Friday, March 4, 2011

Statistical Power


Extract from Gravetter and Wallnau: Statistics for the Behaviour Sciences


Statistical Power


-          Purpose of hypothesis test – determine whether or no particular treatment has an effect

-          Risk of reaching wrong conclusion

o   Type 1 error

§  Reject true null hypothesis

§  Probability of type 1 error -> level of significance chosen

o   Type 2 error

§  Failing to reject false null hypothesis

§  Treatment really has effect, but hypothesis test failed to discover it.

-          Power – probability of reaching correct decision

o   Probability test will correctly reject false null hypothesis

o   More powerful  -- more readily it will detect treatment effect when one really exists

-          When treatment effect exists à two results

o   Fail to discover existing treatment effect (type 2 error)

o   Correctly detect – and reject false null)

-          Power = 1 – β

-          Null hypothesis is rejected whenever sample data is in critical region.

 

Power & Size of Treatment Effect

 

-          Power depends on size of treatment effect

-          Large effect – easy to detect – power will be high

-          Rather than refer to power as single value – examine different values of power associated with different magnitudes of treatment effects.

-          Critical region

 

 

-          Alpha level

o   Reducing alpha level reduces power of test

§  0.05 to 0.01

§   Z value changes from 1.96 to 2.58

o   Critical boundary moves to the right – therefore this reduces critical region where null hypothesis is rejected

o   Less of the treatment distribution

 

 

 

 

-          One tailed vs two tailed

o   Changing from regular two tailed test to one tailed test

o   With 2 tailed and α = 0.05, critical region on rhs begins at z = 1.96

o   Changing to 1 tailed, critical value moves left to a value of z = 1.65

o   Therefore increasing critical region

 

 

-          Sample size

o   Larger the sample, the better it will represent sample

o   Increased sample size deuces std error

 

No comments: