As I understand your major concern is that the estimated model is dynamic, so standard panel data estimators, such as fixed effects and random effects  are biased. Standard approach to addressing this problem is to apply an instrumental variable estimator, such as that proposed by Arellano and Bond (1991) or Arellano and Bover (1995) - these estimators are asymptotically consistent, but their properties are unsatisfactory in the case of short samples as yours. In such a case it is possible to correct the bias of the standard estimators without affecting their efficiency e.g. by corrected least square dummy variable estimator proposed by Bun and Kiviet (2002) and modified for the analysis of the unbalanced panels by Bruno (2005). This estimator is available for STATA as user written procedure “xtlsdvc”.

I am not sure if you can use GMM with some variables are I(0) and some I(1) . I would put my money in favour of ARDL. However, If you find any literature which uses GMM for such variables then cite here, please. 

Hello Ajay, since you have a short panel (N > T), GMM is most suited....but I may be wrong since there is no 'best' estimation technique....but I will stake it out with GMM in this case.

Use both ARDL and GMM but implement an error test to see if both are ok and which of them have the lowest error.

@ Piotr Cizkowicz : Could you name the titles of the publications by Bun and Kiviet (2002) and Bruno (2005) you were referring to ?

Thanks and best

Tobi

I am currently having a dataset of T=20 and N=15, and will most likely use the LSDVC estimate by Kievit and GMM, according to the paper by Judson, Owen (1999). Unfortunately, I am now questioning whether N=15 may be too small, does anybody have an insight concerning this ?

Enlarging to N=25 would be possible without too many issues.

Thanks already and best regards

Tobi

I am currently running a dataset of T=20 and N=20. is it possible to run GMM or panel ardl(mg, pooled mean group)?

which is most appropriate?

in addition what commands do i use to perform diagnostic tests when running PMG?

Well. ARDL is used for long time series with large frequency with given cross sectional units where unit root is the issue, otherwise stationarity of the series is not the concern. GMM vs. ARDL is not a matter of choice. These are already defined for specified data and model design.

I am analyzing a panel data set containing 17 years data for 107 countries and export value is my dependent variable while having 8 independent variables. I want to asses the long run short run relationship of my dependent and independent variables. I found that my dependent variable and six independent variables are stationary at level where as other 2 inependent variables are stationary at I(1). My first question is can I apply ARDL model for my data? My data in natural log form because of my model is in natural log form. My another problem is I have many negative log values in my data set and many missing values as well. I want to know whether this type of daat will support to perform ARDL test in STATA? Is there any limitation of number of independent variables can be accomodated in ARDL model? I have stucked with my master thesis at this point.

yes you can use small T and large N,

" The first empirical work of the present study is to check the cross-sectional dependence between Yit and Xit. Pesaran (2004) proposed CD test which can be applied when N is larger than T. Since our study includes 23 cross-sectional data (N) and 15 years’ time period (T)."

Hello everyone

I have a panel data that is same as Ajay Massand (N=21, T=20), mixed stationary at I(0) and I(1). Can you help me to find a model which fits my data?

I have try to run with ARDL:

ardl lnlabpro lngdp lncpi ln2cpi lnwr lnw, lags(1 2 1 1 0 0) ec

and got the error result as bellow:

sample may not include multiple panels

r(459);

Could you show me what does the error report mean?

Thank you

Sau

I have a panel data for studying determinant of financial inclusion with T=5 and N=15. Should I use panel ARDL or panel GMM?

SYSTEM GMM is better for panel crossectional data(N>T), while PANEL ARDL MODEL for Timeseries panel data(T>N), moreover, there is a formula for converting negative values to logarithmic check it.

Soufiane Mahjoubi Don't worry about CD in case of N>T. CD is a problem when T>N

Sau Mai xtset i.d years, yearly

On Stata first declare the dataset to the panel data

@Muhammad Dawood

I have to estimate spatial durbin model without Log, it's that possible or no? Because, some variable is ommited with collinearity problem.

The ARDL model is suited for data with N < T. Also the dependent variable must not be stationary at levels.

I have a panel data with T=10 and N=9. CanI use fixed effects method?

Abdullahi Baba Mohammed

Fixed effects approach is better when controlling for heterogeneity.

Here is a great resource from Sohag Kazi concerning N and T.