Imputation files not found

[amarkovitz]

Hello,

I ran what appeared to be a successful imputation procedure, but when I type realcomImputeLoad in stata, I am getting the following error:

file imputeiter100. not found

If it's helpful to diagnose this error, here is my model in stata:
realcomImpute bmi bmixsplinea1 bmixsplinea2 bmixsplinea3 bmixsplinea4 bmixsplinea5 bmiat18 ///
splinea1 splinea2 splinea3 splinea4 splinea5 parous HDL ///
parousxsplinea1 parousxsplinea2 parousxsplinea3 parousxsplinea4 parousxsplinea5 ///
post_preg_ind post_time h3 ses ///
using TestMIInput.dat, replace numresponses(7) ///
level2id(pid) cons(cons)

When I ran the MCMC, I obtained the attached log and it saved the attached impvals.txt file. It also created a storedata.mat file which as far as I can tell doesn't contain anything in it. Was there some issue in creating the imputed datasets? How were they supposed to be stored? As far as I can tell, the documentation doesn't get into this level of detail.

Thank you,
Amanda

[ChrisCharlton]

When you run the imputation Realcom-Impute should save a file containing the imputed responses for each chosen imputation. The names of these files should correspond with the text in the fourth line of the impvals.txt file and should be in the same directory as this file. The storedata.mat file should contain the state of the software when you last used it, in theory allowing you to carry on from that point at a later date, this is not however used by external software. If the imputeiter* files are missing then it may that these was an estimation problem. The output that you posted seems to contain a lot of zeros, so it might be worth looking at the earlier output to see whether there were any further messages from when it was running the MCMC chains. It may also be worth checking the MCMC trajectories to ensure that they look okay.

[amarkovitz]

Hello Chris,

Thank you for the fast reply. It does seem like there might have been estimation problems, although I did not see anything in the log to indicate this. When I click "view results" I do get error messages (attached). I did monitor one of the MCMC trajectories while it was running, but could not find anywhere in the documentation that explained what to look for in these trajectories. I've attached an example of one. Is there other documentation I could use to learn more about how to evaluate these trajectories?

I've tried many different versions of this model with different predictor variables and haven't yet been able to sucessfully run it without some type of error. Are there any tips for getting the model to fit? I'm predicting a level 2 variable (BMI at age 18) and I have values for BMI at other ages, as well as other variables that will end up in my final model. There are some missing values for BMI at other ages, although not many, so I included them as response variables.

Thank you,
Amanda

[ChrisCharlton]

It looks like something definitely is going wrong there, you shouldn't get any error messages while estimating and the value of the beta you are monitoring is staying extremely close to zero. I would suggest that you initially check the files that have been output from Stata for use in Realcom-Impute to ensure that they match what you expect from your data. I would also suggest checking that the value used to correspond to "missing" within these files matches that specified within Realcom-Impute before you start the iterations. For general advice on MCMC and interpreting the chains see the first chapter of the MLwiN MCMC guide (http://www.bris.ac.uk/cmm/media/softwar ... mc-web.pdf)

[amarkovitz]

Hello Chris,

Thank you for that information, that will be very helpful. In an effort to try to figure out what might be going wrong, I tried to simplify the problem. I ran MI again with just two variables, BMI at age 18 which is missing for about 44% of people and current BMI which I made sure was not missing for anyone. I am now getting an immediate error upon running the MCMC despite confirming that the missing value was set to 9999 in the dataset and in REALCOM. I also tried it with a few other non-missing predictors in the model as well. I feel like there must be something very basic that I'm missing, any suggestions are greatly appreciated!

Here is my stata code:
sort pid

replace bmiat18=9999 if missing(bmiat18)

drop if bmi==.

realcomImpute bmiat18 bmi ///
using TestMIInput2.dat, replace numresponses(1) ///
level2id(pid) cons(cons)

[ChrisCharlton]

The interface for setting the missing value can sometimes be confusing. Can you confirm that you clicked the button labelled "set missing value to:" after typing in the new value, as I think that the error message you are getting is suggesting that it isn't seeing any missing data?

[amarkovitz]

Hello Chris,

Unfortunately, I did try that and received the same error. I've also tried a few different missing values (ex. STATA's default -9.999e+29, 9999, and 99999) and also have tried running the model with different response and auxiliary variables to see if was something wrong with the variables I'm using, but am having the same problem. Is there anything else you would suggest I try?

Thank you,
Amanda

[ChrisCharlton]

If you have access to MLwiN you could try working through the example in the Realcom-Impute manual (http://www.bris.ac.uk/media-library/sit ... tation.pdf) to see whether you get the same problems there. If not then if you are able to replicate the problem using a publicly available dataset then we can look into it further and see whether there is an undiscovered bug in the software.

[amarkovitz]

Hello Chris,

I tried to use a publicly available dataset and I did not receive the error about missing data, but it is having the estimation problem that I initially wrote about:
. realcomImputeLoad
file impvals.txt not found

Perhaps it just isn't a very good example and if so please let me know an example that should run (I tried having less missing data and it was still a problem). Here is the STATA code I used:

/*Use a publicly available multilevel STATA dataset*/
use http://www.stata-press.com/data/r12/productivity.dta

/*Create a level-2 (state-level) variable with missing data*/
/*Mean unemployment by state*/
by state, sort: egen meanunemp = mean(unemp)

/*Missing at random conditional on "public" variable*/
replace meanunemp=. if public>50000

/*Create dataset for REALCOM-Impute*/
gen cons = 1
sort state
cd "H:\HUNT\Datasets_intermediate" //Same folder that REALCOM-Impute starts in
realcomImpute meanunemp public gsp ///
using TestMIInput.dat, replace numresponses(1) ///
level2id(state) cons(cons)

realcomImputeLoad

Thank you,
Amanda

[ChrisCharlton]

I just tried replicating your example and it seemed to work for me:

Code to set up imputation in Stata:

Code: Select all

. /*Use a publicly available multilevel STATA dataset*/
. use http://www.stata-press.com/data/r12/productivity.dta
(Public Capital Productivity)

. 
. /*Create a level-2 (state-level) variable with missing data*/
. /*Mean unemployment by state*/
. by state, sort: egen meanunemp = mean(unemp)

. 
. /*Missing at random conditional on "public" variable*/
. replace meanunemp=. if public>50000
(113 real changes made, 113 to missing)

. 
. /*Create dataset for REALCOM-Impute*/
. gen cons = 1

. sort state

. cd "O:\Realcom-stata" //Same folder that REALCOM-Impute starts in
O:\Realcom-stata

. realcomImpute meanunemp public gsp ///
> using TestMIInput.dat, replace numresponses(1) ///
> level2id(state) cons(cons)
1 responses: meanunemp
2 auxiliary variables: public gsp
(note: file TestMIInput.dat not found)
Data saved to TestMIInput.dat
(note: file TestMIInput_wts.dat not found)

Realcom-Impute output:

Code: Select all

User selected O:\Realcom-stata\TestMIInput.dat

c_nimpute =

     1


c_naux =

     4


c_vartype =

     1


ans =

     2


c_betaconstraint =

     0
     1


fid =

     3


Imputation_method =

     0


Multivariate estimation using Gibbs  and MH sampling - Normal/probit models

Independence assumed with standard probit model for unordered categories

 Nov 14 2006

Two-level responses version

NRTYPE =

     1


BETASTART =

     0
     0
     0


BETASTART_2 =

   Empty matrix: 1-by-0


SIGMASTART =

    1.3530


 Burn in =    100        Sample =   1000

       no level 1, level 2 units =    816,      48

iteration     10
iteration     20
iteration     30
iteration     40
iteration     50
iteration     60
iteration     70
iteration     80
iteration     90
iteration    100
iteration    110
iteration    120
iteration    130
iteration    140
iteration    150
iteration    160
iteration    170
iteration    180
iteration    190
iteration    200
iteration    210
iteration    220
iteration    230
iteration    240
iteration    250
iteration    260
iteration    270
iteration    280
iteration    290
iteration    300
iteration    310
iteration    320
iteration    330
iteration    340
iteration    350
iteration    360
iteration    370
iteration    380
iteration    390
iteration    400
iteration    410
iteration    420
iteration    430
iteration    440
iteration    450
iteration    460
iteration    470
iteration    480
iteration    490
iteration    500
iteration    510
iteration    520
iteration    530
iteration    540
iteration    550
iteration    560
iteration    570
iteration    580
iteration    590
iteration    600
iteration    610
iteration    620
iteration    630
iteration    640
iteration    650
iteration    660
iteration    670
iteration    680
iteration    690
iteration    700
iteration    710
iteration    720
iteration    730
iteration    740
iteration    750
iteration    760
iteration    770
iteration    780
iteration    790
iteration    800
iteration    810
iteration    820
iteration    830
iteration    840
iteration    850
iteration    860
iteration    870
iteration    880
iteration    890
iteration    900
iteration    910
iteration    920
iteration    930
iteration    940
iteration    950
iteration    960
iteration    970
iteration    980
iteration    990
iteration   1000
iteration   1010
iteration   1020
iteration   1030
iteration   1040
iteration   1050
iteration   1060
iteration   1070
iteration   1080
iteration   1090
iteration   1100
Error using save
Unable to write file storedata: permission denied.

Error in mcmcdriver (line 1192)



Error in gui_2levmixresp>pushbutton16_RunMCMC_Callback (line 264)



Error in gui_mainfcn (line 96)



Error in gui_2levmixresp (line 46)



Error in @(hObject,eventdata)gui_2levmixresp('pushbutton16_RunMCMC_Callback',hObject,eventdata,guidata(hObject))



Error while evaluating uicontrol Callback

Code to load imputed data into Stata:

Code: Select all

. realcomImputeLoad

10 multiple imputations successfully loaded

  Style:  wide
          last mi update 31aug2016 16:19:12, 0 seconds ago

  Obs.:   complete          703
          incomplete        113  (M = 10 imputations)
          ---------------------
          total             816

  Vars.:  imputed:  1; meanunemp(113)

          passive:  0

          regular:  0

          system:   1; _mi_miss

         (there are 12 unregistered variables)

Can you confirm that you are setting the output directory (via the Impute button in Realcom) to be the same as Stata is using?