Chapters 22-23 Flashcards

1
Q

The problem of multiple comparisons

A

the most difficult challenge of statistics

-it is impossible to interpret P values without knowing how many comparisons

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
2
Q

Assuming two comparisons, what is the chance of obtaining at least one statistically significant conclusion (false positives) by chance?

A

1-0.9025=0.0975

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
3
Q

Assuming two comparisons, what is the chance of obtaining zero statistically significant conclusion (false positives) by chance?

A

0.95x0.95=0.9025

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
4
Q
With K independent
comparisons,
• The chance that
none will be
significant
A

is 0.95^K

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
5
Q

The chance that one or
more will be significant
when null is actually
true =

A

1- 0.95^K

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
6
Q
Multiple
comparisons can
be used to
generate
hypotheses,
• But not
A

in
experiments to
test hypotheses.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
7
Q

Simply report uncorrected P values but clearly state

A

how many comparisons were made and that no

corrections were made to P values.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
8
Q

Or designate one test as the

A

primary outcome and
all others (secondary outcomes) as exploratory
analyses.
• must be a priori.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
9
Q

Or if all of the comparisons produced similar

significant results,

A

Then all would lead to the same conclusion
whether just one or many comparisons were
performed.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
10
Q

Familywise Error Rate

A
Previously, α was
used to represent
the Type I Error rate
per comparison.
• With multiple
comparisons, we
need a familywise
error rate.
How well did you know this?
1
Not at all
2
3
4
5
Perfectly
11
Q

Here the significance level is redefined to be the

A

chance of obtaining one or more statistically

significant conclusions if all null hypotheses are true.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
12
Q

Bonferroni Correction

A
The most common approach is
to divide the value of α by the
number of comparisons.
• For 20 comparisons:
• 0.05/20 = 0.0025
• i.e. α /K (where K is the
number of comparisons)
How well did you know this?
1
Not at all
2
3
4
5
Perfectly
13
Q
Now there is only a 5% chance
of seeing (with a Bonferroni correction)
A
one or more
significant results (if all nulls
are true) in the whole study.
How well did you know this?
1
Not at all
2
3
4
5
Perfectly
14
Q

Defining Groups With the Data

A
Groups must be
defined before
seeing the data.
• If defined by the
data, then multiple
comparisons have
already occurred.
How well did you know this?
1
Not at all
2
3
4
5
Perfectly
15
Q
When something unusual
happens, it is tempting to
calculate the probability
that this event could
happen in this place.
• But what you really need to
know is
A

the probability that
this event could happen in
one of the places that you
can observe.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
16
Q

Researchers may
develop a model to
explain an event
using

A

multiple
explanatory
(independent)
variables).

17
Q

If you choose individual explanatory variables from a large set of
candidates, and then just use the best in the model, you have
performed multiple comparisons.

A

Some of those variables were

probably associated with the dependent variable just be chance.

18
Q
If you keep adding
samples until you get
a significant result,
but then stop adding
samples every time
you get significant
results.
A

You have bias the results in favor if Type I Errors.

19
Q

Data Torture/Data Diving

A
Example:
• Just keep running
different analyses
until you get a
significant result.
• This is multiple
comparisons.
The chances of a Type I Error are substantially
increased.