On Jun 9, 11:19 pm, djh <halitsk...@att.net> wrote: > In the course of constructing the "control" tables for the S63-R > random dicodon set, I found two errors in the S63-R dicodon table > and "UCP" table whihc were used to obtain the results stated in my > previous post. > > Below is a restatement of the previous results after the two > corrections were made to the dicodon table and "UCP" table. > > Note that there are now no CI overlaps for slope or intercept, > regardless of which way "c" is computed. > > So I assume that these results are "better" than the results > previously stated, but again, that is merely ignorance and naivete > talking (based on absence of CI overlaps.) > > So please let us know what you think from an experienced professional > perspective ... thanks again. > > I. Results with "c" calculated as "simplified c" (original way) > > Regression: ln(c/e) on ln(c/L) > Fold: a1; Length interval 24-27 > "c" calculated as "simplified c" > S63-R S63 > > N 175 227 > Coeff 0.649108783 0.724923464 > Int 1.451994253 2.177729802 > Int CI Low (95%) 1.158243022 1.919638566 > Int CI High (95%) 1.745745485 2.435821037 > Slope 1.037876447 1.424769155 > Slope CI Low (95%) 0.855355439 1.246915339 > Slope CI High (95%) 1.220397455 1.602622970 > > II. Results with "c" calculated as "averaged c" (new way) > > Regression: ln(c/e) on ln(c/L) > Fold: a1; Length interval 24-27 > "c" calculated as "averaged c" > S63-R S63 > > N 175 227 > Coeff 0.654330479 0.748241963 > Int 1.532903339 2.259414374 > Int CI Low (95%) 1.223162987 2.005368037 > Int CI High (95%) 1.842643691 2.513460710 > Slope 1.087213506 1.472866628 > Slope CI Low (95%) 0.898660633 1.301307846 > Slope CI High (95%) 1.275766381 1.644425409
Translating those values into heteroscedastic t-tests of the differences between the two sets gives the following results:
t df p Simplified c Intercept -3.66068 372.8 .000288 Slope -2.99411 388.5 .00293
Averaged c Intercept -3.57722 359.4 .000395 Slope -2.98379 378.8 .00303
The differences between the two regression lines appear to be real, regardless of how c is defined. However, I would be more comfortable withholding final approval until I've checked the raw data for anything that might invalidate the analysis.
Repeating the analysis 71 more times can create problems due to the number of tests involved, but if the sample sizes and the differences between the lines in the 71 new sets are about as big as they are here then there should be no difficulty concluding that the differences are real. Also, even if the evidence is not as strong in some of the new sets, there are other analyses that can be done. (Doing 142 t-tests is probably the simplest way to approach things, but it is almost certainly not the best way.)