Originally Posted by
derjodel
Not at all. Indeed the p value (if one is to use the dreaded p value in order to accept hypothesis) should be adjusted and fairly small. In essence, we could go in blind over all data available, but adjust the p value according to degrees of freedom available to avoid false positives.
Now given that the flight data sample is huge (say, all flights in the past 3 years), the probability of false positive is actually small anyway. So is there a statistically significant difference between # of flights that did/didn't crash in the group of flights which are Boeing in low visibility on final/approach?
And btw, picking a hypothesis up front makes no difference. Imagine 20 different statisticians would decide to test a hypothesis over flight data, each unbeknown and independent of each other. Each would choose a different parameter, but p of only 0.05. Those results would not be different, but there would indeed likely be at least 1 false positive due to stupidly large p value (0.05 is really astronomical when you think about it).
I really don't want to argue with you, but using a tighter p-value is no cure for improper statistical design. You have no model for the distribution of the errors because you chose a hypothesis to fit your limited choice of data. Indeed, the hypothesis you chose fit the data exactly -- so you have a degenerate error distribution. Further, you seem to have a very limited view as to what the relevant data are. For example, the XL Air crash at Perpignan was also a nose-over -- but that was a A320. Your "model" cannot explain that at all.
https://en.wikipedia.org/wiki/XL_Air...ny_Flight_888T