MerQuR+Statistics Glossary

Statistics Glossary → Lag length selection

Lag length selectionconcept

Deciding how many past periods (lags) to include in autoregressive models (AR, VAR, the ADF unit-root test), usually with the help of information criteria.

A VAR model explains each variable by the values of all the variables over the last p periods; the ADF test adds lagged differences to its regression to remove autocorrelation in the residuals. If p is too small, the remaining autocorrelation biases the coefficients and the tests; if it is too large, unnecessary parameters make the estimates imprecise and reduce the power of the tests. The usual approach is to fit every model from 0 up to a chosen maximum lag on the same sample and pick the lag that minimizes a criterion such as AIC, BIC (SC), Hannan-Quinn (HQ) or the final prediction error (FPE). In the general-to-specific approach, one starts from the maximum lag and drops one lag at a time while the last lag is not significant.

The criteria can disagree: BIC is consistent and tends to choose shorter lags, while AIC and FPE choose longer ones and are often better for forecasting. For the comparison to be valid, all candidates must be fitted to the same observations (dropping as many initial observations as the maximum lag). The residuals of the model with the chosen lag should still be checked for autocorrelation. The maximum lag is itself a choice; for the ADF test, for example, the rule ⌈12(n / 100)^¼⌉ is widely used.

An everyday example

An economist studies the relationship between inflation and interest rates with a VAR on monthly data. Trying lags from 1 to 12, AIC suggests 4 lags and BIC suggests 2. If the residuals of the 2-lag model still show clear autocorrelation, the 4-lag model is preferred; if not, the simpler 2-lag model is enough.

In MerQur

It appears as 'Lag selection' (Manual, AIC, HQ, SC (BIC), FPE) and 'Maximum lag' in VAR; 'ADF specification: lag criterion (AIC / BIC / t-stat / fixed lag) and maximum lag' in Time Series; and 'VAR lag K' and 'ADF lag selection' in Cointegration. The panel unit-root tests of Panel Data Regression include 'Unit root: lag selection' (fixed lag, AIC, SIC, Hall).

Analyses where this concept appears

Panel Data RegressionTime SeriesVAR (Vector Autoregression)Cointegration (Johansen / Engle-Granger)

Related concepts

Information criteria (AIC, AICc, BIC)Unit root tests (ADF, Phillips-Perron)