best when estimating ARDL model because of the following reasons:

1. Lag length criteria: Appropriate no. of lags for each of the independent variable and the

most parsimonious model is chosen automatically.

2. It estimates Pesaran et al., ARDL model which may include I(1) and I(0) variables (but not

I(2))

3. Tests for co-integration using bound test approach is provided for in the module

4. It includes a provision of estimating the error term (co-integrating coefficient), short run

and long run coefficients directly.

1. First calculate the F-Value by Bound testing approach, by getting the F-value you can be

in position whether cointegration exist among your indicators or not. After confirmation the

cointegarion in your model you can get ARDL(Long run) and ECM (short run) results,

2- Command is “ardl depvarriable indepvar1 indepvar2 indepvar3 … , aic ec regstore

(ecreg)”

Other general command are:

“estat dwatson” (Durbin Watson statistics, at 1st order autocorrelation).

“estat archlm” (ARCH LM test for higher order autocorrelation)

“estat bgodfrey” (Breusch Godfrey LM test for higher order autocorrelation)

“estat hottest” (Breusch Pagan Heteroscedasticity test)

“estat ovtest” (Ramsey RESET test)

“estat vif” (Test for the Multicollinearity

the dependent variable and lags and leads for othe variables too. And it may contain both

the long run and short run(ecm) daynamics.

requires large observations can be applied let alone ARDL that can produce robust result

even in small observations.

thing authors on academic papers put all diagnostic test result from first model ARDL (the

variables of model at level) on last result, but on reality the diagnostic result first model (for

Ardl Bound Testing) and last model (cointegrating form and long-run coefficient) is different.

i was asking in this forum and another forum statistic but nobody tell me how to do. most of

them suggest me to present diagnostic result and testing robustness for first model (which

all variables at level). i seach byself and then i found that "the keys" is "equation longrun" on

"cointegration form" from first model. we must "generate" equation long-run and making new

variable with name ECT. then estimate cointegrating model on least square to present short

and long run coefficient (final result).

https://nomanarshed.wordpress.com/2015/08/16/estimating-ardl-with-cointegrating-bounds-

in-stata/

https://www.otexts.org/fpp/8/

best when estimating ARDL model because of the following reasons:

1. Lag length criteria: Appropriate no. of lags for each of the independent variable and the

most parsimonious model is chosen automatically.

2. It estimates Pesaran et al., ARDL model which may include I(1) and I(0) variables (but not

I(2)) ******* this for me was a Eureka moment*****

3. tests for co-integration using bound test approach is provided for in the module

4. It includes a provision of estimating the error term (co-integrating coefficient), short run

and long run coefficients directly

Seye Olasehinde-Williams commneted> Serial correlation is not a problem in ardl if you

choose sufficient lags.

Tella Oluwatoba Ibrahim commented> practical experience has shown that the problem can

be solved changing the lag selection. I am sure he didn't use 1-1 lag model... so he needs to

be careful in lag selection to prevent the problem of micronumerousity

Seye Olasehinde-Williams commented> Check Stock & Watson page 612

Zia Eco Marwat commented > its not aproblem because ARDL handel the serial correlation

prob ...

Oussama BA commented> The aim of ardl model is to remove serial correlation. But you

must choose the correct lags. If you chooses your lags as suggested by Aic and Sbc criteria

your model must be good. But think about choosing different lags for your different variables.

Sheikh Muzammil Naseer commented> Just run the model with default lag...

Sayed Hossain commented> The error correction term (-1.292) here is negative and

significant meaning that there is a long run causality running from independent variables to

dependent variable. It also confirms that all the variables are cointegrated or have long run

relationship. We can also say that about 129.27 percent gap between long run equilibrium

value and the actual value of the dependent variable(FDI) has been corrected. It can be

also said that speed of adjustment towards long run equilibrium is 129.27 percent annually

(provided data is annual). Also we can say that system corrects its previous period

disequilibrium at a speed of 129.27% annually. But the speed or adjustment at 129.27%

seems to be over adjusted or may not be practical.

Himmy Khan commented> The error correction coeff should not be lesser then (-1). Not

good for the model. Recheck your data and model.

December Man commented> speed of adjustment to equilibrium is 129%.

Abdul Rahman Nizamani commented> And it also means there is a significant long run

relationship among the variables.

Noman Arshed commented> It is over correcting. Not a sustainable equilibrium.

Olasehinde Timilehin commented> Noman Arshed had said it all. Something must be

wrong....error correction term shows that... .equilibrium convergence does not exist....Model

may needs remodification...Linear model may not be best option. There may be need to

correct for break...Almost, it can reveal the true nature of the data (leave it and find its

Causes)

Jijie Housburg commented> Normally we should get -1<ect<0

Olasehinde Timilehin commented> because there must be short run dynamic that drive the

economy towards a steady state (long run).. Since the coefficient of equilibrium correction is

not valid economically.Long run relationship is not feasible here

Nadia Ameer commented> This might becoz of not selecting appreciate lag length....I also

faced same problems in my model...

Saud Ahmad error correction term must be between 0 to -1

Udegbunam Norris Chinonso commented> Equilibrium convergence does not exist.

Aleem Akhtar commented> You can choose whatever lags you want on the basis of lowest

AIC/SIC values. Moreover if you are not getting significant results, you can change lags in

both options.

Saud Ahmad commented> You can use the automatic selection criteria initially. And then

you can reduce your model by testing ristrictions through wald test. As your data is not large

enough so I suggest to take less number of lags to avoid degree of freedom problem.

Usually for annual data 1 or 2 lags are enough and for quarterly 4 lags are enough.

David Mendy commented> The error correction coefficient, estimated at -0.2060 is highly

significant, has the correct negative sign, and imply a low speed of adjustment to

equilibrium. According to Bannerjee et al. (2003) as cited in Kidanemarim (2014), the highly

significant error correction term further confirms the existence of a stable long-run

relationship. moreover, the coefficient of the error term (ECM-1) implies that the deviation

from long run equilibrium level of ( dependent variable ) of the current period is corrected by

20.60% in the next period to bring back equilibrium.

Pareeshay Jahanxeb Khan commented> Is it convergent or divergent to the equilibrium??

Imran Rjn commented> i am sure, it is convergent...

Meo School of Research

Shishir Shakya

Noman Arshed

Hossain Academy Note

Univariate Models |

Multivariate Models |

Panel Data Model |

basically, the are two school of thoughts. the rigid proponents of 0 to -1 and others which see

nothing wrong from -1 to -2 like Narayan Kumar(2006).

Ade Kutu

Afolabi Luqman

Abdullah Sonnet

Asad Zaman

Atiq Rehman

Burcu Özcan

Ghumro Niaz Hussain

Muhammad Anees

Mohammad Zhafran

Muzammil Bhatti

Monis Syed

Mine PD

Moulana N. Cholovik

Muili Adebayo Hamid

Nicat Gasim

Najid Iqbal

Nasiru Inuwa

Noman Arshed

Rapelanoro Nady

Saeed Aas Khan Meo

Seye Olasehinde-Williams

Suborno Aditya

Sayed Hossain

Shishir Sakya

Sheikh Muzammil Naseer

Tella Oluwatoba Ibrahim

Younes Azzouz