Hi guys, I am currently running tests on some GARCH models. For each model I have estimated one model where all parameters are freely estimated, and then I have estimated these exact specifications, but where one parameter is restricted to the value of 1.
I want to test the impact of the restriction on the significance of the model so I calculate the loglikelihood ratio test statistic as: -2*((LLur) - (LLr)) ~ X^2 , where if the test statistic exceeds the chi-squared critical value restricted model is preferred, whereas if the the test result is not significant (e.g., the null) the unrestricted model is preferred. Furthermore, if the test result is large but negative this result is not significant - e.g., the test statistic must exceed the critical value, not just be larger in an absolute sense.
Is my logic here correct? And why do some LR tests use 2((LLur) - (LLr)) rather than -2((LLur) - (LLr))?
I apologise for asking a relatively rudimentary question, I would just like to be 100% sure in my results and have struggled to find answers/clues in the forums.
Thanks for any pointers!
I want to test the impact of the restriction on the significance of the model so I calculate the loglikelihood ratio test statistic as: -2*((LLur) - (LLr)) ~ X^2 , where if the test statistic exceeds the chi-squared critical value restricted model is preferred, whereas if the the test result is not significant (e.g., the null) the unrestricted model is preferred. Furthermore, if the test result is large but negative this result is not significant - e.g., the test statistic must exceed the critical value, not just be larger in an absolute sense.
Is my logic here correct? And why do some LR tests use 2((LLur) - (LLr)) rather than -2((LLur) - (LLr))?
I apologise for asking a relatively rudimentary question, I would just like to be 100% sure in my results and have struggled to find answers/clues in the forums.
Thanks for any pointers!
