models involves one latent factor predicting a single outcome (and TYPE = MISSING is specified) that pairwise deletion is being used?
2. Would it be correct that when one of my models involves multiple latent factors predicting a single outcome (and TYPE = MISSING is specified) that some other method of handling missing data is being used?
3. If so, could you please describe this other method of handling missing data (i.e., what is being done to handle the missing data)?
4. Is it problematic that I already listwise deleted (a small) part of the sample to estimate the latent factors, and am then using the "MISSING" command on missing data on an outcome variable being predicted by the latent factors? What sort of an impact do you think this mixed approach towards dealing with missing data in my study will have on my results? (I am trying to get a sense of whether I should go back and re-estimate these models with the missing data on the variables used to estimate the latent factors re-included in the data set)
bmuthen posted on Monday, October 03, 2005 - 10:40 am
1. Pairwise deletion is used with categorical outcomes and Type = Missing.
2. The number of factors does not influence the missing data handling, only whether or not observed covariates are present.
4. Listwise deletion is obtained when Type = Missing is not used. Type = Missing gives pairwise deletion with categorical outcomes. I would recommend using Type = Missing throughout.
Kim Henry posted on Thursday, July 10, 2008 - 11:46 am
I want to make sure that I understand how mplus handles missing data when type=WLSMV? If I understand correctly, only information from the X variables is used when dealing with missing data on the outcome variables (that is, for example, when estimating the regression of Y1 on X1-X3, Y2 is not used to help consider missingness on Y1. Is this correct? And, when I write this up, is their a term for this type of consideration of the missingness. For example, in MLR - missing data is dealt with by using full information maximum likelihood.
There is not a term for this that I know. WLSMV with covariates x works in 4 steps: univariate probit regression of each u on the x's using all people with data on that u (and the x's), bivariate probit regression of each pair of u's on the x's using all people with data for that pair, estimation of the weight matrix, and fitting the model using weighted least squares. The first 2 steps use ML estimation. This means that this is better than pairwise present data for the u's because missingness is allowed to be affected by the x's and so can be quite selective. So it has an MAR flavor wrt the x's. But the final results in step 4 are not MAR in the sense that for the u's only pairs of u's are used in the first 2 ML steps, not all of them. So for instance attrition giving missingness for a later u predicted by an early u would not give consistent results. This is the price paid for the simplicity of the WLSMV approach.
Erika Wolf posted on Friday, March 12, 2010 - 8:21 am
Do you have a citation that would be approporiate for describing how WLSMV handles missing data (re--your response above on 7/10/08)? Thank you.
The only place we describe this in on page 7 of the user's guide. You can look for a citation on pairwise present which is the method we use when there are no covariates.
leah lipsky posted on Monday, March 15, 2010 - 12:36 pm
If I'm using WLSMV with missing data, how do I know how many subjects are used in the model (I'm assuming that with pairwise deletion, the model only estimates based on subjects with no missing data)? My full sample is N = 413, and I know there are missing data for my 2 dependent variables, but the output says there are 413 observations. thanks.
Pairwise deletion uses different sample sizes for different pairs of dependent variables, but it sounds like you have only one pair given only 2 DVs. That sounds like none of the 413 has missing on both DVs. I believe in this case that Mplus would delete subjects with missing on both DVs.
leah lipsky posted on Tuesday, March 16, 2010 - 6:53 am
Thanks for your response. I should have clarified that there are 2 IVs and 2 DVs. I checked my variables, and there are 45 subjects who were missing for both DVs, and 2 subjects are missing for both IVs. It sounds like you're saying I should expect the number of observations to be the full sample (n = 413) reduced by the number of subjects who are missing on both DVs (n = 45), but this is not the case (output says # observations = 413). I'd very much appreciate any further advice. Thank you.
In reference to Bengt's post above on 7/10/2008 - I'm wondering what estimation and handling of missing data are used for those outcomes that are not declared as categorical but estimated using WLSMV (because there are categorical predictors that are also dependent variables in the path model). Specifically, my outcome variable is continuous (and with considerable missing data); several of my key predictor variables are categorical or binary, these variables are also dependent variables in the model, which necessitates use of WLSMV. Given this set up, what process is used to estimate Y (continuous outcome) on U (categorical predictor), and how is missing data on Y handled?
I'm a bit confused about earlier discussions in this forum about missing data treatment with WLSMV. I've read several articles that report to use the WLSMV estimator for parameter estimation. At the same time, these papers report using the FIML method to handle missing data. From the technical appendix (WLSMV with missing data) I understand that WLSMV uses unvariate FIML estimates as the first stage estimate "sigma^1".
If WLSMV can use FIML estimates at stage one, using pairwise deletion as missing data treatment doesn't make sense to me (given that MAR holds), as FIML is said to be more efficient under MAR than pairwise deletion.
With TYPE=GENERAL and ESTIMATOR=WLSMV (in MPLus Version 5), does WLSMV use a FIML method or pairwise deletion for missing data treatment?
I'd be very grateful for any advice on this topic. Thank you!
Missing data theory does not apply to the univariate case. Therefore, it is not involved in the univariate FIML estimates that are used as first stage estimates.
WLSMV uses pairwise present for missing. Maximum likelihood and categorical outcomes uses FIML.
Sarah Ryan posted on Thursday, October 06, 2011 - 11:37 am
Regarding your above explanation of missing data handling by WLSMV on Thursday 10/8/2008, let me make sure I understand.
"WLSMV with covariates x works in 4 steps: univariate probit regression of each u on the x's using all people with data on that u (and the x's)...,"
Q1) MEANING THAT IF U IS MISSING, THE CASE IS DROPPED OR MEANING THAT U IS INFERRED GIVEN INFORMATION ON X'S?
"... So it has an MAR flavor wrt the x's. But the final results in step 4 are not MAR in the sense that for the u's only pairs of u's are used in the first 2 ML steps, not all of them. So for instance attrition giving missingness for a later u predicted by an early u would not give consistent results. "
q2) MEANING THAT WE MUST BE CONFIDENT THAT THE ESTIMATES IN STAGE1 WERE THE "TRUE VALUES" (IN PARTICULAR, FOR THOSE MISSING ON THE U IN STAGE1) IN ORDER TO CONCLUDE THAT WE HAVE OBTAINED CONSISTENT RESULTS IN THE FINAL STAGE?
Q1) Meaning that the case is dropped, which would also be the case when FIML is used and there is only 1 DV.
Q2) I think it is the Stage 2 (conditional correlation) estimation that we should worry about. "So for instance attrition giving missingness for a later u predicted by an early u would not give consistent results. "
To avoid missingness with WLSMV you can first do Multiple Imputation. See Topic 9, May 2011 version.
Carolyn CL posted on Wednesday, August 14, 2013 - 10:06 am
Dear Drs. Muthen,
I am estimating a structural equation model with missing data on x's (two of which are continuous latent variables) and y's. Two of my y's are categorical ordinal variables which has lead to the use of WLSMV estimation with theta parametrization. Because I wish to use FIML estimation methods to deal with the missing data, I included an auxiliary variable (family SES at birth) which I allow to correlate with all observed variables (Enders, 2010).
I am having a difficult time clearly articulating the estimation method in my Methods section, as I wish to draw a comparison between the WLSMV and FIML methods.
For the sake of clarity, I re-estimated the method treating the ordinal level y variables as continuous, in order to compare the WLSMV and FIML methods. In both cases, the 'Numer of observations' corresponds to the full sample. In both cases, the number of observed missing data patterns and covariance coverage are the same. In the case of the WLSMV, the Chi-square and df values are smaller. CFI and RMSEA are comparable for both estimation methods. In terms of the parameters, the regression coefficient estimates tend to be slightly larger and the standard errors slightly smaller in the WLSMV method.
Carolyn CL posted on Wednesday, August 14, 2013 - 10:08 am
(NOTE: When running the model using FIML, I get an expected error 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 -0.286D-14. PROBLEM INVOLVING PARAMETER 106.)
Am I correct in the following:
(i) The weighted least squares estimation with missing data method gives parameter estimates that are similar to those using full information maximum likelihood estimation when missing data assumptions are met (Asparouhov & Muthèn, 2010). (ii) Missing data assumptions are that missing data in y are explained by a covariate x (in this case, family SES at birth and other x's). (iii) By using a saturated correlate model (whereby all observed variables are allowed to correlate with an auxiliary variable associated with attrition) all participants who contributed information to the model were retained in the WLSMV analyses (Davey, Shanahan and Schafer, 2001; Enders, 2010). (iv) The estimation of the parameters in WLSMV benefited from the retention of complete and partial data, including that of participants who would have been more likely to desist from the study over time.
Yes, it results in logistic regression as the default if the categorical variables are put on the CATEGORICAL list.
JW posted on Thursday, September 11, 2014 - 7:57 am
Thanks for your reply.
When I use ML I receive the following message, I suspect as I have 15 categorical observed variables:
THE CHI-SQUARE TEST IS NOT COMPUTED BECAUSE THE FREQUENCY TABLE FOR THE LATENT CLASS INDICATOR MODEL PART IS TOO LARGE.
This means I obtain no goodness of fit indeces... Could I report AIC/BIC in the write-up for a paper?
Is there any other way to obtain RMSEA, CFI or TLI?
JW posted on Thursday, September 11, 2014 - 8:50 am
From previous posts, I am under the impression that I should request TECH10 - however, I am not sure which part of the TECH10 output would give me an indication of how good is the fit of the model - could you help pls?
I'm running an EFA with categorical and continuous variables with varying missingness on the indicators (N = 997; 21 variables). I'm using the WLSMV estimator, Type = Missing, and culstering by by community group (24 communities), but I want to ensure I understand how this method handles missingness for EFAs.
My understanding is that this method uses pairwise deletion, is that correct? However, my output indicates that I have 997 observations and I'm a bit confused as there are some variables with missingness on both variables, as was posted above.
Additionally, does clustering change the way missingness is handled in any way?
I apologize for the additional post. I realized that the type that I'm actually using is type = individual and listwise = off (to do pairwise delete). It is my understanding that Type = missing is no longer used and received an error message.
TYPE=MISSING is the default. The full number of observations is printed although each correlation is based on however many people have both variables. Clustering does not change how missingness is handled.
I have read though this forum but still need some guidance on handling NMAR longitudinal missing data (due to attrition). Wave 1 has 1,182 participants and 400 completed Wave 2. My final longitudinal models use cross-lagged SEM and and multi-group analyses. Estimator is WLSMV because I have both continuous and categorical variables. As I understand it, it is not possible to go multi-group analysis with imputed data (and I am not sure imputing data makes sense when 65% of Wave 2 data are missing) and FIML is not possible with WLSMV.
Can you recommend an approach for handling missing data in this case (or point me to resources to help make this decision)? Thank you very much.
Ping Kuo posted on Monday, August 24, 2015 - 6:35 am
Hello I'm running a CFA of three-factor model using WLSMV. (a) If I did not use Type function, is type=missing default? (b) In the situation above, the pairwise deletion is used. Is it correct? Thanks.
a. Yes, this has been the case for some time. b. Yes, when the model has no covariates.
Ping Kuo posted on Monday, August 24, 2015 - 8:40 am
Hello, Thanks for your quick responses. If I test longitudinal measurement invariance using WLSMV ( I did not use the type function), is the pairwise deletion used? In my LMI model, the same factors at difference time points are correlated.