使用EViews，运行稳健的最小二乘回归，我无法进行MM估计吗？

Question

I have developed a fairly simple multivariate regression econometrics model. I am now attempting to run Robust Regressions (EViews calls them Robust Least Square). I can easily run a Robust Regression M-estimation. But, every time I run a Robust Regression MM-estimation I run into the same error: "Maximum number of singular subsamples reached." I have played around with the MM-estimation specifications by increasing/decreasing number of iterations, convergence level, etc... Invariably, I run into the same error.

At an EViews forum, another fellow ran into the exact same problem for both MM-estimation and S-estimation. The forum moderator indicated that if a model has the presence of dummy variables without that many observations, such estimations may not reach convergence and generate the error as mentioned above. My model does have dummy variables. And, some of them do not have that many observations (8 consecutive ones out of a time series data with 217 observations). However, I am unclear if this is a limitation of EViews or if this is truly an algorithm limitation. I may attempt to rerun MM-estimation in R. And, see if it is feasible.

Following up on the above, I did just that. And, ran Robust Regression using R with MASS package using rlm() function. Just as in EViews I had no problem running an M-estimation. Similarly, I firt ran into trouble when attempting an MM-estimation. Just as in EViews I got an error message stating the regression/simulation did not reach convergence after 20 iterations. So, I reran my MM-estimation by first eliminating all my dummy variables. As predicted, it worked. Next, I added just one single dummy variable at a time and each time I reran my MM-estimation. I did that to observe when the MM-estimation model would break down. To my surprise, it never did. And, now I eventually could run my MM-estimation with all the dummy variables. I don't know why I could not run it at first with all the dummy variables in at once (maybe I did an error in coding).

This leads me to conclude that R is somewhat more flexible than EViews on this count. After closer inspection, I noticed that the EViews M-estimation I ran was of the bisquare type (vs. the regular Huber one). This makes a big difference. When I did run in R an M-estimation of the bisquare type I almost got the exact same results as EViews. There were small differences between the two. This can be expected given that the solving process is iterative.

Answer 1

As you can read in my commentary, I did quite a bit of work on the issue. In the end, I am unclear why EViews methodically crashes when running a Robust Regression of the MM-estimation type with a model that has a few dummy variables. I feel like it should not. The exact same model using the same Robust Regression methodology was solvable in R with the MASS package and rlm function using method = "MM".

In case you find yourself in similar circumstances, I would advise you to attempt to do the Robust Regression MM type in R instead. I don't know what is the relative resiliance of this process in SAS, SPSS, Python, STATA and other similar software. Hopefully, any of those is more resiliant than EViews on this count.

It is not unlikely that this type of model can actually cause a software to crash (after numerous iterations, the algorithm is not converging towards a solution). But, if my experience is any indicator R has a much higher resiliance threshold than EViews on this count.