Convergence not achieved question

[ChrisF]

Good Morning,

I am having an issue determining a problem I am having when I run xtnbreg. To be brief, I am running two models: The first model has no issues; however, in the second model I convert my two main independent variables to change variables and the model does not achieve convergence. Instead I get a continuous number of iterations and nothing else has changed. I have played with the data and pulled out each control variable individually and I believe I have determined which control variable may be the issue. In addition, I have checked for collinearity with my variables and there is no issue.

My big problem here is that the control variable that I believe is an issue is measured on a scale from 0 to 1 (calculated as a decimal i.e. .787, etc.) and I need to keep it this way. I have multiplied this control by 1,000 to eliminate the decimal just to see if this would have an impact and the models worked.

Any thoughts, suggestions, ideas, etc are much appreciated.

[Clyde Schechter]

Why do you have to keep your variable measured on a scale from 0 to 1? That sounds like saying you have to keep a distance measured in meters, not millimeters. It just doesn't make any sense. The scale of numeric variables sometimes produces numerical difficulties in estimation procedures, and changing scale is a common and innocuous solution.

If the issue is that in reporting results, your variable is conventionally reported on a 0 to 1 scale and you need the coefficient in that metric, consider that when v1000 = 1000*v1, then _b[v1000] = _b[v1]/1000. And a similar relationship applies to the standard errors. So you can easily calculate what the coefficient in the 0-1 version would have been had you been able to achieve convergence with it.

[Joao Santos Silva]

I fully agree with the points made by Clyde, but I would like to understand the problem better. Could it be the case that your other independent variables have large values and the problematic regressor is the only one that is "small"? If that is the case, you may be able to side-step the issue by rescaling the other independent variables so that all your regressors have comparable magnitudes (this is generally a good practice to avoid serious numerical issues).

All the best,

Joao

[Richard Williams]

Here is my paraphrasing of an argument Scott Long has made concerning convergence problems:

Scaling of variables. Scaling can be important. The larger the ratio between the largest standard deviation and the smallest standard deviation, the more problems you will have with numerical methods. For example, if you have income measured in dollars, it may have a very large standard deviation relative to other variables. Recoding income to thousands of dollars may solve the problem. Long says that, in his experience, problems are much more likely when the ratio between the largest and smallest standard deviations exceeds 10. (You may want to rescale for presentation purposes anyway, e.g. the effect of 1 dollar of income may be extremely small and have to be reported to several decimal places; coding income in thousands of dollars
may make your tables look better.)

[Martin Paul]

Dear statalist

I am facing a similar problem with my technical efficiency model.

I estimated using the half normal and exponential specifications. with the half normal model, everything goes fine and the results are very meaningful. However with the exponential it takes like forever with about 16000 iterations. In the end, I get the r430 signal with results with a lot of insisgnificant variables.

Does this point to something?

[Richard Williams]

Martin, I have no idea what a technical efficiency model is or what commands estimate it. It could help if you showed your code and output. Even better might be using dataex to provide an extract from the data and a replicable example. See #12 in the FAQ.

My generic advice is to add the -difficult- option. Often it does nothing but sometimes it works miracles.

My next suggestion is to start with a model that is as simple as possible and gradually add variables to it. You may be able to identify a variable that is causing you grief.

[Richard Williams]

Incidentally, if I were you I would start a new thread rather than add to one that is 3 years old. Make the title something like "Convergence not achieved in technical efficiency model." Then, if there are some technical efficiency modeling experts out there, they may be more likely to see the post and respond. You could include a link back to this post just in case something that was said earlier is still relevant.