Problems with new MLWiN version

[michaellawton]

Hey runmlwin

I had a runmlwin model that was working back in OCt 2012. I am currently writing up a paper and when I try to run the model with the same command I get the following message

"V gone -ve definite for block 0 iteration aborted(fixed pass)"

any idea what is going on here?

Best wishes
Michael Lawton

[GeorgeLeckie]

Hi Michael,

Would you be able to paste your runmlwin syntax and output so that we can try to see what the problem might be?

Best wishes

George

[michaellawton]

My syntax is below (it is only the final model that has problems with the message)

Code: Select all

matrix a= (1,1,0)
. runmlwin score cons fp1 fp6 , level2 (patid: cons  ) level1(newids: cons ) nopause maxiterations(1250)  tolerance(1)
. runmlwin score cons fp1 fp6 , level2 (patid: cons fp1) level1(newids: cons ) nopause maxiterations(2550) initsprevious tolerance(1)
. runmlwin score cons fp1 fp6 , level2 (patid: cons fp1 fp6 ) level1(newids: cons fp1, elements(a)) nopause maxiterations(2550) initsprevious  tolerance(1)

Also below is the output from the final model which I last ran in October 2012

Code: Select all

Warning: getversion plugin could not be loaded
MLwiN ... multilevel model                      Number of obs      =      1589
Normal response model
Estimation algorithm: IGLS

-----------------------------------------------------------
                |   No. of       Observations per Group
 Level Variable |   Groups    Minimum    Average    Maximum
----------------+------------------------------------------
          patid |      359          1        4.4         25
-----------------------------------------------------------

Run time (seconds)   =       1.50
Number of iterations =         39
Log likelihood       =  -2126.533
Deviance             =  4253.0659
------------------------------------------------------------------------------
       score |      Coef.   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
        cons |   2.634763   .3224213     8.17   0.000     2.002829    3.266697
         fp1 |   .1588309   .0315073     5.04   0.000     .0970777    .2205841
         fp6 |   -.149601   .2813045    -0.53   0.595    -.7009476    .4017456
------------------------------------------------------------------------------

------------------------------------------------------------------------------
   Random-effects Parameters |   Estimate   Std. Err.     [95% Conf. Interval]
-----------------------------+------------------------------------------------
Level 2: patid               |
                   var(cons) |   8.670482   1.849217      5.046084    12.29488
               cov(cons,fp1) |   .0859842   .1618544     -.2312446    .4032129
                    var(fp1) |   .0820835   .0186486      .0455328    .1186342
               cov(cons,fp6) |  -5.380843   1.625468     -8.566702   -2.194983
                cov(fp1,fp6) |  -.6000037   .1650493     -.9234944    -.276513
                    var(fp6) |   7.136051   1.596572      4.006828    10.26527
-----------------------------+------------------------------------------------
Level 1: newids              |
                   var(cons) |   .4020671   .0255776       .351936    .4521982
               cov(cons,fp1) |  -.0032306   .0006861     -.0045754   -.0018858
------------------------------------------------------------------------------

[GeorgeLeckie]

HI Michael,

Thanks for posting the commands and output

You are trying to fit a random-intercept and two random coefficients, but you only have ~4 measurements per individual, and less than 4 in many cases. This is quite demanding.

I am concerned with your specification of the maxiterations(1250) option, this is an exceedingly high number of iterations to specfiy for any MLwiN model. Rarely do you need more than 50 iterations, particularly when the response variable is continuous.

I am also concerned with your specification of the tolerance(1) option, this is a very coarse tolerance. The model will stop iterating once all parameters change by less than 1 point moving. The default if 0.001 which works fine for most models.

A general thing you can do when fitting complex models is to build up to them gradually using the parameter estimates from the previous model as starting values for the next model. Make use to the initsprevious option to do this. You might want to split your Model 3 into Model 3a and Model 3b where in Model 3a you introduce the random coefficient on fp6 and in Model 3b you enter fp1 into the random part of the model at level-1

Note, make sure to always specify the corr option to check that all your correlations lie between -1 to 1.

It is hard to say more without looking at the data. See how you get on. If the above doesn't help you work out the problem then email me the data and dofile offline and I will take a look.

Best wishes

George