Random Effects with Full Information Maximum Likelihood

[roxy kallsen]

Hello Statalist Users,

I am trying to run a random-effects model with full information maximum likelihood. I used the xtdpdml command suggested by Moral-Benito, Allison, & Williams (2016). Please see the command below.

xtdpdml infractions, inv(mile age oyas length race) pre(contact pwas therapy) ylag(0) fiml re

Infractions is my dependent variable coded as 0 (no infractions) or 1(infractions)
The time-invariant variables were added as suggested by the Hausman test (failed to reject the null hypothesis)
Contact (family contact), pwas (parental warmth and acceptance scale), and therapy (family therapy) are my main IVs. I added these are predetermined variables suggesting that they are sequentially or weakly exogenous variables
Ylag(0) was added since there are only two time points
re was added to indicate that I want a random-effects model
FIML was added to handle missing data

I have two questions.

1. How can I check if my IVs are strictly exogenous or weakly exogenous? This will determine how these variables will be entered in the command.
2. I ran this command and got output but also an error that reads: "Warning, convergence not achieved" (please see attached)

Thank you in advance for your help.


Roxy

[Phil Bromiley]

You will increase your chances of a helpful answer by following the FAQ on asking questions - provide Stata code in code delimiters, readable Stata output, and sample data using dataex.

If you only have two periods per panel and 124 panels, then dynamic panel data analysis is questionable. With y lag(0) you don't have a dynamic model at all so I'm not even sure why xtdpdml - you have no lagged dv anyway so why not stay with xtreg? Once you move to xtreg, then there are many test options available (search this listserve for more info).

[Phil Bromiley]

Note also, without convergence, you can't interpret your results.

[Richard Williams]

With y lag(0) you don't have a dynamic model at all so I'm not even sure why xtdpdml - you have no lagged dv anyway so why not stay with xtreg?

As noted in https://journals.sagepub.com/doi/abs...urnalCode=stja,

With a strongly balanced panel, no missing data, and no effect of lagged y, xtdpdml produces results that are almost identical to xtreg. Of course, there are missing data with this dataset, making the use of xtdpdml with FIML desirable.

So, missing data handling is one reason for sometimes preferring xtdpdml over xtreg.

How can I check if my IVs are strictly exogenous or weakly exogenous? This will determine how these variables will be entered in the command.

The models are nested. Run both and do a chi-square or BIC contrast between them.

I ran this command and got output but also an error that reads: "Warning, convergence not achieved" (please see attached)

Quite possibly the program's fault! We decided that the re option should be handled a bit differently. I have a fix ready and need to send it to SSC. (If you are in a mad rush, email me for the fixed version.) But, having said that, I think you almost always want fe anyway, as xtdpdml can estimate the effects of time invariant variables. fe always wins when I test it against re.

[roxy kallsen]

Phil Bromiley Thank you for the advice. My apologies I did not post a readable STATA output. I will learn how to do this. I did, however, use the delimiters <> for the code. I am not sure why it is not showing properly.
Richard Williams thank you for your reply. I sent you an email. I look forward to hearing back from you.