Analysis of Power
Download
Report
Transcript Analysis of Power
In SAS
March 11, 2013 - Mariya Cheryomina
power (π) + β = 1
β = (1- π) = probability of accepting false Ho
(ie. reject true Ha)
- probability of Type II error
– false positive
Power (π) = (1- β) = probability of detecting
a difference when a difference does exist
- probability of accepting true Ha (ie. reject
false Ho)
– how sensitive your test is to the existing
difference between the compared samples
Generally, the minimal sufficient (acceptable)
value of power is 0.80
π
≥ 0.80
1) Before gathering data
To determine the minimal sample size
needed to have desired power in statistical
testing (to detect a particular effect size)
2) After gathering data
To determine the magnitude of power that
your statistical test will have given the sample
parameters (n and s) and the magnitude of
the effect that you want to detect
Sample size (n)
Standard deviation (s)
Alpha level (α )
Size of effect/difference that you want to
detect
Type of statistical test performed
One-sample t-test
One-way Anova
proc power;
onesamplemeans
mean = ____
ntotal = ____
stddev = ____
power = ____ ;
run;
(type of statistical test you want the power to be calculated for)
(difference you are interested in detecting)
(sample size)
*One of the four variables must be left blank – this is what you want SAS to
calculate
You are asked to determine whether the use of a new soil type
leads to a significantly different average height of young pine
trees on plantation (compared to the historical/hypotherical
mean height of 110cm recorded with the old soil type). You
want to conduct a one-sample t test with a 2-sided α = 0.05.
You select150 trees (n). You decide that the minimal difference
in height worth addressing is 8cm (effect size)
α = 0.05
s = 40
n = 150
“mean” = 8
(effect size)
What will the power of your statistical test be?
Ho: = 110cm
Ha: ≤ 102cm OR ≥ 118cm
SAS text:
proc power;
onesamplemeans
mean = 8
ntotal = 150
stddev = 40
power = .;
run;
*Unless you indicate otherwise, SAS will automatically assume that α = 0.05
There is a there is a
probability of 0.68 that the
t test will produce a
significant result indicating
a difference in mean tree
height of at least 8cm
Sample size
Standard deviation
You are interested in finding out how the changes in samples size (n),
standard deviation (s) and minimal effect size of interest (“mean”)
will effect the power of your one-sample t- test:
proc power;
onesamplemeans
mean = 5 10
ntotal = 150
stddev = 30 50
power = .;
plot x=n min=100 max=200;
run;
Increasing sample size does
not infinitely increase power
Effect size
Example 2:
You want to compare the average diversity of Canada’s native
bee species in four types of habitat:
1)
Urban
2)
Agricultural : Monoculture plantations (with pesticide use)
3)
Agricultural : Organic farms (no pesticide use)
4)
National parks
Ho: μ1 = μ2 = μ3 = μ4
Ha: one of μ is different
What is the minimal number of sites of each
habitat type that needs to be surveyed (ie.
minimal sample size of each group) to detect
whether a significant difference in bee species
diversity exists between any of the four
habitat types?
(using a one-way Anova test with α = 0.05)
The desired power of 0.9.
*Analysis of power usually involves a number of simplifying
assumptions
Assumptions:
Average number of bee species surveyed per site in Canada is
~35 species with a standard deviation of ~ 10 species
Based on available research, you predict the following average
bee species diversity for each habitat type:
1)
Urban - 25 species
2)
Monoculture plantations - 30 species
3)
Organic farms– 40 species
4)
National parks – 45 species
*Assume that all groups have the same stdev (s)
All numbers used in example were invented for
the purpose of the exercise
proc power ;
onewayanova
groupmeans = 25| 30 | 40 | 45
stddev = 10
alpha = 0.05
npergroup = .
power = .9;
run;
SAS finds the group
sample size that
gives a power
(actual power)
closest to the power
you desire (ie. to
the nominal power)
You will need to survey at least 7 locations with each of the
four habitat types (ie. 7 urban sites, etc.) to detect the desired
(significant) difference between the mean diversities of bee
species found at each of the habitat types
Now you are trying to detect a
smaller difference between
sample means (at a significant
level)
Observe the new minimal
sample size for each group
Sample size (↑)
X approaches when n approaches N
Standard deviation (↓)
Difference between samples is less likely to cur simply due to random
sampling effects
(α ) (↑)
Higher α leads to lower β which results in higher power (ie. the more
willing you are to reject Ho, the less likely you are to accept false Ho,
which leads to a higher probability of detecting a truly existing
significant difference)
Minimal effect size (↑) (difference that you want to detect)
A large difference between samples is less likely to occur due to random
variability between samples than a small difference is
When you are expecting a large effect size,
but are not fully confident that the true effect
is as large, use a larger sample size (ie. one
that the analysis of power suggests for
detecting a smaller effect )
Analysis of power for Anova (in SAS):
http://www.ats.ucla.edu/stat/sas/dae/fpower.htm
Analysis of power for one-sample and two-sample t-tests (in
SAS):
http://support.sas.com/rnd/app/papers/power.pdf