ARDL / bounds testing Model

Cointegration analysis traditionally requires that all variables be integrated of the same order. The Johansen (1991) procedure assumes a pure I(1) system; the Engle-Granger (1987) two-step method requires the same. In applied macro, this precondition is rarely met cleanly. Unit-root tests have limited power, some series appear borderline I(0)/I(1), and pre-testing for integration order introduces its own specification risk. Pesaran, Shin, and Smith (2001) - hereafter PSS - proposed the bounds-testing approach: estimate an ARDL model in error-correction form and test for the significance of the lagged-level terms. The test provides two sets of critical values - one assuming all regressors are I(0), one assuming all are I(1) - and the sample F-statistic is compared to both. If it exceeds the upper bound, cointegration is confirmed regardless of integration order. If it falls below the lower bound, no cointegration. If it falls between the bounds, the test is inconclusive.

The ARDL model itself is old - Hendry and von Ungern-Sternberg (1981) used distributed-lag specifications throughout the 1970s and 1980s. What PSS contributed was an inferential framework that works in the I(0)/I(1) gray zone. The model reparameterizes the standard $ARDL(p, q_1, ..., q_k)$ into a conditional error-correction form where the long-run coefficients appear explicitly as ratios of lagged-level coefficients and the short-run dynamics appear as differenced terms.

Estimation is straightforward OLS after selecting the lag orders (p, $q_1$, ..., $q_k)$ by information criteria. The bounds test is an F-test on the joint significance of the lagged levels of $y_t$ and $x_t$ in the error-correction equation. Pesaran and Shin (1999) showed that the ARDL approach produces consistent estimates of the long-run parameters even when the regressors are endogenous, provided the lag orders are correctly specified.

The ARDL bounds test is the workhorse cointegration test in applied development economics, energy economics, and agricultural economics. Google Scholar returns over 80,000 citations to PSS (2001), making it one of the most cited econometrics papers ever. It is the default approach in EViews, Stata's ardl command, and the R package dynamac.

The $ARDL(p, q_1, ..., q_k)$ model takes a single dependent variable $y_t$ and k regressors $x_{1t}$, ..., $x_{kt}$. Each regressor enters with its own lag order $q_j$, selected by AIC, BIC, or similar criteria. The model includes p lags of $y_t$ and up to $q_j$ lags of each x_jt. Deterministic components (intercept, restricted trend, unrestricted trend) are included depending on the case (PSS tabulated five cases).

The error-correction reparameterization rewrites the ARDL as: Delta $y_t =$ alpha_0 + pi_y * $y_{t-1} +$ sum(pi_j * $x_{j,t-1}) +$ short-run differenced terms + $u_t$. The coefficient pi_y is the speed of adjustment: it must be negative for the ECM to be stable. The long-run coefficients are recovered as beta_j = -pi_j / pi_y. The bounds test is an F-test for $H_0$: pi_y = pi_1 = ... = pi_k = 0 (no levels relationship).

Output includes: (1) the estimated long-run coefficients beta_j with standard errors (the delta method or Bewley transform), (2) the short-run coefficients on the differenced terms, (3) the error-correction speed pi_y, (4) the bounds F-statistic and its position relative to the I(0) and I(1) critical value bounds, and (5) diagnostic tests (serial correlation, heteroskedasticity, functional form, normality).

The World Bank and IMF use ARDL bounds testing to study long-run determinants of economic growth, financial development, and energy consumption in developing countries where time series are short (T = 30-60) and integration orders are uncertain. The single-equation framework is attractive because it avoids the Johansen procedure's requirement for a well-specified VAR with sufficient degrees of freedom.

Energy economists use the ARDL to test the Environmental Kuznets Curve (EKC) hypothesis: whether pollution first rises and then falls with per-capita income. The long-run coefficient on income-squared must be negative and significant. The bounds test confirms whether a genuine long-run relationship exists or whether the quadratic pattern is spurious.

The ARDL cannot handle more than one cointegrating vector in a single equation. When the system has multiple long-run relationships (e.g., a money-demand equation and a Phillips curve that share variables), the Johansen VECM is needed. The ARDL also requires that no variable is I(2), which rules out nominal price series in hyperinflationary episodes. With very short samples (T < 30), the bounds F-statistic has low power and the critical values from PSS (2001) may not apply - Narayan (2005) tabulated small-sample critical values.

Extensions include the nonlinear ARDL (NARDL) of Shin, Yu, and Greenwood-Nimmo (2014), which decomposes regressors into positive and negative partial sums to test for asymmetric long-run and short-run effects. The quantile ARDL (QARDL) extends the framework to quantile regression. Panel ARDL estimators (Pesaran, Shin, and Smith 1999) handle cross-country heterogeneity through mean-group and pooled mean-group estimators.