Linear VARs assume that a one-standard-deviation monetary shock produces the same output response in a deep recession as in a boom. That symmetry is a maintained hypothesis, not a finding. Threshold VARs drop it. The idea traces to Tong's (1990) self-exciting threshold autoregression (SETAR) for univariate series and was extended to the multivariate case by Tsay (1998). Hansen (1996) supplied the inferential machinery: because the threshold parameter vanishes under the null of linearity, conventional Wald and LM tests have non-standard distributions, and Hansen's bootstrapped sup-LM statistic became the workhorse test for threshold existence.
The mechanism is direct. Pick an observable variable - the output gap, credit spread, or transition indicator - and call it the threshold variable. The model searches over candidate cutoff values and splits the sample into two (or more) regimes. Each regime gets its own VAR coefficient matrix and its own innovation covariance. The switch is instantaneous: once the threshold variable crosses the cutoff, the entire parameter vector flips. This is what separates the Threshold VAR from its smooth-transition cousin, which allows gradual migration between regimes.
Estimation proceeds by concentrated least squares. For each candidate threshold, the sample splits and regime-specific OLS is fast. The concentrated sum of squared residuals is then minimized over the threshold grid. Hansen (1997, 2000) showed that the threshold estimator is super-consistent under regularity conditions, converging at rate T rather than the usual root-T. Confidence intervals for the threshold come from inverting a likelihood-ratio statistic, not from the asymptotic normal approximation.
Central banks and fiscal authorities use Threshold VARs to study asymmetric policy transmission. Balke (2000) documented that credit-channel effects strengthen when the credit spread exceeds a threshold, a finding that linear models miss entirely. Lo and Zivot (2001) applied the framework to threshold cointegration in the law of one price, using disaggregated price data; threshold adjustment in interest rates is due to Enders and Granger (1998) and Enders and Siklos (2001). More recent work nests Threshold VARs inside Bayesian frameworks with stochastic volatility to handle both regime switches and time-varying uncertainty.