I've been trying to use the msm package in R to model an 6 state, multi-state model of a disease. My data set, in total, contains about 22,000 subjects, with slightly over 81k observations.

I had a lot of trouble getting my data prepped because my data was pretty messy. However, now that my data is all ready, when I run the model I get an overflow error:

MD.msm <- msm( PatientState_num ~ Years, subject=IDNumber, data = MD_Death2, qmatrix = Q2.crude, deathexact = 6)

Error in Ccall.msm(params, do.what = "lik", ...) : 
  numerical overflow in calculating likelihood

I cannot seem to find any resources that would help me diagnosis what is causing the overflow nor how I might correct it.

The closest bit of advice that I can find from Page 8 of Multi-State Models for Panel Data: The msm Package for R (https://www.jstatsoft.org/article/view/v038i08) states:

To facilitate convergence, the "BFGS" quasi-Newton optimization algorithm is used (see the documentation for the R function optim()), and the maximum number of iterations is increased to 10000. The -2 log-likelihood is also divided by 4000, since it takes values around 4000 for plausible ranges of the parameters. This ensures that optimization takes place on an approximate unit scale, to avoid numerical overflow or underflow

However, if I just use those options I still get the overflow error.

Any suggestions?

In case you need to look at the crude initial values for transition intensities they are:

Q2.crude <- crudeinits.msm(PatientState_num ~ Years, IDNumber, data=MD_Death2, qmatrix=transitions_allowed)

> Q2.crude
           1          2           3            4          5          6
1 -5.1078374  2.8763872  1.35671107  0.130212308  0.5666499 0.17787694
2  0.7587895 -2.4108875  0.05652063  0.583858120  0.9748396 0.03687971
3  1.0173350  1.2362397 -3.07569242  0.255661062  0.3328495 0.23360723
4  0.3807051  0.7235918  0.16135846 -1.265655417  0.0000000 0.00000000
5  0.7331575  0.4638941  0.21511807  0.004390165 -1.4941194 0.07755958
6  0.0000000  0.0000000  0.00000000  0.000000000  0.0000000 0.00000000

Also if you need to see the state table:

statetable.msm(PatientState_num, IDNumber, data=MD_Death2)
from     1     2     3     4     5     6
   1    73 10802  5095   489  2128   668
   2  5370 12681   400  4132  6899   261
   3  2491  3027  1331   626   815   572
   4   151   287    64     1     0     8
   5  1002   634   294     6   101   106

2 Answers 2


I have had the same problem. In my case i solved it by first differentiate time within each subject and take the minimum value. That can be done using aggregate():

diffs <- aggregate(cbind(minDiff=Years)~IDNumber, FUN=function(x) min(diff(x)),data=MD_Death2)

If you then merge this with your data you can easily exclude subjects in which there are very small differences in time between each observation.

MD_Death2 <- merge(sampleData,diffs,by='IDNumber',all.x=T)

In my case, after excluding subjects that had less than 0.01 years between two observations (which in my case were likely errors) solved the problem. If still does not work you might also try increasing the fnscale option further, (and maximum iterations if necessary), eg pass


to msm().


It's hard to deal with questions of this type if one doesn't have the data to actually debug the specific error (numerical or otherwise). An alternative that you might consider (and it may help you in understanding your transitions better) is to break up the estimation into different parts. It's a lot more manual work compared to running one simple estimation in MSM but may provide additional insights to your analysis.

Here's how I would do it:

  1. Transform your data into "long format", i.e, each id contributes as many records to your data set as the number of time intervals that its been observed for.
  2. Create a categorical variable that represents the mutually exclusive states you've described and a lagged variable that represents this state value in the prior period.
  3. Split your consolidated data set into 5 separate datasets, based on the 5 "from" states you have, based on the value of the prior state variable. Each observation in this data subset will have the same value for the prior state but different values for the "to" state, unless they are censored in which case there is no transition.
  4. For each of these, estimate either a multinomial logistic, or separate binary logits, with the dependent variable being either the multi-category "to" state (or a single category if you use a binary logit). In either case, the prior state becomes your reference category (e.g., if Y is the dependent variable, it is set to 0 with the to categories being either 0, 1, 2, 3, 4 or 0, 1 (binary case, with 1 being the chosen "to" state.)).
  5. This framework will allow you to explore and estimate other possible features of the data such as "duration-dependence" as a function of the periods. If you want to explore this, you should also create a "time in state" variable that is set to 0 at the beginning of each spell for every id in your data.
  6. This framework should also allow for "multiple" spells in the data where the same id can transition between states.
  7. Given that some of your states are sparse, you'll find plenty of warnings about perfect separation etc., In those instances, you can start with a constant term and then add variables based on some measure of model errors.
  8. In order to make the task manageable and reduce code duplication, you should write a function that takes as input a from and a to state and outputs the estimates from the GLM.
  9. Other nice tools such as glmnet, cross-validation etc also become available to you.

Sorry I couldn't help you with your original question but hope this helps you think through a path forward for your research.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.