Lily Wang posted on Sunday, February 05, 2012 - 12:23 pm
Hi, Drs Muthens,
I encounter a problem (accidental termination, to be specific) when trying to impute the data. The data is a national data (TYPE=COMPLEX). I wrote:
DATA IMPUTATION: impute=V1 V2 V3; save=CCimpute*.dat;
Mplus terminates when doing second imputation (as shown in the DOS window) without any of further notice or error message. The output window does not pop up as usual. If I open the output file manually, the file only contains the syntax I wrote.
And you want to remove the Within= line since that would mis-specify the variables as not having any between-level variance.
2) You should only use
Between = Alevel2;
during imputation and estimation phases.
FN briere posted on Friday, March 30, 2012 - 10:06 am
Thank you, this is very useful.
I gather from a different post that to for models with random slopes and cross-level interactions, it is necessary to switch to a H0 approach. I wish to do something as close as possible to H1, but including random slopes.
I would tend to do this by specifying a H0 imputation model with random slopes and all correlations between variables at the two levels.
e.g. Usevariables = x1 x2 y1 Alevel2 ; cluster = SCHOOL; between = Alevel2 ; MISSING are all (-100);
ANALYSIS: TYPE = TWOLEVEL; ESTIMATOR = BAYES;
MOdel: %WITHIN% s | y1 on x1; y1 with x2; x1 with x2; %BETWEEN% s y1 x2 x1 Alevel2 with s y1 x2 x1 Alevel2;
DATA IMPUTATION: IMPUTE = x1 x2 y1 Alevel2; NDATASETS = 5; SAVE = twolevel*.dat;
Does that seem like a correct approach? Thank you for your time again,
I think that's ok. It sounds like your primary interest is in getting imputed data for some later investigation, not estimating the model parmeters. You can estimate the model parameters without imputing.
Assuming imputing is the primary interest, in general it may be a good idea to impute from a model that is as close to the "true" model as possible. With twolevel settings, however, it can be difficult to get convergence with a very unrestricted model (a model close to H1), mainly due to having many between-level parameters. That's why our UG gives an example of imputing from a simpler model than the later analysis model. How far apart these two models can be is an interesting research question.
FN briere posted on Saturday, March 31, 2012 - 11:39 am
Thanks, always very useful.
One last question which I think may also benefit others.
Given that my main interest is more in specific cross-level interactions than in random effects, another option may be to run an H1 imputation including cross-level interaction terms. I tried that and results from models estimated on imputed files (second step) look fine. I am thinking to go this way, as I did get some convergence problems with H0 models with random effects.
I don't see the distinction between cross-level interactions and random effects. Your model above would have a cross-level interaction if you regress the random slope s on Alevel2.
FN briere posted on Saturday, March 31, 2012 - 12:59 pm
Yes, I wasn't clear. I am wondering about different strategies to obtain something similar: defining a cross-level interaction term as a variable to be included in an H1 imputation v. specifying a random slope s to be regressed on Alevel2 in a H0 model. I get no convergence problems with the first strategy, but I do get some with the second. This is why I ask.
Dear Dr. Muthen i am doing multilevel SEM using multiple imputation and incorporating the complex survey design. The data comes from Add health and my sub-population for the complex survey analysis are students attending 9th to 11th grades in wave 1.
My outcome is a factor composed by GPA in math, reading, science and social studies. I performed multiple imputation to deal with missing data however, i still have some missing data because students did not take one or more of the four courses during the school year.
I am not sure how Mplus deals with this valid missing cases. I am considering out of my sub-population all cases that have missing in all four course, but i would like to keep cases that have taken at least one course. Would it be ok to have those cases just specifying the value for missing (for example, missing are -9999) or does mplus expect non-missing for all variables in each imputed data set? (in this case, how do i deal with valid missing cases)
It sounds like the above case presented by Fernando involves variables that have missing values because no values were imputed for those variable. Could we instead have variables that have values imputed for some cases but left missing for other cases?
The example I'm thinking of is a survey with a skip pattern such that there are items that only a subset of respondents answer. Let's say 60 out of 100 of respondents are presented with an item but only 50 answer them. I would want to impute for those 10 who were presented the item but elected not to respond. Is this possible? Perhaps through some kind of "if" statement associated with DATA IMPUTATION?
You can impute the data but not use the imputed values for the subgroup. Some kind of code in the analysis part of the imputed data should work (if 9999 is the missing value and you have both the imputed Yimp and the non-imputed Y in the same file):
define: if (group==1 .and. Y=9999) then Yimp=_missing;
This will restore the missing value where you need it.
Hope you are doing smoothly in fall semester. Can I ask you for your advice on how to impute cluster variable using multiple imputation(TYPE=COMPLEX)?
Here is my syntax:
VARIABLE:NAMES = econpr2 invpar latediv tr addep3 parrej delq bmi3 ill3 biosex4 hos smoking alcohol eat ill4d psu never ed; Missing are all(999); DATA IMPUTATION: IMPUTE = econpr2 invpar latediv (c) tr (c) addep3 parrej delq bmi3 ill3 biosex4 (c) hos smoking alcohol eat ill4d psu never (c) ed (c); NDATASETS = 10; SAVE = illimp*.dat;
ANALYSIS: TYPE = BASIC; OUTPUT: TECH8;
With this syntax, I was able to get perfect dataset without any missingness. However, finally, I found a problem because psu is cluster variable(school ID). With above syntax, program treat that variableas continuous variable. Although this variable has quite huge range(1-371), this is not a continuous variable. So, I am wondering whether there's certain way that I can treat this variable as cluster variable within the context of Multiple imputation.
Dear Dr. Muthén, I have several questions regarding MI for multilevel models.
1.) I want to include cluster means as predictors in the analysis. If I use a H1 imputation and do not specify (level 1) variables as within, then it is not necesarry to include cluster means (of the level 1 variables) on the between level? Is this correct?
2.) I have 3-level data (pupils, classes, schools). There is no missing data for class and school level variables. I want to estimate the effect of a 0/1 coded treatment variable at class level on performance. Additionally I'm interested in cross level interactions (effect of school level variables on the effect of treatment on performance).
2a.) I tried type = threelevel and estimator = bayes, but the model does not converge. Would it be adequate to use type = basic twolevel (cluster = class) in order to achieve convergence? (there are no problems for type = basic and type = basic twolevel).
2b.) Due to convergence problems it seems also not possible to include the crosslevel interactions. Would it be adequate for H1-imputation to include simple product terms (VariableLev3xVariableLev2, ...)?
I have tried to run a complex twolevel model with a cross-level interaction based on 10 imputed datasets. The model without the cross-level interaction runs fine. However, when including the interaction, it seems as if the data is not used, and hence I don’t get any results (both the fit indices and the estimates are zero). The output also says: Number of replications Requested 10 Completed 0
My input file reads as follows: DATA: FILE IS TotalFile19t.4.bmsrclist.dat; type = imputation; CLUSTER = TherID cohort; MISSING ARE ALL (999); WITHIN = zgender Language therexp athome inschool noarrest; BETWEEN = team1-team26 teamyexp; CATEGORICAL = athome inschool noarrest (0-1);
Thank you for your reply. I have analysed a single data set and received the error message "THE ESTIMATED BETWEEN COVARIANCE MATRIX COULD NOT BE INVERTED." When looking into the manual I saw that centering was used in cross-level interactions. Therefore, I have centered therexp using grandmean. 1) The analyses on every single datafile now runned without errors. However, when using the 10 imputed datasets together, still only 4 of the replications were completed.
2) Also, in the output, I see the estimates of all paths (my model includes more paths than mentioned above), including the between-level path s on teamyexp. However I don’t see the estimates for the path famav19 on therexp. I was wondering whether this is correct or whether it is an error in my model. Is there any way I could get the estimates for this specific path?
3) In the output with the model estimates I also have a column ‘Rate of Missing’. What does this column mean? Since I have imputed all my data, I wouldn't expect any missings (I also get this column in the analysis without the cross-level interaction, which is running without problems).
I hope you could provide me with some suggestions to move on.