Applied
-
This problem makes use of the
Carseatsdataset in theISLPpackage.-
(a) For each quantitative variable in the dataset besides
Sales, fit a linear model to predictSalesusing that quantitative variable. Report the p -values associated with the coefficients for the variables. That is, for each model of the form Y = β 0 + β 1 X + ϵ , report the p -value associated with the coefficient β 1. Here, Y representsSalesand X represents one of the other quantitative variables. -
(b) Suppose we control the Type I error at level α = 0 . 05 for the p -values obtained in (a). Which null hypotheses do we reject?
-
(c) Now suppose we control the FWER at level 0 . 05 for the p -values. Which null hypotheses do we reject?
-
(d) Finally, suppose we control the FDR at level 0 . 2 for the p -values. Which null hypotheses do we reject?
-
-
In this problem, we will simulate data from m = 100 fund managers.
rng=np.random.default_rng(1)
n,m=20,100
X=rng.normal(size=(n,m))
596 13. Multiple Testing
These data represent each fund manager’s percentage returns for each of n = 20 months. We wish to test the null hypothesis that each fund manager’s percentage returns have population mean equal to zero. Notice that we simulated the data in such a way that each fund manager’s percentage returns do have population mean zero; in other words, all m null hypotheses are true.
-
(a) Conduct a one-sample t -test for each fund manager, and plot a histogram of the p -values obtained.
-
(b) If we control Type I error for each null hypothesis at level α = 0 . 05, then how many null hypotheses do we reject?
-
(c) If we control the FWER at level 0 . 05, then how many null hypotheses do we reject?
-
(d) If we control the FDR at level 0 . 05, then how many null hypotheses do we reject?
-
(e) Now suppose we “cherry-pick” the 10 fund managers who perform the best in our data. If we control the FWER for just these 10 fund managers at level 0 . 05, then how many null hypotheses do we reject? If we control the FDR for just these 10 fund managers at level 0 . 05, then how many null hypotheses do we reject?
-
(f) Explain why the analysis in (e) is misleading. Hint: The standard approaches for controlling the FWER and FDR assume that all tested null hypotheses are adjusted for multiplicity, and that no “cherry-picking” of the smallest p-values has occurred. What goes wrong if we cherry-pick?