stata neuling möchte fehlende werte ersetzten

[Phillip Thies]

as part of a research project, we would like to determine predictors for the outcome of a particular form of therapy. To this end, we collected pre- and post-data using various PROM questionnaires. however, some of these questionnaires contain missing values because not all questions were answered. Now I would like to replace these missing values. I would like to use either FIML or MI, as I think these are the most useful methods. however, my experience with stata is very limited (I have only been working with it since the start of the project). Both are difficult for me. The theoretical background underlying both calculations is mostly clear to me. However, I don't understand why I still seem to have missing data after FIML - this shouldn't really be the case, should it?
In the following code, I compare a sum score before the FIML with a score after the FIML at the end (Table at the end). The latter should actually take on a different value by replacing the missing values. what am I doing wrong here - or is my train of thought not correct? I hope you understand what I mean

Code:
* Apply FIML to handle missing values in fess_1_pre - fess_10_pre
sem (fess_1_pre fess_2_pre fess_3_pre fess_4_pre fess_5_pre ///
fess_6_pre fess_7_pre fess_8_pre fess_9_pre fess_10_pre), method(mlmv)

* Generate new variable fess_score_pre_new as the sum of non-zero items
egen fess_score_pre_new = rowtotal(fess_1_pre fess_2_pre fess_3_pre fess_4_pre ///
fess_5_pre fess_6_pre fess_7_pre fess_8_pre ///
fess_9_pre fess_10_pre) if fess_1_pre != 0 & ///
fess_2_pre != 0 & fess_3_pre != 0 & fess_4_pre != 0 & ///
fess_5_pre != 0 & fess_6_pre != 0 & fess_7_pre != 0 & ///
fess_8_pre != 0 & fess_9_pre != 0 & fess_10_pre != 0

* Documentation:
* - The sem command applies FIML to handle missing values in the specified variables.
* - The egen function with rowtotal calculates the sum of the non-zero items for each observation/id.
* - The if condition excludes any observations where any of the fess_1_pre to fess_10_pre items are 0.
* - The new variable fess_score_pre_new now contains the recalculated sum score, excluding 0 values.
* - The existing fess_score_pre variable is preserved.


* Create a table comparing the two variables
tabstat fess_score_pre fess_score_pre_new, ///
statistics(n mean sd min max) ///
columns(statistics) ///
longstub

Output:
* Using Full-Information Maximum Likelihood (FIML) in Stata to handle missing
> values
. * Ensure missing values are coded as "." for FIML to work correctly
.
. * Apply FIML to handle missing values in fess_1_pre - fess_10_pre
. sem (fess_1_pre fess_2_pre fess_3_pre fess_4_pre fess_5_pre ///
> fess_6_pre fess_7_pre fess_8_pre fess_9_pre fess_10_pre), method(mlmv)
(1 all-missing observation excluded)

Exogenous variables
Observed: fess_1_pre fess_2_pre fess_3_pre fess_4_pre fess_5_pre fess_6_pre
fess_7_pre fess_8_pre fess_9_pre fess_10_pre

Fitting saturated model:
Iteration 0: log likelihood = -1987.5859
Iteration 1: log likelihood = -1986.2606
Iteration 2: log likelihood = -1986.2261
Iteration 3: log likelihood = -1986.2261

Fitting baseline model:
Iteration 0: log likelihood = -2341.374
Iteration 1: log likelihood = -2341.3686
Iteration 2: log likelihood = -2341.3686

Fitting target model:
Iteration 0: log likelihood = -1986.2261
Iteration 1: log likelihood = -1986.2261

Structural equation model Number of obs = 142
Estimation method: mlmv

Log likelihood = -1986.2261

-------------------------------------------------------------------------------
| OIM
| Coefficient std. err. z P>|z| [95% conf. interval]
--------------+----------------------------------------------------------------
mean(fe~1_pre)| 3.753521 .1059557 35.43 0.000 3.545852 3.96119
mean(fess_2~e)| 4.297576 .0997105 43.10 0.000 4.102147 4.493005
mean(fess_3~e)| 3.59276 .116653 30.80 0.000 3.364125 3.821396
mean(fess_4~e)| 3.818664 .1002849 38.08 0.000 3.62211 4.015219
mean(fess_5~e)| 4.139852 .105352 39.30 0.000 3.933366 4.346338
mean(fess_6~e)| 3.415493 .1026542 33.27 0.000 3.214294 3.616691
mean(fess_7~e)| 3.570423 .1307653 27.30 0.000 3.314127 3.826718
mean(fess_8~e)| 3.514085 .1009457 34.81 0.000 3.316235 3.711934
mean(fess_9~e)| 3.626761 .1070836 33.87 0.000 3.416881 3.83664
mean(fe~0_pre)| 3.690141 .1011652 36.48 0.000 3.491861 3.888421
--------------+----------------------------------------------------------------
var(fess_1_~e)| 1.594178 .1891941 1.263333 2.011666
var(fess_2_~e)| 1.389289 .1664312 1.098556 1.756965
var(fess_3_~e)| 1.925182 .2288001 1.525141 2.430153
var(fess_4_~e)| 1.42291 .1690777 1.127283 1.796063
var(fess_5_~e)| 1.563497 .1864577 1.237615 1.975189
var(fess_6_~e)| 1.49638 .1775876 1.185831 1.888256
var(fess_7_~e)| 2.428139 .2881671 1.924219 3.064027
var(fess_8_~e)| 1.446985 .1717255 1.146687 1.825925
var(fess_9_~e)| 1.628298 .1932434 1.290372 2.054721
var(fess_10~e)| 1.453283 .172473 1.151678 1.833873
--------------+----------------------------------------------------------------
cov(fess_1_~e,|
fess_2_pre)| .6959802 .1384627 5.03 0.000 .4245983 .9673621
cov(fess_1_~e,|
fess_3_pre)| 1.025174 .1704203 6.02 0.000 .6911559 1.359191
cov(fess_1_~e,|
fess_4_pre)| .6929782 .1392014 4.98 0.000 .4201485 .965808
cov(fess_1_~e,|
fess_5_pre)| .6199708 .1425125 4.35 0.000 .3406515 .8992901
cov(fess_1_~e,|
fess_6_pre)| .6587483 .1409086 4.68 0.000 .3825724 .9349241
cov(fess_1_~e,|
fess_7_pre)| .4082027 .1686215 2.42 0.015 .0777107 .7386948
cov(fess_1_~e,|
fess_8_pre)| .7464293 .1420154 5.26 0.000 .4680843 1.024774
cov(fess_1_~e,|
fess_9_pre)| .8657508 .1534881 5.64 0.000 .5649197 1.166582
cov(fess_1_~e,|
fess_10_pre)| .5151756 .1348499 3.82 0.000 .2508747 .7794765
cov(fess_2_~e,|
fess_3_pre)| .707832 .150346 4.71 0.000 .4131592 1.002505
cov(fess_2_~e,|
fess_4_pre)| .7702579 .1356952 5.68 0.000 .5043002 1.036216
cov(fess_2_~e,|
fess_5_pre)| .8982935 .1456122 6.17 0.000 .6128987 1.183688
cov(fess_2_~e,|
fess_6_pre)| .7120622 .1356932 5.25 0.000 .4461084 .9780161
cov(fess_2_~e,|
fess_7_pre)| .5947767 .1634216 3.64 0.000 .2744762 .9150772
cov(fess_2_~e,|
fess_8_pre)| .6619475 .1327698 4.99 0.000 .4017235 .9221715
cov(fess_2_~e,|
fess_9_pre)| .7167607 .1410359 5.08 0.000 .4403354 .9931859
cov(fess_2_~e,|
fess_10_pre)| .473951 .1267695 3.74 0.000 .2254874 .7224147
cov(fess_3_~e,|
fess_4_pre)| .8986428 .1588445 5.66 0.000 .5873133 1.209972
cov(fess_3_~e,|
fess_5_pre)| .7904607 .1602389 4.93 0.000 .4763982 1.104523
cov(fess_3_~e,|
fess_6_pre)| .9649799 .1638709 5.89 0.000 .6437988 1.286161
cov(fess_3_~e,|
fess_7_pre)| .5669175 .1883615 3.01 0.003 .1977358 .9360992
cov(fess_3_~e,|
fess_8_pre)| .9383406 .1616449 5.80 0.000 .6215223 1.255159
cov(fess_3_~e,|
fess_9_pre)| 1.103945 .1759019 6.28 0.000 .7591836 1.448707
cov(fess_3_~e,|
fess_10_pre)| .545249 .1485431 3.67 0.000 .2541098 .8363882
cov(fess_4_~e,|
fess_5_pre)| .7329144 .1404756 5.22 0.000 .4575872 1.008242
cov(fess_4_~e,|
fess_6_pre)| .7285103 .1371322 5.31 0.000 .4597362 .9972844
cov(fess_4_~e,|
fess_7_pre)| .7971047 .1699422 4.69 0.000 .4640242 1.130185
cov(fess_4_~e,|
fess_8_pre)| .8854802 .1417811 6.25 0.000 .6075943 1.163366
cov(fess_4_~e,|
fess_9_pre)| .9516822 .1506946 6.32 0.000 .6563261 1.247038
cov(fess_4_~e,|
fess_10_pre)| .6128256 .1311902 4.67 0.000 .3556975 .8699537
cov(fess_5_~e,|
fess_6_pre)| .8925968 .1486674 6.00 0.000 .601214 1.18398
cov(fess_5_~e,|
fess_7_pre)| .6602347 .1735039 3.81 0.000 .3201734 1.000296
cov(fess_5_~e,|
fess_8_pre)| .6551925 .138805 4.72 0.000 .3831397 .9272453
cov(fess_5_~e,|
fess_9_pre)| .828412 .1518199 5.46 0.000 .5308505 1.125974
cov(fess_5_~e,|
fess_10_pre)| .6294073 .1380864 4.56 0.000 .3587629 .9000518
cov(fess_6_~e,|
fess_7_pre)| .5517258 .1665266 3.31 0.001 .2253396 .8781121
cov(fess_6_~e,|
fess_8_pre)| .9131621 .1453288 6.28 0.000 .6283229 1.198001
cov(fess_6_~e,|
fess_9_pre)| .922684 .1521651 6.06 0.000 .6244459 1.220922
cov(fess_6_~e,|
fess_10_pre)| .6005753 .1336211 4.49 0.000 .3386827 .8624679
cov(fess_7_~e,|
fess_8_pre)| .8616842 .1731233 4.98 0.000 .5223687 1.201
cov(fess_7_~e,|
fess_9_pre)| .7269887 .1776657 4.09 0.000 .3787702 1.075207
cov(fess_7_~e,|
fess_10_pre)| .6415394 .1665801 3.85 0.000 .3150484 .9680303
cov(fess_8_~e,|
fess_9_pre)| 1.121454 .1595279 7.03 0.000 .808785 1.434123
cov(fess_8_~e,|
fess_10_pre)| .7578853 .1373099 5.52 0.000 .4887629 1.027008
cov(fess_9_~e,|
fess_10_pre)| .792799 .1452271 5.46 0.000 .5081592 1.077439
-------------------------------------------------------------------------------
LR test of model vs. saturated: chi2(0) = 0.00 Prob > chi2 = .

.
. * Generate new variable fess_score_pre_new as the sum of non-zero items
. egen fess_score_pre_new = rowtotal(fess_1_pre fess_2_pre fess_3_pre fess_4_pre
> ///
> fess_5_pre fess_6_pre fess_7_pre fess_8_pre
> ///
> fess_9_pre fess_10_pre) if fess_1_pre != 0
> & ///
> fess_2_pre != 0 & fess_3_pre != 0 & fess_4_
> pre != 0 & ///
> fess_5_pre != 0 & fess_6_pre != 0 & fess_7_
> pre != 0 & ///
> fess_8_pre != 0 & fess_9_pre != 0 & fess_10
> _pre != 0

.
. * Documentation:
. * - The sem command applies FIML to handle missing values in the specified var
> iables.
. * - The egen function with rowtotal calculates the sum of the non-zero items f
> or each observation/id.
. * - The if condition excludes any observations where any of the fess_1_pre to
> fess_10_pre items are 0.
. * - The new variable fess_score_pre_new now contains the recalculated sum scor
> e, excluding 0 values.
. * - The existing fess_score_pre variable is preserved.
.
.
. * Create a table comparing the two variables
. tabstat fess_score_pre fess_score_pre_new, ///
> statistics(n mean sd min max) ///
> columns(statistics) ///
> longstub

Variable | N Mean SD Min Max
-------------+--------------------------------------------------
fess_score~e | 143 36.91608 9.619886 0 58
fess_score~w | 143 36.91608 9.619886 0 58
----------------------------------------------------------------

.
end of do-file

[Maarten Buis]

FIML does not replace missing values. It just uses all the available information, including from observations with partially missing information, in the estimation. So this is something you do in the final model. This is also the logic behind MI: MI does not recover the missing values. They are assumed to be lost forever. The imputation is just a computational trick to average out those missing values, so the information in the observed values can be used.

[Phillip Thies]

I appreciate your answer. I thought that (regardless of whether it is FIML or MI) missing values in the dataset are replaced by multiple sets of simulated values based on the observed data and a defined imputation model, or that FIML incorporates all available information from the observed data and the pattern of missing data directly into the estimation process. I could then use these values for further calculations, right?

[Maarten Buis]

That depends (which is short for "in theory yes, in practice probably not")

[Phillip Thies]

But what could be an alternative approach then?

[Maarten Buis]

That depends on what you exactly want to do and why.

[Phillip Thies]

We’d like to determine predictors for the outcome of a particular form of therapy using a regression model. Some of the potential predictor variables have missing data, that’s why i wanted to perform fiml or mi before the building the regression model.

[Maarten Buis]

So, you can do that regression with sem and use fiml