[Federico Belotti]

Hi -- I'd like to bump this thread. I'm still interested if there's any more to say about Nick's third question above:

• What's the best way to assess high leverage observations when running XSMLE? Should I just estimate a pooled OLS model and then calculate DFBETA and Cook's D or is there another approach?

The post-estimation options after XSMLE don't include, as far as I can tell, any help with leverage, cook's d, etc. Are there any other options?

Ethan

Ethan:

As you correctly pointed out, the post-estimation options of -xsmle- do not include leverage-points detection measures. As far as I know, unless you code it by yourself, there are no other options. My view is that, since you can obtain the hessian matrix after an -xsmle- estimation using the -posthessian- option, something like equation (2.2) in https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3495605/ might be helpful and quite simple to code once the leave-one-subject out estimates have been obtained.

Federico

[Tim Umbach]

Guys, I've got another question: When I use the SDM Model with time fixed effects the ML estimator does not converge - it stops after iteration 100 with the statement 'convergence not achieved'. (It converges when I forego the time fixed effects) Even after scaling up iteration size (e.g. 5000) it does not converge.

Do you have any idea what I can do about it? Respectively, do you think the results are reliable?


Thanks a lot!

Hi Bastian,

no, these results are not reliable, as 100 (or 5000) iterations are a random cut-off point. So the log-likelihood hasn't converged and therefore it is dubious to claim that you have found a maximum. When I encounter this, it is usually because I have mis-specified the model or have chosen bad starting values. So here is what you can do:
1. Use the difficult option of xsmle , which uses a different optimization algorithm for the not concave parts. Or you could also experiment with the different optimization algorithms, see help maximize. But while this is the easiest route, it's not the most likely to succeed.
2. Try different starting values via from()
3. Try a different specification of the model. I've encountered this issue for example when I've use the same variable twice as regressors by accident, thereby creating collinearity.

I hope that helped,

Tim

[wanhaiyou]

Hi Bastian,

no, these results are not reliable, as 100 (or 5000) iterations are a random cut-off point. So the log-likelihood hasn't converged and therefore it is dubious to claim that you have found a maximum. When I encounter this, it is usually because I have mis-specified the model or have chosen bad starting values. So here is what you can do:
1. Use the difficult option of xsmle , which uses a different optimization algorithm for the not concave parts. Or you could also experiment with the different optimization algorithms, see help maximize. But while this is the easiest route, it's not the most likely to succeed.
2. Try different starting values via from()
3. Try a different specification of the model. I've encountered this issue for example when I've use the same variable twice as regressors by accident, thereby creating collinearity.

I hope that helped,

Tim

Hi Tim,

I have the same problem.Thanks very much for your suggestions.

convergence not achieved
Computing marginal effects standard errors using MC simulation...

I don't understand the comment ##3：Try a different specification of the model.
Why not choose the model using the related test? For example, we can employ the LR test to choose SDM and SAR model.

Other suggestions?

Thanks again for your help!

Bests,
wanhaiyou

[Tim Umbach]

Hi Tim,

I have the same problem.Thanks very much for your suggestions.

convergence not achieved
Computing marginal effects standard errors using MC simulation...

I don't understand the comment ##3：Try a different specification of the model.
Why not choose the model using the related test? For example, we can employ the LR test to choose SDM and SAR model.

Other suggestions?

Thanks again for your help!

Bests,
wanhaiyou

Hi wanhaiyou,

I meant the third suggestion more in terms of what variables to include,whether to include them in form of logs or not, whether to include time lags and in what form etc.

It is important to note, however, that the choice of model should NOT purely rely on tests. Instead you should have a theoretical model on how each regressor relates to the regressand in spatial terms and then test this theoretical hypothesis with a test.

The optimization can also fail if you have many outliers (check via scatterplots). If this is the case, you could include an outlier dummy in your estimation. Deleting them is not a good idea, as xsmle needs a balanced panel. In the end it is also possible that the theoretical distribution you haven chosen to maximize (Gaussian with xsmle) is wrong. In this case you have a serious problem I'm afraid I can't help you with. To check this I use histograms and kernel density estimation, although there are also formal tests of normality.

I hope this is of some help,

Tim

[wanhaiyou]

Hi wanhaiyou,

I meant the third suggestion more in terms of what variables to include,whether to include them in form of logs or not, whether to include time lags and in what form etc.

It is important to note, however, that the choice of model should NOT purely rely on tests. Instead you should have a theoretical model on how each regressor relates to the regressand in spatial terms and then test this theoretical hypothesis with a test.

The optimization can also fail if you have many outliers (check via scatterplots). If this is the case, you could include an outlier dummy in your estimation. Deleting them is not a good idea, as xsmle needs a balanced panel. In the end it is also possible that the theoretical distribution you haven chosen to maximize (Gaussian with xsmle) is wrong. In this case you have a serious problem I'm afraid I can't help you with. To check this I use histograms and kernel density estimation, although there are also formal tests of normality.

I hope this is of some help,

Tim

Thanks very much for your valuable comments, Tim.


Bests,
wanhaiyou

[Bastian Longer]

The Lee and Yu (2010) transformation approach in the case of a general model with both individual and time effects cannot be implemented using -xsmle-. You need to code it from pag. 169-172 of L.-f. Lee, J. Yu, Journal of Econometrics 154 (2010).

@ ALL

Sorry Frederico, but I have no idea how to code the transformation approach in the case of a general model (I use XSMLE and a SD Model) with both individual and time effects from Lee and Yu. I just had a look in the paper and they state that these effects can be eliminated by taking deviations from time and cross section means... So what they mean I guess, is a double de-meaning procedure, right?

I would have to demean all my variables in order to eliminate my individual and time fixed effects... is that right? Can you help me with this?

Thanks a lot!