[Kong Chantha]

Daer prof,
1-for ECM, we use the sing (+) or (-) in front of ECM coefficient? But I see many articles used (+) in front. Can you explain ?
2- for ECM has many IV variables in model, (ex. 5 variables), so emc(t-1)=Y(t-1) - bX(t) where b is long run coefficient, how about X(t) ? Dose X(t) is representative for all IV variables ? Or we have another ECM(t-1) formula? Thank

[Kong Chantha]

Dear prof,
1-for ecm, the sign used in front of ECM's coefficient is (+) or (-)? I see many articles used (+). Can you explain ?
2-For ecm that has many IV variables in model (e.g. 5 IV variables) what is ecm(t-1) formula? ECM(t-1) = Y(t-1) - b X(t) where b is long run coefficient and X(t) is representative for all Vairables in long run right ? Or another formulas?
Thank

[Kong Chantha]

Dear Prof,

When i use newey to robust Heteroskedasticity it got in valid code as below ? thank

quietly ardl logEGWsecD1 logTXRD1, regstore(ardlreg)
quietly estimates restore ardlreg
local cmdline '"'e(cmdline)""
gettoken cmd cmdline:cmdline
newey 'cmdline' lag(1)

[Kong Chantha]

Dear prof,

After reading your London conference, doing the newey- robust for short run and long run in ardl-ecm, my question is how about speed adjustment? It's still keeps same coefficient before and after robust ? Thank

[ahmed mohamed ak]

If the dependent variable is I(0) and the independent variable is I(1), can the ARDL approach and the bounds test be applied?

[joseph Mgaya]

Greetings,
1) I have a problem in estimating optimal lag in ardl model, i so some videos saying we should find maximum lag for every variable example (varsoc Y =2, varsoc X1=1, varsoc X2=4) and use the maximum value of all variables (ardl Y X1 X2, maxlags(4) aic), other say we use maximum lag of the dependent variable ((ardl Y X1 X2,maxlags(2) aic) another say that we use maximum lag of each variable in conducting bound test (ardl Y X1 X2 , lags(2,1,4) ec btest). Whch is the best method?
2) What is normal range of speed of adjustment? is the value less than -1 such as -2.5 ok?
3) I have run some variable in ardl model but the results for bound test show only coefficients the rest is empty even in t and F stats, what might be the problem?
4) Can i have a stable ardl model if i have 36 yearly observation and 5 or six independent variable?
Thank you

[Sebastian Kripfganz]

You should not normally use the varsoc command to determine the lag orders for the ARDL model. The varsoc command is for VAR models. For the bounds test, you do not want to be too restrictive with the lag order; you do not normally want to omit any relevant dynamics (i.e. significant lags).

A speed-of-adjustment coefficient less than -1 is highly unusual; a coefficient less than -2 indicates system instability. This and the other problems you encountered appear to be related to the small sample size. I am afraid, with only 36 observations, you cannot estimate a sophisticated model with 5 independent variables and lags of them. If you want to keep all independent variables, you can at best hope to get some meaningful results with a maximum lag order of 1, but even this might be stretching.

More on ARDL estimation and bounds testing:

• Kripfganz, S., and D. C. Schneider (2023). ardl: Estimating autoregressive distributed lag and equilibrium correction models. Stata Journal 23 (4), forthcoming.

[joseph Mgaya]

My ardl model has passed every test (eg Jarque-Bera normality test, White test and LM test) except the Ramsey reset test. Is it still okay to present the results? I have no missing variables in my time series. The cusum is also within the 5% bound.

[Sebastian Kripfganz]

If the Ramsey test for correct functional form rejects the model, then you could try adding squared regressors to your model. Whether this is necessary or not depends on your specific research context and the literature you are addressing.

[joseph Mgaya]

Thank you for your reply, I am really new at this
1)How can I add these squared regressors in my ardl ?
2)Am having four independent variables in my model do i add four?
3) Wont it affect the model because i have only 6 number of observations?

[Sebastian Kripfganz]

You would create squared variables before running the ARDL regression, e.g.

gen x2 = x^2

Whether you want/need to do this for all or just some variables is an empirical question that I cannot answer.

Did you really mean to say "6 observations"? With such few observations, you essentially cannot do anything.

[joseph Mgaya]

sorry i meant 36 observations

[Sebastian Kripfganz]

This is still very small. You should keep your model specification simple to avoid overfitting; specification tests are of limited use here and should not be overemphasized.

[Francisco Neto]

Dear all,

I am currently making use of the ardl command in stata, and had a few questions if you do not mind. While running the ardl command, I have used the "ec" representation (not the "ec1") as some of my variables have optimal q* = 0. Based on Kripfganz and Schneider (2023) the attached equation should be considered (Equation 6). This means that long run coefficients enter at time t and not t-1.

If I ignore the intercept and linear trend can I show that the deviations from equilibrium are (e_t-1 = y_t-1 - θx_t) or am I showing it as e_t instead of e_t-1 now? Please see attached the equation that the authors mention to be the one that is actually estimated on this case (Equation 7). I am now wondering three things. (1) how do I show it explicitly when presenting the two equations what exactly the error correction term (ECT) is in my case and also what subscript for it t or t-1? (2) as I show my short-run results together with the results from alpha (ADJ under the stata output) how should I name the variable that is associated with alpha, ECT(-1) or would it be different given that my long run coefficients enter the model at time t and not t-1? (3) the result I get under ADJ in the output is -0.85. how would I interpret this adjustment coefficient in my case?

Apologies for the long question, but would really appreciate any support.

Best,
Francisco.

[Sebastian Kripfganz]

1. In terms of notation, it is up to you how to define the error correction term. It is most intuitive to define e_t-1 = y_t-1 - θ x_t-1. You could then define e*_t-1 = y_t-1 - θ x_t. Do not just change the time subscript, but rather introduce a different notation (e.g., by adding as asterisk or a tilde or whatever you like).
2. In your results table, you could then also refer to it as ECT(-1) or ECT*(-1) respectively. Just make sure to properly define these terms.
3. Reported is the negative speed-of-adjustment coefficient. A speed-of-adjustment coefficient of 0.85 implies that 85% of a deviation from the long-run equilibrium will be corrected within the next period. This does not depend on whether you choose e_t-1 or e*_t-1. (The interpretation of the other short-run coefficients change slightly, as they are either the effects conditional on a deviation of lagged y from the equilibrium value prescribed by X in the previous or current period.)