Corrections (generally for non-orthogonal) Flashcards
Boole’s inequality
The probability of one of a set of events can never exceed the sum of their individual probabilities
Bonferonni set bounds on this:
e.g. 3 comparisons, each with probability of Type 1 error a=.05
so overall the probability of a Type 1 error CANNOT exceed (.05+.05+.05)= .15
Bonferonni correction
a/c (c= number of comparisons)
e.g. a=.05, c=5 so .05/5=.01 use this as the significance threshold (a=.01)
Problems with bonferonni correction
:( the more comparisons done, the lower the alpha is so it is harder to observe significant effects (reduced Type 1 error but increases Type 2)
:( it can be overly conservative
Holm procedure (multistage bonferroni)
For controlling the FW error rate
1) Scale t for all contrasts, arrange the t values in rank order
2) Check the largest first against the critical value in Dunn’s (1961) table corresponding to contrasts (=a/c)
3) if it is significant, the next largest t has an alpha based on c-1
E.g. 5 groups:
1st: a=.05, c=5, so .05/5=.01 (like bonferonni)
2nd: a=.05, c=4, so .05/4=.0125
3rd: a=.05, c-3, so .05/3=.0167
etc. stop when a non significant result is researched (minimised number of analyses, might not do all 5)
Benefit of the Holm procedure
:) more LIBERAL threshold for each successive comparison, you are more likely to find a significant effect
when to use FDR
When there are lots of significant effects, it is better to control proportion of errors (false discovery rate), rather than the probability of making one error (familywise error rate)
FDR (false discovery rate)
Used to control for FW error rate (Benjamini & Hochberg, 1995)
V= number of true null hypotheses rejected (false positives) R= total number of nulls rejected FDR= V/R if R>0
If R=0 (there are no significant effects) so don;t bother with it
Where does the new criterion come from in FDR
The new criterion is based on number of tests significant, relative to their rank.
1) rank p values in order
2) look at the table with ranked p values, find where p
advantages and disadvantages of FDR
:) more power than FW error corrections
:) you can bear more errors when there are more nulls rejected
:( higher rates of a Type 1 error than FW error corrections