What would be the difference of estimating a variable inside and outside a VAR model? Namely, if I know the relevant explanatory variables to model a certain variable in a time series framework, what's wrong with estimating a distributed lag instead of a VAR?

Edit: For me it's clear the limitations in terms of forecasting and impulse-response analysis. I'm not sure about the impacts on coefficients and residuals.

EDIT: I believe that an answer for my question is that it would be a specification error if we estimated an ADL model for a set of variables (in which only one would enter endogenously) that we know to affect each other in a contemporaneous way. In that case, some type of VAR is the correct approach since we'd need to take into account the multivariate covariance matrix of the residuals for appropriate inference.

  • $\begingroup$ Can you give us any more detail about your precise scientific question? $\endgroup$
    – mdewey
    Aug 10, 2016 at 19:33
  • $\begingroup$ @mdewey specifically, I'd like to know the consequences regarding coefficient estimates and residuals. Thanks! $\endgroup$ Aug 10, 2016 at 19:36
  • $\begingroup$ Related thread. $\endgroup$ Aug 10, 2016 at 19:43
  • $\begingroup$ @RichardHardy Thanks, but the thread doesn't tackle my specific doubt concerning the consequences on coefficients estimates, though it mentions the need for different inference techniques. $\endgroup$ Aug 12, 2016 at 4:14
  • $\begingroup$ @lucasfariaslf, what do you want to use the model for? Inference on some hypothesis invovlving testing of model coefficiens? Forecasting? I am trying to understand what exactly the question is. "Estimating a variable" is not quite clear. $\endgroup$ Aug 12, 2016 at 5:28

1 Answer 1


I am having a difficulty understanding the actual question, but let me provide some clarification that could be helpful. Take two time series, $x_t$ and $y_t$.

Distributed lag DL($q$) model:

$$ y = \beta_0 + \alpha_1 x_{t-1} + \dotsc + \alpha_q x_{t-q} + \varepsilon_t. $$

Autoregressive distributed lag ARDL($p,q$) model:

$$ y = \beta_0 + \alpha_1 x_{t-1} + \dotsc + \alpha_q x_{t-q} + \beta_1 y_{t-1} + \dotsc + \beta_p y_{t-p} + \varepsilon_t. $$

One equation of vector autoregressive VAR($r$) model:

$$ y = \beta_0 + \alpha_1 x_{t-1} + \dotsc + \alpha_r x_{t-r} + \beta_1 y_{t-1} + \dotsc + \beta_r y_{t-r} + \varepsilon_t. $$

Take $r=\max(p,q)$, set $\alpha_j=0$ for $j>q$ and set $\beta_j=0$ for $j>p$; you get that one equation of this restricted VAR($r$) is the same as ARDL($p,q$). Furthermore, set $\beta_j=0$ for $j=1,\dotsc,p$ to get an DL($q$) model.

From this perspective, the models are not that different. If you do not care about forecasting (which is straightforward with VAR but less so with DL or ARDL because the latter two do not give forecasts for $x_t$), the DL, ARDL and one equation of a VAR allow you to do essentially the same thing.

(The coefficients of a DL or an ARDL model may be restricted, e.g. by Koyck or Almon lag structures, but in principle this could also be done in a VAR model.)

  • $\begingroup$ Thanks for the further clarification. I suspect my doubt regarding the implications of the chosen structures has more to do with the hypothesis made on the error terms, since the only immediate difference reported in many threads is the straightforward forecasting solution that VAR offers. The motivation for my question was in the hope that a similar discussion as the one in the 1980 paper by Sims (i.e. Simultaneous Equations x VAR) existed for ARDL and VAR. $\endgroup$ Aug 22, 2016 at 4:32

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.