Hello,
I am a data analyst trying to carry out several research analyses that require mixed effects models. I am working with very large administrative data sets (~1.5 million to 31 million observations) and am having many challenges getting the specified models to converge.
I find it difficult to intelligently troubleshoot these models when I don't have a good understanding of what error codes and startingvalues and startgrid options actually do/mean. Being able to critically troubleshoot these models is important to me because it can take hours/days for each model I am trying to run.
I am hoping you may help explain to me what the following scenarios mean and help me build intuition (or a mental flowchart) around the following scenarios so I can make choices about what to do next:
Basic things:
1. When the model is (concave) or (backed up) and is on its hundredth or more iteration
2. The model is stuck on the same value for its log psuedo-likelihood after many iterations
3. When I get the error codes after running a meglm model like "initial values not feasible" or "discontinuous region..."
I have done as the meglm handbook and many forum posts recommend. For instance,
1. I try startvalues(iv), then startvalues(constantonly), etc.,often in combination with startgrid (). I also will try differing values for startgrid (e.g., startgrid(2))
2. I will also fit a basic regression model (not mixed effects), store the values in a matrix, and use those values in my mixed effects model using the from ( ) command in combination with startvalues and startgrid
These options have been helpful, however, whatever I choose seems rather arbitrary. When I get the model to converge, I don't understand what made the difference. Can you help me understand what scenarios call for what actions?
Thank you for any advice you may have!
Amy
I am a data analyst trying to carry out several research analyses that require mixed effects models. I am working with very large administrative data sets (~1.5 million to 31 million observations) and am having many challenges getting the specified models to converge.
I find it difficult to intelligently troubleshoot these models when I don't have a good understanding of what error codes and startingvalues and startgrid options actually do/mean. Being able to critically troubleshoot these models is important to me because it can take hours/days for each model I am trying to run.
I am hoping you may help explain to me what the following scenarios mean and help me build intuition (or a mental flowchart) around the following scenarios so I can make choices about what to do next:
Basic things:
1. When the model is (concave) or (backed up) and is on its hundredth or more iteration
2. The model is stuck on the same value for its log psuedo-likelihood after many iterations
3. When I get the error codes after running a meglm model like "initial values not feasible" or "discontinuous region..."
I have done as the meglm handbook and many forum posts recommend. For instance,
1. I try startvalues(iv), then startvalues(constantonly), etc.,often in combination with startgrid (). I also will try differing values for startgrid (e.g., startgrid(2))
2. I will also fit a basic regression model (not mixed effects), store the values in a matrix, and use those values in my mixed effects model using the from ( ) command in combination with startvalues and startgrid
These options have been helpful, however, whatever I choose seems rather arbitrary. When I get the model to converge, I don't understand what made the difference. Can you help me understand what scenarios call for what actions?
Thank you for any advice you may have!
Amy
