PROC IRT Won't Converge

[Rover17]

Hi All,

I am trying to use PROC IRT procedure. I am getting the following warning messages. After the following warning messages, I provide my code below. 

WARNING: The M step for the EM algorithm fails to converge.
Number of iterations or number of function calls
exceeds limit.
WARNING: The estimation algorithm did not converge.

proc irt data=mald.calibration_file_d link=logit itemstat outmodel=mald.initial_parameters_file_d guessprior=(0.2,20)
slopeprior=(1,2) fconv=.01 technique=EM maxiter=1000 maxmiter=1 noad qpoints=11;
group group;
var C9_21 C8_22 C8_23 C8_24 C10_25 C9_26 C10_27 C12_28 C12_29
C13_30 C11_31 C11_32 C12_33 C12_34 C13_35 C12_36 C11_37 C11_38 C11_39
C12_40;
model C9_21 C8_22 C8_23 C8_24 C10_25 C9_26 C10_27 C12_28 C12_29
C13_30 C11_31 C11_32 C12_33 C12_34 C13_35 C12_36 C11_37 C11_38 C11_39
C12_40/resfunc=threep;
run;

Can someone please suggest how to correct my code so that the model converges?

Thank you,

Aaron

[ballardw]

Without any knowledge of your data it is hard to make a pointed suggestion.

A default concept when a limit is exceeded is to see if you can change the limit. Since the warning is the M step you might consider increasing the MAXMITER which from the documentation:

specifies the maximum number of iterations in the maximization step of the EM algorithm. By default, MAXMITER=1.

I might try setting it to larger values though I wouldn't start with anything very large.

Other considerations would be things like the convergence criteria. I do see that the documentation states that the FCONV parameter is not used by the EM algorithm. Perhaps you want to use a different convergence criterion.

You may also want to consider that making a model converge does not mean that the result is actually good model.