tsDyn (version 0.7-60)

TVAR.LRtest: Test of linearity

Description

Multivariate extension of the linearity against threshold test from Hansen (1999) with bootstrap distribution

Usage

TVAR.LRtest(data, lag=1, trend=TRUE, series, thDelay = 1:m, mTh=1, thVar, nboot=10, plot=FALSE, trim=0.1, test=c("1vs", "2vs3"), model=c("TAR", "MTAR"), hpc=c("none", "foreach"), trace=FALSE, check=FALSE)

Arguments

data

multivariate time series

lag

Number of lags to include in each regime

trend

whether a trend should be added

series

name of the series

thDelay

'time delay' for the threshold variable (as multiple of embedding time delay d) PLEASE NOTE that the notation is currently different to univariate models in tsDyn. The left side variable is taken at time t, and not t+1 as in univariate cases.

mTh

combination of variables with same lag order for the transition variable. Either a single value (indicating which variable to take) or a combination

thVar

external transition variable

nboot

Number of bootstrap replications

plot

Whether a plot showing the results of the grid search should be printed

trim

trimming parameter indicating the minimal percentage of observations in each regime

test

Type of usual and alternative hypothesis. See details

model

Whether the threshold variable is taken in level (TAR) or difference (MTAR)

hpc

Possibility to run the bootstrap on parallel core. See details in TVECM.HStest

trace

should additional infos be printed? (logical)

check

Possiblity to check the function by no sampling: the test value should be the sme as in the original data

Value

• A list containing:

  -The values of each LR test

  -The bootstrap Pvalues and critical values for the test selected

Details

This test is just the multivariate extension proposed by Lo and Zivot of the linearity test of Hansen (1999). As in univariate case, estimation of the first threshold parameter is made with CLS, for the second threshold a conditional search with one iteration is made. Instead of a Ftest comparing the SSR for the univariate case, a Likelihood Ratio (LR) test comparing the covariance matrix of each model is computed.

$$LR_{ij}=T( ln(\det \hat \Sigma_{i}) -ln(\det \hat \Sigma_{j}))$$ where $\hat \Sigma_{i}$ is the estimated covariance matrix of the model with i regimes (and so i-1 thresholds).

Three test are avalaible. The both first can be seen as linearity test, whereas the third can be seen as a specification test: once the 1vs2 or/and 1vs3 rejected the linearity and henceforth accepted the presence of a threshold, is a model with one or two thresholds preferable?

Test 1vs2: Linear VAR versus 1 threshold TVAR

Test 1vs3: Linear VAR versus 2 threshold2 TVAR

Test 2vs3: 1 threshold TAR versus 2 threshold2 TAR

The both first are computed together and avalaible with test="1vs". The third test is avalaible with test="2vs3".

The homoskedastik bootstrap distribution is based on resampling the residuals from H0 model, estimating the threshold parameter and then computing the Ftest, so it involves many computations and is pretty slow.

References

Hansen (1999) Testing for linearity, Journal of Economic Surveys, Volume 13, Number 5, December 1999 , pp. 551-576(26) avalaible at: http://www.ssc.wisc.edu/~bhansen/papers/cv.htm

Lo and Zivot (2001) "Threshold Cointegration and Nonlinear Adjustment to the Law of One Price," Macroeconomic Dynamics, Cambridge University Press, vol. 5(4), pages 533-76, September.

See Also

setarTest for the univariate version. OlsTVAR for estimation of the model.

Examples

data(zeroyld)
data<-zeroyld

TVAR.LRtest(data, lag=2, mTh=1,thDelay=1:2, nboot=3, plot=FALSE, trim=0.1, test="1vs")