To make sure this is not a mistake over coding - fit a single -level model in MLwin- do this by simply ticking off  or removing the level 2 variance term

There is a lot of help on logistic models here

http://www.bristol.ac.uk/cmm/software/mlwin/mlwin-resources.html#discrete

Check you are using the same reference code for the categorical predictors in MLwin as in R

The customised predictions tool in MLwiN will allow you to change the reference category without re-running the model

see https://www.researchgate.net/publication/234015221_Manual_supplement_for_MLwiN_Version_2.14

Data Manual supplement for MLwiN Version 2.14

Dear professor Jones,

Thank you very much for your help, I will try to read the manual supplement. I fitted a single level model, even changed the reference category but in any case the sign is still minus (which does not make sense because if the reference group changes then the sign should change as well). I am using 2.31 version which has always worked very well. 

Thank you for all your help!

Do you get the same sign flip problem going from glm in R to lme4's glmer in R? Is the outcome variable actual zeroes and ones? I have seen these sign switches occur when a program reads a binary variable as characters and recodes as zeroes and ones but the other way than expected.

Let us know what happens.

The current release is version 2.35; see

http://www.bristol.ac.uk/cmm/software/mlwin/

If you are only using higher level variables - that is about Schools, you can collapse the data to proportions (number of passes divide by  number of passes + number of fails) and fit a Binomial Logit model and not a Bernoulli with groups of individuals (cells)  at level 1 and schools at level 2 - there will be no loss of information just quicker to estimate and you will get the same results

see https://www.researchgate.net/publication/260772180_Developing_multilevel_models_for_analysing_contextuality_heterogeneity_and_change_using_MLwiN_Volume_2

and

https://www.researchgate.net/publication/23539033_Multilevel_perspectives_on_modeling_census_data

Book Developing multilevel models for analysing contextuality, he...

Article Multilevel Perspectives on Modeling Census Data

Is this a logistic or ordered logistic regression. Sign changes can can arise in ordered logistic regression from changes in parameterization of the model (as well as changed in reference category).

Thank you very much for your answers. 

Professor Jones, I will make proportions and re run the model. I haven't even able to do it yet because in my job they assigned my something else to do. But probably tomorrow I will tackle this issue. Thank you very much for your answer.

On the other hand, the outcome is actual zeros and ones and in this case the ordering does not apply I just have two groups 0 is fail and 1 is pass.  Thank you very much for your ideas and advice.  

When I re run the model i will let you know what happens. Thanks a lot!

Dear all,

Thank you very much for all your help. I managed to run my model using a Binomial logit model. It worked very well. 

I was trying to get the model fit indices (DIC in this case because MLwiN uses the quasi likelihood algorithm). Eventhough I get the number my model does not converge (all numbers are in blue). Can I still follow that DIC numbers to determine the best fitting model or I cannot use them? I tried with 5000 and 10000 iterations using MCMC and it does not converge.

Thank you very much for all your help!

Juan, it is a while since I have used MLwiN but convergence in MCMC estimation in any package is different from ML and related methods. Essentially you need to look at the diagnostics to determine whether the chains have converged. It is quite possible to get models that give good estimates of some parameters but not others (e.g., variances close to zero). Unless the model is very simple and there is plenty of data I don't think that 5000 or 10000 iterations is likely to produce good estimates. I've run models with 500,000 or more iterations to get healthy diagnostics before (and have colleagues who have data sets that require greater numbers).

The only guarantee with MCMC is that you'll eventually get good estimates if you wait long enough.However, there are tricks that can improve efficiency.

It would also be sensible to run the model in glmer() as you've got the data in R already.  You should get similar parameter estimates in both cases. MLwiN has some advantages (such as giving you DIC).

I will add to Thom's useful posting

1) The MLwiN estimates will never turn green but will always remain blue - as they are converging towards a distribution not a point

2) MlwiN has useful diagnostics where you can examine the trajectories of the estimates for convergence - ie no trend - and whether they have been run long enough once they have reached this equilibrium distribution,

3) Bill Brown has done a lot of work to speeding up the MCMC procedures and making these estimates less correlated - you have to request these as they are not the defaults - Bill explains them in detail in the MLwiN MCMC manual and these are additions to the original manual;  readers may not realize that these procedures can be applied to the models and methods discussed in earlier chapters. They often make a considerable difference

In relation to points 2 and 3 - the MCMC Chapter in volume 1 of my training manual  explains for a very simple two-level model how to judge convergence and how to speed things up

If you are in r and have MLwiN on your machine you can make use of MLwiN's fast algorithms through R2MLwiN

http://www.bristol.ac.uk/cmm/software/r2mlwin/

https://www.researchgate.net/publication/260771330_Developing_multilevel_models_for_analysing_contextuality_heterogeneity_and_change_using_MLwiN_Volume_1_updated_June_2015_Volume_2_is_also_on_RGate

Book Developing multilevel models for analysing contextuality, he...

Dear all,

Thank you very much for your help. I will re read volume 2 to get more hints. Also, I have already downloaded Bill Brown's manual so I will also read it to make a decision.

Thank yuo very much for your help. Any other difficulty I will let you know. Thank you very much for your help!

the MCMC key stuff is in Volume 1

Dear Professor Jones,

Thank you! Yes! I was reading volume one. After doing my analysis it turns out that my two level with only direct effects has a better model fit than the one with interactions (contrary to my single level model). 

I am learning quite a lot about this. I think I will keep reading about MCMC and bayesian statistics because they are very useful and gives great insights. 

Thank you very much for your help

Hi everyone, I encountered a similar problem recently, where the estimation for a variable changed sign. It is a cross-classified logit model: level 1 patient nested in level 2 hospital, cross classified with level 3 neighbourhood. The estimation for one neighbourhood variable in a single level model is negative changed to positive when using cross-classified structure. both models are estimated in MLwiN. What is the possible reasons for it? (The inter-hospital var is 5.3, much larger than inter-neighbourhood var, 0.06).