Original Message:
Sent: Mon July 10, 2023 11:27 AM
From: David Dwyer
Subject: How does the /MISSING Subcommand work in MIXED Procedure?
Hi @Timo Schurr
I'll have to let our more statistically inclined colleagues address your question of whether the estimates should change. But here are my thoughts:
1) If all you had was valid values and System-Missing, then /MISSING=INCLUDE and /MISSING=EXCLUDE would yield the exact same result. The /MISSING subcommand applies to User-Missing values only. Cases with System-Missing values are already excluded, always.
2) If you exchanged all of your System-Missing observation for the same single User-Missing value, then I would expect /MISSING=EXCLUDE to yield the same results as if you left it all System-Missing. I'd posit that since you exchanged the single internal System-Missing value for a single User-Missing value, you aren't adding any new variability to the model overall. Thus I guess I would not be surprised the estimates did not change.
3) By the same token, if you went variable by variable exchanging the internal System-Missing value for one that was meaningful within the range of each variable, then yes, I'd expect the estimates to change when you specify /MISSING=INCLUDE.
You've also asked "How exactly does the estimation of missing values work within the /MIXED command? " . As far as I know the MIXED command does not do any estimation of missing values. If you are interested in actually replacing the System-Missing value with the series mean, some regression estimate, some maximum likelihood estimate, etc. please see the Missing Values Analysis option in IBM SPSS Statistics Professional and IBM SPSS Statistics Premium editions.
I hope this helps!
------------------------------
David Dwyer
SPSS Technical Support
IBM Software
Original Message:
Sent: Mon July 10, 2023 06:26 AM
From: Timo Schurr
Subject: How does the /MISSING Subcommand work in MIXED Procedure?
Hi @David Dwyer
Thank you for your answer!
You are right.
I have replaced all 114 system-missing values with user-missing values, because I wanted to try out to what extend the estimates change by including the user-missing values.
I am still a bit confused, what the /MISSING command exactly does, since the estimates do not change at all, although I have changed all system-missing to user-missing values?
Shouldn't the estimates be different?
How exactly does the estimation of missing values work within the /MIXED command?
By excluding the missing values in the dependent and independent variable the way it right now does, the /MIXED model is not different from a repeated measures ANOVA?
Thank you very much!
------------------------------
Timo Schurr
Original Message:
Sent: Fri July 07, 2023 01:46 PM
From: David Dwyer
Subject: How does the /MISSING Subcommand work in MIXED Procedure?
Hi @Timo Schurr
Here the the description of the /MISSING subcommand From the SPSS Statistics Syntax Reference Guide.
Is it possible that you have user-missing values defined for some variables, while at the same time having system-missing values on others? System-missing variables are always excluded. So I'm hypothesizing that the same cases are being excluded because each has at leas one system-missing value across the variables in your analysis. Thus, /MISSING=INCLUDE or /MISSING=EXCLUDE wouldn't make a difference. The /MISSING subcommand is differentiating user-missing from system-missing.
I hope this helps
------------------------------
David Dwyer
SPSS Technical Support
IBM Software
Original Message:
Sent: Thu July 06, 2023 03:28 AM
From: Timo Schurr
Subject: How does the /MISSING Subcommand work in MIXED Procedure?
Hey everyone,
I'm currently having trouble figuring out what exactly the /MISSING subcommand of the MIXED procedure does.
My data set consists of 820 participants, with a total of three measurements:
At baseline, 650 people participated
Follow-up one 600 people participated
Follow-up two 550 people participated
This results in a total of 1800 observations.
I have a dependent variable and an independent variable. The dependent variable contains 81 missing entries, the independent variable contains 99 missing entries, and when both are combined, there are 114 missing entries (because one part overlaps). So there are 1686 complete cases.
If I now execute the following syntax:
MIXED Distress
/CRITERIA=DFMETHOD(SATTERTHWAITE) CIN(95) MXITER(1000) MXSTEP(10) SCORING(1)
SINGULAR(0.000000000001) HCONVERGE(0.00000001, RELATIVE) LCONVERGE(0, ABSOLUTE) PCONVERGE(0,
ABSOLUTE)
/FIXED=| SSTYPE(3)
/METHOD=REML
/PRINT=CPS DESCRIPTIVES G R SOLUTION TESTCOV
/RANDOM=INTERCEPT | SUBJECT(id) COVTYPE(VC).
SPSS tells me that those 114 cases have been excluded.
If I'm adding now the /MISSING=INCLUDE subcommand, the 114 cases are included.
In both cases, the output of the estimated parameters is exactly the same.
1.What exactly does the /MISSING command do?
2. I always assumed that missing values are estimated by ML / REML (which is the advantage of linear mixed models over repeated measures ANOVA). Shouldn't the parameter estimates be different?
Thanks a lot!
------------------------------
Timo Schurr
------------------------------