hlme computing error

[LeaChauveau]

Dear developers,

Thanks a lot for providing the lcmm package! I'm currently learning how to use it. Most of the time, it works well but sometimes when I'm trying to compute more than 2 classes with several class memberships in hlme, I'm facing this error:

Numerical problem by computing fn value of function is : NaN

The error has already been documented, and you advised to try with different initial values in argument B. I'm currently using the estimated parameters of the 1-class hlme model but I've read that I could specified my own vector values. However, I do know exactly which values I should choose and if this would fix the error. (I'v also tried to change the max number of iterations)

Here is my current code:

hlme1 <- hlme(Y ~ X1 + c1+ c2+ c3+ c4, random = ~ X1, subject = "id", ng = 1, data = df.analyses)
summary(hlme1)

hlme2 <- hlme(Y ~ X1 + c1+ c2+ c3+ c4, random = ~ X1, subject = "id", ng = 2, data = df.analyses, mixture = ~ X1, classmb = ~ X2+ X3, B = hlme1)
summary(hlme2)

hlme3 <- hlme(Y ~ X1 + c1+ c2+ c3+ c4, random = ~ X1, subject = "id", ng = 3, data = df.analyses, mixture = ~ X1, classmb = ~ X2+ X3, B = hlme1)
summary(hlme3)

hlme4 <- hlme(Y ~ X1 + c1+ c2+ c3+ c4, random = ~ X1, subject = "id", ng = 4, data = df.analyses, mixture = ~ X1, classmb = ~ X2+ X3, B = hlme1)
summary(hlme4)

hlme5 <- hlme(Y ~ X1 + c1+ c2+ c3+ c4, random = ~ X1, subject = "id", ng = 5, data = df.analyses, mixture = ~ X1, classmb = ~ X2+ X3, B = hlme1)
summary(hlme5) 

The error occurs at hlme3, hlme4 and hlme5. Curiously, when changing Y values, models hlme3 and hlme4 are computed.

Surprisingly, when I'm scaling X3, hlme3 is computed but the variance-covariance matrix of the random-effects seems empty. When I'm completely removing X3, all 5 models are computed (but no p-values from hlme4).

Although I don't think a high number of classes is relevant to my analysis, I'd still like to be able to compare models to support my choice of a lower number of classes.

Do you have any suggestion to fix the error?

Thank you in advance!

[VivianePhilipps]

Hi,

this should be fixed by changing the initial values. See my answer in #220. Here you could try by running the models with B=random(hlme1), or better with gridsearch, to start from different sets of initial values.

Viviane