2018 Fall Q6 F Statistics
Hello,
Based on the F Statistic formula, The numerator is the Unscaled Deviance =phi * 2*(ll sat - ll model) .
Therefore in this question Ds = 2(-1000-(-1500))*1.75 = 1750, and Db = 2(-1000-(-1465))*1.75 = 1627.5
and F = (1750-1627.5)/(2*1.75) = 35
The sample answer seem to be using the scaled deviance, leading to an answer of 20. Would you please advise which is correct, the paper or the question? Or is there a reason why the scaled deviance is used in this question instead of the unscaled deviance.
Thank you lots!
Comments
Hi,
This question was defective/ambiguous on the 2018 exam as the source text at the time didn't clearly define scaled and unscaled deviance. The correct F-test answer uses unscaled deviance. This is noted in part a of the Examiner's Report for this question.
Remember, the F-test should use unscaled deviances when comparing against the critical value.
However, in this case because the estimated dispersion parameters for models A and B are the same, we could compare the scaled deviances for Model A and Model B to determine which of Model A and Model B is best but we can't use the critical value provided to perform an F-test in this situation.
Thank you!