[Carlo Lazzaro]

Eszti:
-fe- estimator gets rid of time-inavriant predictor; hence, as far as time-invariant predictors (observed or not) are concerned, there's no omitted variable bias.
Unfortunately, if the incorrect specification refers to time-varying predictor, your regression model might still be at risk of omitted variable bias and endogeneity.
That said, if you mean "how could I estimate time-invariant predictors that omitted due to -fe-machinery?" and -hausman- test points you to -fe- specification, there's nothing you can do.

[Dr. Iqra Yaseen]

Clyde Schechter Carlo Lazzaro

the results I get after using xtcd2 command is

Unbalanced panel detected, test adjusted.
 Missing values imputed for CD*.
 Missing values imputed for CD*.
 Missing values imputed for CD*.
 Missing values imputed for CD*.
 Missing values imputed for CD*.
 Missing values imputed for CD*.
 Missing values imputed for CD*.
 Missing values imputed for CD*.
 Missing values imputed for CD*.
 Missing values imputed for CD*.

Testing for weak cross-sectional dependence (CSD)
   H0: weak cross-section dependence
   H1: strong cross-section dependence
------------------------------------------------------------------------
               |    CD            CDw           CDw+          CD*
---------------+--------------------------------------------------------
ln_STI_HF      |   143.58        0.15          0.15             .
               |  (0.000)     (0.884)       (0.884)       (   .)
ln_STI_EP      |   143.58        0.15          0.15             .
               |  (0.000)     (0.884)       (0.884)       (   .)
ln_STI_SP      |   143.58        0.15          0.15             .
               |  (0.000)     (0.884)       (0.884)       (   .)
ln_STI_EnP     |   143.58        0.15          0.15             .
               |  (0.000)     (0.884)       (0.884)       (   .)
ln_trade_mi~s  |   142.18        0.07          0.07             .
               |  (0.000)     (0.942)       (0.942)       (   .)
ln_patent_a~d  |   143.58        0.15          0.15             .
               |  (0.000)     (0.884)       (0.884)       (   .)
ln_currency~d  |   143.58        0.15          0.15             .
               |  (0.000)     (0.884)       (0.884)       (   .)
ln_income_d~f  |   143.58        0.15          0.15             .
               |  (0.000)     (0.884)       (0.884)       (   .)
ln_WTUI_IMF    |   143.58        0.15          0.15             .
               |  (0.000)     (0.884)       (0.884)       (   .)
ln_discap      |   143.58        0.15          0.15             .
               |  (0.000)     (0.884)       (0.884)       (   .)
------------------------------------------------------------------------
p-values in parenthesis.
References
  CD:        Pesaran (2015, 2021)
  CDw:       Juodis, Reese (2021)
  CDw+:      CDw with power enhancement from Fan et. al. (2015)
  CD*:       Pesaran, Xie (2021) with 4 PC(s)

are the results OK (as i get same results for all variables) or do i need to relook.

i get "insufficient joint observation" error while using xtcdf command

thanks

[Clyde Schechter]

Sorry, but I can't help you here. I am completely unfamiliar with -xtcd2- If it were an official StataCorp command, I would make a fairly confident guess that a message like "insufficient joint observation" should be construed as a warning that the results cannot be relied on. But as it is user-written, I can't even be sure of that much.

[Carlo Lazzaro]

Sybil:
let's hope that JanDitzen comes across this thread and has the time to lend you a hand.
That said, it is not advisable to work with such complex and theoretical demanding estimators without having a pretty precise idea of what they do behind the curtain.

[Dr. Iqra Yaseen]

Clyde Schechter Carlo Lazzaro
thanks for your prompt reply.
so
1. what can be done to avoid "insufficient joint observations" error. the panel id of my data groups (respondent country, minerals and time together).
2. if not above what are the other alternatives for pre-estimating cross section dependence for panel data with N > T. i have already tried xtcd, xtcsd, xtcdf with no satisfactory results.
3. also used pwcorrf command , but i doubt whether this command is valid and reliable to be used for estimating cross section dependency in panel data.

[JanDitzen]

Thanks Carlo Lazzaro for linking me to this post. As you pointed out, as one of the authors of one of the mentioned programs I have some insights.

sybil arqi, what are the dimensions of your data? How large is N, how large is T? Is the panel balanced enough so you can calculate correlation across units?

The CD test is based on the correlation across units. If you panel is unbalanced and there are not enough joint time periods, it might be impossible to calculate the correlations across units. Hence the message "insufficient joint observations". If this is the case, I am afraid there is very little you can do.

The results from xtcd2 are very suspicious because you obtain the same value for each variable. xtcd2 does not check if there is enough data to calculate the cross-correlations, however if there are missings in the data matrix, it tries to adjust the standard CD test and calculate the cross-correlations unit by unit. Still, I believe, there is something wrong in your case, so the first thing I would do is check if the dimensions are fine.

Hope this helps.

[Dr. Iqra Yaseen]

thank you for your response.

Regarding the data dimensions, the data consists of 14 countries (N) studied over 9 years (T), weakly balanced.

with these dimensions, if I can't use xtcdf command and with xtcd2 command showing suspicious results, what alternatives stand out to be used in this case.
please help.

[JanDitzen]

Thanks for the clarification.

The CD test and the concept of strong cross-section dependence require "large" N and T and a sufficient amount of joint observations. Neither is the case in your dataset. In the best case, a cross-correlation between two units is based on 9 observations. Any outlier will influence the correlations heavily. From the outputs you presented, it is likely that there are many combinations in which there is no sufficient number of observations to even calculate the correlations.

In a nutshell: I would strongly disencourage the use of xtcd2, the CD test and the concept of strong cross-section dependence in your setting. What to do depends on your research question. If you main aim is to estimate a model, I would ignore strong cross-section dependence (respectively factors) and try to estimate a FE model with unit and time fixed effects. In my experience, this is likely to do the job.

[Dr. Iqra Yaseen]

JanDitzen thank you so much for your time and effort.