I'm working with a trait.glm analysis from the mvabund package in R. I have two datasets that look exactly the same with the same amount of environmental variables and the same functional traits, but with different species. After running the analysis I do a prediction analysis to get the R-squared values. The R value indicates that my models are a good fit, but when running one of my datasets with the glm I always encounter warnings of non-convergence. Does that mean my model is invalid even though my R values look fine?
Thanks in advance!
Hello,
Warnings of non-convergence in a trait.glm analysis do not always indicate that the model is not valid, but they do suggest that there may be some problems with the model fitting or the data. Non-convergence means that the algorithm used to estimate the model parameters did not reach a stable solution within a given number of iterations. This can happen for various reasons, such as:
- The model is overparameterized or misspecified, meaning that there are too many parameters to estimate or the model does not fit the data well.
- The data are sparse, skewed, collinear, or have outliers or influential points, meaning that there are not enough observations or variation to estimate the parameters reliably or some observations have a disproportionate effect on the estimates.
- The algorithm is sensitive to the initial values, optimization method, or convergence criteria, meaning that the choice of starting values, optimization algorithm, or stopping rule can affect the convergence of the algorithm.
Depending on the cause of non-convergence, you may need to try different strategies to fix it, such as:
- Simplifying or modifying the model, such as removing unnecessary terms, adding regularization or penalty terms, transforming variables, or changing the link function or distribution family.
- Checking and cleaning the data, such as removing missing values, outliers, or influential points, standardizing or centering variables, or reducing collinearity or multicollinearity.
- Changing the algorithm settings, such as increasing the maximum number of iterations, changing the initial values, optimization method, or convergence criteria, or using a different package or function that implements a different algorithm.
You can find more information and examples about how to deal with non-convergence warnings in these web pages (1-3).
However, even if you get non-convergence warnings, you may still get some results from the model fitting. These results may not be invalid, but they may not be reliable or accurate. You should always check the results for plausibility and consistency with other models that do converge. You should also report any non-convergence warnings and how you addressed them in your analysis.
(1) r - What does it mean when glm algorithm doesn't converge but still .... https://stats.stackexchange.com/questions/464601/what-does-it-mean-when-glm-algorithm-doesnt-converge-but-still-gives-results.
(2) Why am I getting "algorithm did not converge" and "fitted prob .... https://stackoverflow.com/questions/8596160/why-am-i-getting-algorithm-did-not-converge-and-fitted-prob-numerically-0-or.
(3) Understanding and fixing false convergence errors in glmmTMB + lme4. https://stackoverflow.com/questions/62478569/understanding-and-fixing-false-convergence-errors-in-glmmtmb-lme4.
Thank you very much for your answer, I've done several of these steps but unfortunatley I still encounter the non-convergence warnings. I, ll continue with exploring my data sets and the model. Thank you once again for your response!
Sometimes the problem is in the data themselves and do what you will it won't converge. I had cases when changing only one data point with 0.01 up or down converges.
- Badges
- Science topic
- Similar topics
- Computer Science
- Computer Communications (Networks)
- Internet