Wednesday, March 6, 2013

ARDL Models - Part I

I've been promising, for far too long, to provide a post on ARDL models and bounds testing. Well, I've finally got around to it!

"ARDL" stands for "Autoregressive-Distributed Lag". Regression models of this type have been in use for decades, but in more recent times they have been shown to provide a very valuable vehicle for testing for the presence of long-run relationships between economic time-series.

I'm going to break my discussion of ARDL models into two parts. Here, I'm going to describe, very briefly, what we mean by an ARDL model. This will then provide the background for a second post that will discuss and illustrate how such models can be used to test for cointegration, and estimate long-run and short-run dynamics, even when the variables in question may include a mixture of stationary and non-stationary time-series.

In its basic form, an ARDL regression model looks like this:

  yt = β0 + β1yt-1 + .......+ βkyt-p + α0xt + α1xt-1 + α2xt-2 + ......... + αqxt-q + εt

where εt is a random "disturbance" term.

The model is "autoregressive", in the sense that yt is "explained (in part) by lagged values of itself. It also has a "distributed lag" component, in the form of successive lags of the "x" explanatory variable. Sometimes, the current value of xt itself is excluded from the distributed lag part of the model's structure.

Let's describe the model above as being one that is ARDL(p,q), for obvious reasons.

Given the presence of lagged values of the dependent variable as regressors, OLS estimation of an ARDL model will yield biased coefficient estimates. If the disturbance term, εt, is autocorrelated, the OLS will also be an inconsistent estimator, and in this case Instrumental Variables estimation was generally used in applications of this model.

In the 1960's and 1970's we used distributed lag (DL(q), or ARDL(0,q)) models a lot. To avoid the adverse effects of the multicollinearity associated with including many lags of "x" as regressors, it was common to reduce the number of parameters by imposing restrictions on the pattern (or "distribution") of values that the α coefficients could take.

Perhaps the best known set of restrictions was that associated with the Koyck (1954) for the estimation of DL (∞) model. These restrictions imposed a polynomial rate of decay on the α coefficients. This enabled the model to be manipulated into a new one that was autoregressive, but with an error term that followed a moving average process. Today, we'd call this an ARMAX model. Again, Instrumental Variables estimation was often used to obtain consistent estimates of the model's parameters.

Frances and van Oest (2004) provide an interesting perspective of the Koyck model, and the associated "Koyck transformation", 50 years after its introduction into the literature.

Shirley Almon popularized another set of restrictions (Almon, 1965) for the coefficients in a DL(q) model. Her approach was based on Weierstrass's Approximation Theorem, which tells us that any continuous function can be approximated, arbitrarily closely, by a polynomial of some order. The only question is "what is the order", and this had to be chosen by the practitioner.

The Almon estimator could actually be re-written as a restricted least squares estimator. For example, see Schmidt and Waud (1973), and Giles (1975). Surprisingly, though, this isn't how this estimator was usually presented to students and practitioners.

Almon's approach allowed restrictions to be placed on the shape of the "decay path" of the gamma coefficients, as well as on the values and slopes of this decay path at the end-points, t=0 and t=q. Almon's estimator is still included in a number of econometrics packages, including EViews. A Bayesian analysis of the Almon estimator, with an application to New Zealand imports data, can be found in Giles (1977), and Shiller (1973) provides a Bayesian analysis of a different type of distributed lag model.

Dhrymes (1971) provides a thorough and very general discussion of DL models.

So, now we know what an ARDL model is, and where the term "Autoregressive-Distributed Lag" comes from. In the next post on this topic I'll discuss the modern application of such models in the context of non-stationary time-series data, with the emphasis on an illustrative application with real data.


References

Almon, S., 1965. The distributed lag between capital appropriations and net expenditures. Econometrica, 33, 178-196.

Dhrymes, P. J., 1971. Distributed Lags: Problems of Estimation and Formulation. Holden-Day, San Francisco.

Frances, P. H. & R. van Oest, 2004. On the econometrics of the Koyck model. Report 2004-07, Econometric Institute, Erasmus University, Rotterdam. 

Giles, D. E. A., 1975. A polynomal approximation for distributed lags. New Zealand Statistician, 10, 22-26.

Giles, D. E. A., 1977. Current payments for New Zealand’s imports: A Bayesian analysis. Applied Economics, 9, 185-201.

Johnston, J., 1984. Econometric Methods, 3rd ed.. McGraw-Hill, New York.


Koyck, L. M., 1954. Distributed Lags and Investment Analysis. North-Holland, Amsterdam.

Schmidt, P. & R. N. Waud, 1973. The Almon lag technique and the monetary versus fiscal policy debate. Journal of the American Statistical Association, 68, 1-19.

Shiller, R. J., 1973. A distributed lag estimator derived from smoothness priors. Econometrica, 41, 775-788.


© 2013, David E. Giles

49 comments:

  1. If the disturbance term is not autocorrelated, why would OLS produce biased estimates?

    ReplyDelete
    Replies
    1. Econometrics 101 - e.g., Y(t) = a + bY(t-1) + e(t), where e(t) is i.i.d. and homoskedastic. Clearly, E[e(t).Y(t)] is not zero; and so E[e(t).Y(t+s)] is not zero, for all non-negative s, and for all t. Hence OLS is biased in finite samples, but the bias vanishes asymptotically (as long as the errors are indeed serially independent). In the case of no drift (a = 0), the bias of the OLS estimator of 'b' is -sb/n, to O(n^-2), where 'n' is the sample size. I believe this was first established by J. S. White, in "Biometrika", 1961.

      Delete
  2. Looking forward to the second part of this and modern uses of ARDL models. I was under the impression that they were relatively old-school models that were put into the dustbin once ARIMA and ARIMAX models became easy to fit.

    ReplyDelete
  3. "Let's describe the model above as being one that is ARDL(p,q), for obvious reasons."

    Sorry to ask the obvious, but why aren't p and q the same? What would be an example of using a different number of lags for the y term and the x term.

    And, any model that includes y (even where p=1) And an x explanatory variable, but also any number of other explanatory variables (z, w, etc.) also with any number of lags would be considered an ARDL?

    And finally, the term autoregressive seems descriptive, but the term distributed lag to describe the other regressors which may only have a lag of 0 (?) isn't intuitively descriptive to me? Not central question to be sure, but it would be helpful to understand the term.

    Thanks!
    Dan

    ReplyDelete
    Replies
    1. Dan - it's the same as in a VAR model - we don't necessarily want the same lag length on all variables. We need to think about the economics of the problem too.

      Yes, we can have additional explanatory variables with their own maximum lag lengths.

      Usually, we'd have several lags of the x (and other explanatory) variables. IN this case the "shape" of the distribution of the weights (coefficients) as we go back through the lags is of interest.

      Delete
  4. Is it possible to study the log-run relationship between three macro-economic variables using ARDL model? Should Engle-Granger Johansen co-integration technique be preferred over ARDL model when we are dealing with three variables?

    ReplyDelete
    Replies
    1. Yes, you can certainly use the ARDL methodology with three or more varaibles, possibly integrated of different orders. The example that's coming in Part II of this post will do just that.

      Delete
  5. Sir
    Eagerly waiting for Part II
    Abhijit

    ReplyDelete
  6. Hi, so in essence, this method can be used instead of a VECM approach, where the variables show cointegration, but aren't all I(1)?

    Looking forward to part 2!

    ReplyDelete
  7. Sania Ashraf- Doctoral Scholar Islamic FinanceJune 12, 2013 at 11:09 PM

    Dear Sir,
    what is the minimum required observations for ARDL estimation?

    ReplyDelete
    Replies
    1. There's no simple answer - it depends on the frequency of your data (monthly, quarterly, annual) and on the number of lags you need to properly specify the model.
      DG

      Delete
    2. Thank you. This has given me a better understanding to ARDL but is there a main difference between this and ARIMA .
      Can't wait for the Part 2

      Delete
    3. ARIMA is for a single time-series. For Part II, see http://davegiles.blogspot.ca/2013/06/ardl-models-part-ii-bounds-tests.html

      Delete
    4. Sir,
      For ARDL if the frequency of data is monthly what would be the minimum number of observation required?

      Delete
    5. I'd be wanting at least 10 years of data - n>120. See below - the tests are only asymptotically valid.

      Delete
  8. Can we apply ardl approach when number of observations are 10?

    ReplyDelete
    Replies
    1. The tests have only asymptotic (large n) relevance, so "no". More generally, an time-series models using only n=10 is unlikely to be of much use.

      Delete
  9. sir u are very affable and full of knowledge.i have read many of ur posts. sir plz tell me that whether we have to find out Durdin h stat in ARDL to cheak out serial Auto correlation or not?if not then DW-stat value detect serial auto correlation correctly and LM test?plz comment sir.

    ReplyDelete
    Replies
    1. The DW test is inappropriate, given the lagged dependent variables. The LM test can be used to test for various orders of AR() and MA() processes in the errors. Keep in mind that this is only an asymptotically valid test, as is the h-test for the AR(1) case.

      Delete
  10. Can Y(t) be I(0) ? or do have to run a unit root to be sure that Y(t) is not I(0)for ARDL?
    Thanks
    Marc

    ReplyDelete
    Replies
    1. See an earlier query - it will usually be I(1), but it need not be.

      Delete
    2. Hi Sir, thanks for your prompt reply.

      I did notice though that some authors are quoting that the dependent variable has to be I(1), (Trade Liberalisation, Financial Development and Economic Growth: Evidence from Pakistan (1980-2009)
      Rao Muhammad Atif , Abida Jadoon, Khalid Zaman , Aisha Ismail, Rabia Seemab,
      Journal of International Academic Research (2010) Vol.10, No.2.).

      I went back and read the referenced Pesaran et al. (2001), article but could not really find that the dependent variable has to be I(1).
      (Pesaran, M.H., Y. Shin., and Smith R. (2001) Bounds testing approaches to the analysis of level relationships, Journal of Applied Econometrics, 16, 289-326.)

      I am a meteorologist, so I might have missed something in the Pesaran article, but it looks like I have to concur with you too that Yt can be I(0) also.

      Regards
      Mark

      Delete
    3. This comment has been removed by the author.

      Delete
    4. I can see nothing in the Pesaran paper on the bounds testing that required the Y variable to be I(1). However, if you proceeded with a Y variable that appeared to be I(0) and then the bounds test gave a clear outcome of cointegration, then you have a conflict. You can't have cointegration unless the variables are non-stationary to begin with.

      So, if you are really, really confident in your unit root testing, and you feel that Y is I(0), in that case it would make little sense to do the bounds testing in the first place!

      Delete
  11. can i use ARDL model for 21 observations ?

    ReplyDelete
    Replies
    1. That's up to you, but in my view that's a very short sample, especially if you are testing for cointegration along the way.

      Delete
  12. Dr. Giles,

    Have you encountered any sort of ARDL of the form?:

    y(t) = a + b*y(t-1) + c(t)*x(t) + e(t)

    with c(t) = k + p*c(t-1) + i(t) (c is an unobservable random coefficient)
    x(t) is observable co-variate. The correlation between x(t) and i(t) may be non-zero, but no correlation between e(t) and either i(t) or x(t)

    thanks!

    ReplyDelete
  13. Tahir Muhammad NaveedJanuary 3, 2014 at 3:45 AM

    Once we decided the lag length of distributed lag model (lag length of explanatory variable), and if the coefficients of a lagged explanatory variable are changing signs, some are positive and some negative. If we add them up, can we say that it is the "net effect" on that explanatory variable.
    Thanks

    ReplyDelete
    Replies
    1. No, we can't because they're measured at different points in time. That's the point of the dynamics.

      Delete
  14. Would it be possible to post on Nonlinear ARDLs. There is a lot of two-step Engle-Granger stuff on asymmetric adjustment in ECMs but I don't see much on asymmetric adjustment in ARDLS or one-step error correction models.

    ReplyDelete
    Replies
    1. Good suggestion - I'll see what I can do. Just need more hours in the day..... :-)

      Delete
    2. In that spirit...this blog is truly great and many of these intricate posts must take a pretty long time to create. Thanks for sharing all your knowledge, I truly appreciate it.

      Delete
    3. Tom, thanks fr the comment and the sentiment. They do take time, but it's fun, and it's nice to "give back".

      Delete
    4. thank u dr, i have a question.
      how can we solve the low degree of freedom risk in ARDL for short-time series?

      Delete
    5. You can't - except by getting more data!

      Delete
  15. Thank You Sir,,, Can I use ARDL Bounds Test even when my data fractionally integrated such as I(1) AND I(2) or more???

    Regardsss

    ReplyDelete
    Replies
    1. I've already made it clear in the post that you can't use it if any of the data are I(2). (BTW, I(1) and I(2) are different from "fractionally integrated".)

      Delete
  16. Dear Sir, I want to investigate this relation y = f(x,d) where y and x two series , y~ I(2) and x~ I(1) by ADF test, d = dummy veraible, from Jan 2003 to June 2011 =0 , and from July 2011 to Dec 2013 = 1. My observations are 132.
    I want to examine the long equilibrium relationship between y and x, and the impact of the dummy veriable on y.
    I got no cointegration between y and x, using cointegration method an error correction model.
    Please how can i analyz this data using dynamic models such as ARDL "remember one series is ~I (2)" or any other model.
    Thank You

    ReplyDelete
    Replies
    1. Ahmed - you could second-difference the Y variable, first-difference the X variable, and use OLS.

      Delete
    2. Thank You Dr. Giles...

      Delete
  17. Dear Sir,
    I am currently conducting a forecast on the demand on petroleum( and its allied products, petrol and diesel) in India. Can this model be used. Eagerly awaiting your inputs.
    regards
    Aditya

    ReplyDelete
    Replies
    1. Yes, as long as none of your series are I(2).

      Delete
  18. Dear Sir
    am working on the determinants of electricity consumption, can i use ardl methodology and how?

    ReplyDelete
    Replies
    1. See here: http://davegiles.blogspot.ca/2013/06/ardl-models-part-ii-bounds-tests.html

      Make sure that none of your variables are I(2).

      Delete
  19. Dear sir,
    am working on income inequality, can i use ARDL as i have only 27 annual observations. Also does ARDL itself takes care of problem of endogeinity. and what about, if there is multicollinearity among explanatory variables, can we still use ARDL. is any eviews code available to run ARDL.
    thanku

    ReplyDelete
  20. See: http://davegiles.blogspot.ca/2014/06/some-questions-about-ardl-models.html

    ReplyDelete