The effect of a variance shift on the Breusch-Godfrey's LM test

In this article, we study the impact of an abrupt change in variance on the Breusch-Godfrey's LM test for autocorrelation. It is demonstrated by Monte Carlo simulations that a break in variance can generate spurious rejections of the null hypothesis of no serial correlation. Hence, a researcher might conclude that the error terms are serially correlated when in fact the contrary is true. It has been found that the likelihood of making this mistake depends on three factors: (i) break size, (ii) break location and (iii) the number of lagged terms included in the LM test.