This web page is for error corrections and other changes to the slides and homeworks.
- (posted Nov 17). Simplified
notes on Chapter 8
by using R function
loglm
in R packageMASS
rather than R functionloglin
in base R. - (posted Nov 16, after class). An error noticed in class about which models are on the AIC and BIC lists for notes about glmbb, seat belt use example. It said something wrong about where the best according to BIC was according to AIC. Now fixed.
- (posted Nov 09, after class). An error noticed by a student in class
class Wednesday, Nov 11 in the list of models selected
by R function
glmbb
in the notes on Chapter 8 in Agresti one model was listed as ∅ when it should have been L ∗ S. Fixed. - (posted Nov 09, after class). There were several errors and infelicities noticed in class Monday, Nov 09. All four have now been fixed.
- (posted Oct 21, after class). There was an error in the the addendum to the lecture notes for Agresti Chapter 4 where test = "LRT" was inadvertently left off the code for the tests. Now fixed.
- (posted Oct 14, after class). There was an error found by a student in the section about Rao confidence intervals for the Poisson distribution. The first two displayed equations in that section are OK, but in the third displayed equation the right-hand side should obviously be μ zα ⁄ 22 and it wasn't (lost the square). And this has consequences for all the equations and computer code in this section. Now fixed.
- (posted Sep 25, after class). There was an error found in questions
after class and in an e-mail from a student seen after class, the errors
were different but the issue was the same. R function logl
defined in the
section about graphing the log likelihood
was broken. It has now been fixed. The problem is that in this context
we should define 00 = 1 and 0 log(0) = 0. And R does not
> 0 * log(0) [1] NaN
although Mathematica does if asked in the right wayIn[1]:= Limit[x Log[x], x -> 0] Out[1]= 0
Hence our repaired R function calculates as before but then changes the result from NA or NaN to 0. Then it returns this fixed up result.Sorry I couldn't figure all this out in class. I had to stop talking and think. Now we have a function that actually always works (I hope).
- (posted Sep 25, after reviewing the recording). In answer to a question
about the treatment effect I mistated what it is.
The treatment effect when discussing a mean is
δ = (μ − μ0) ⁄ σwhere μ is the true unknown mean (population mean), σ is the true unknown standard deviation (population standard deviation), and μ0 is the value of μ hypothesized under the null hypothesis. The reason for talking about the
treatment effect
instead of the three things that go into it is because the power of the test only depends only on δ rather than those three things separately. Then the estimated treatment effect isˆδn = ( n − μ0) ⁄ sn