REML analysis / Iterations Converged in the Gradient

dwillingmyre — Thu, 05 Oct 2023 15:20:35 GMT

What is the significance of the statement "Iterations Converged in the Gradient" in the Iterations output within the REML Variance Component Estimates Report?

Also, what is the impact on using this REML analysis to evaluate residual error if the iterations do not converge in the gradient? Is this an indication that the dataset was not amenable to the analysis?

Thanks!

Re: REML analysis / Iterations Converged in the Gradient

Mark_Bailey — Fri, 06 Oct 2023 15:50:41 GMT

The REML procedure is a 'search algorithm' or 'numerical optimization.' These methods are used when a closed-form solution does not exist. Such methods are generally not guaranteed to find the global optimum, but they often take steps to minimize converging to a local optimum. Furthermore, the finite precision of the numerical routines and processor can lead to various errors (not an error message, but a departure from the true value). Lastly, most methods quit when one of several convergence criteria is met. That is to say, it is not an exhaustive search.

Lack of convergence might produce 'good enough' results, but it is best not to trust the results in such a case. Failure to converge can be caused by the nature of the data or the convergence criteria are too strict.

topic REML analysis / Iterations Converged in the Gradient in Discussions

REML analysis / Iterations Converged in the Gradient

Re: REML analysis / Iterations Converged in the Gradient