Update: xtdcce2, xtcse2, xtcd2 and new program: xtdcce2fast

[JanDitzen]

Thanks to Kit Baum, an update to xtdcce2, xtcse2 and xtcd2 is available on SSC.

The update includes the following changes:

xtdcce2 (now version 3.0)

• new options: clustercrosssectional() and globalcrosssectional(). clustercrosssectional() creates cross-sectional lags for subsamples (for examples country specific cross-section averages in a dataset with regions across countries) and globalcrosssectional() creates cross-sectional averages taking all observations of the dataset into account, i.e. ignoring if statements.
• improved output for estimated long run coefficients
• various bug fixes and speed improvements (see help for more details)

xtdcce2fast (new!)
is an optimized version for speed and large datasets. In comparison to xtdcce2 it does not perform any collinearity checks does not support pooled estimations and instrumental variable regressions. It also stores some estimation results in mata rather than e() to circumvent some restrictions on matrix dimensions in Stata.

xtcse2 (now version 1.02)

• new option: nocenter, to avoid variables being centered around zero
• estimation of exponent of residuals following Bailey, Kapetanious, Pesaran (2019)
• improvements for unbalanced panels
• bug fixes

xtcd2 (now version 2.3)

• implementation of weighted CD test and Power Enhancement Approach (Juodis and Reese 2021)
• improved support for time series operators
• xtcd2 now removes any unit specific means prior to testing, option nocenter avoids this behaviour

More information, details and examples can be found on GitHub.

[Vita Chen]

Dear Prof.Ditzen,

I am using CCEMG model recently, but my data is short panel, with large N（600）small T (18), and is unbalance data. I have 9 independent variables. When I use xtdcce2, it shows "Units (symbol) to be removed due to insufficient numbers of observations:" and "No observations left." . However, xtdcce2fast can run good results. Why is this? What is the principle of xtdcce2fast?

Many thanks Prof.

[JanDitzen]

Hi Vita,

thanks for using xtdcce2. As this is an old thread, I strongly encourage you to update xtdcce2. The latest version is available on GitHub.

xtdcce2 does several data checks prior to estimation. Among those are checks for collinearity and if the sample size is sufficiently large to estimate the model in the first place. While important, those checks are time consuming, especially in very large panels. xtdcce2fast is a speed optimized version of xtdcce2 and I designed it initially for Monte Carlo Simulations. While faster, it does not do any of the aforementioned checks, takes the data as is and estimates the model. Therefore xtdcce2fast can produce estimation results when xtdcce2 does not. My recommendation is always to run the final specification with xtdcce2 as well to check if all checks are passed.

Your time dimension in comparison to the cross-sectional dimension and the number of variables is very small. In a case without cross-section averages, you estimate 10 parameters (9 coefficients plus fixed effect). Most likely you are adding all variables as cross-section averages, meaning that you try to estimate a model with 19 variables. This is more than you have observations per unit and therefore an estimation is not possible. Hence I would strongly discourage the use a CCE(MG) estimator with your current specification. I would either reduce the number of variables or use a fixed T, large N estimator.

Why does xtdcce2fast still produce results? xtdcce2fast partials out the 9 cross-section averages and then estimates the model based on the defactored out variables using the 18 observations over time. In a sense it does not "know" (or check) that in total 19 observations over time are required to run the model. This behavior is intended to speed computations up, but can lead outcomes as the present one.

Hope this helps.

Jan

[Vita Chen]

Hi Prof.

Thanks you for replying.

I want to ask how to use the fixed T large N estimator? like fixed T standard error correction. Can I do it using xtdcce2 command?

Thanks.

Vita

[JanDitzen]

I was mostly thinking of the standard FE estimator, eg: reghdfe in Stata.

See also here:

https://www.statalist.org/forums/for...-fixed-effects

[dilshad jhn]

Hi Jan JanDitzen I would be very grateful if you can clarify this.
I am using CS-ARDL model for my regression analysis. In your 2021 paper in STATA journal on XTDCCE2 command the CS-ARDL command used level variables ( xtdcce2 c if year >= 1962, > lr(L.c L(0/1).y pi L.pi) lr_options(ardl )[/QUOTE]. The command generates both short run and long run variables. However, there is no first difference of the variables that have been used in the command. So, how does it compare to Chudik and Pesaran (2015) CS-ARDL model?
Any clarification will be greatly appreciated.