Lecture 7

Question 1

type I error

Answer 1

• percentage of the null distribution that falls beyond the critical cut off
• 5% of the time (or whatever your α is) when H0 is true.

Question 2

power

Answer 2

• When H0 is false, power is the ability to detect a significant outcome.
• essentially the percentage of the
  alternative distribution that falls beyond the critical cut off
• want a minimum of 80%

Question 3

types of power

Answer 3

a priori
post hoc

Question 4

a priori power

Answer 4

• calculated before collecting data.
• typically involves figuring out the
  expected effect size and sample size for 80% power.

Question 5

post hoc power

Answer 5

• calculated after the statistical result, usually if it is not significant.
• It involves figuring out how likely we were to reject H0 if it
  were actually false.
• From here, we usually assess if our sample size was
  adequate.
• usually not evaluated if the results are significant as, by default, we can consider it to be equal to 1

Question 6

ways to increase power

Answer 6

• decrease spread
• increase group differences
• increase sample size
• increase alpha (cheating)

Question 7

parametric tests

Answer 7

tests whose values can
be generalized to the population.

Question 7

why can we use effect size to determine power

Answer 7

• power is a 3 way relationship between magnitude of the difference between your means, the noise (variance), and sample size
• so, can use effect size because it contains 2/3 elements for power (magnitude of the difference between means, and noise/variance)

Question 8

speed-accuracy tradeoff

Answer 8

that the faster one does
on a task, the less accurate they tend to be

Question 9

what to do when an assumption if violated?

Answer 9

• transformation, which modifies the
  data to try and fit the assumption.
• controversy bc you are technically changing your data (like squaring or square rooting it)
• applying corrections to degree of freedom is best and easiest method
• can also artificially increase sample size (jackknifing & bootstrapping)
• use nonparametric tests (as they don’t rely on parametric assumptions)

Question 10

boostrapping

Answer 10

repeatedly subsampling from your data (with replacement) and
creating a new, artificially enlarged, sample size (e.g., 10,000 samples of n – 1)

Question 11

jackknifing

Answer 11

involves increasing the sample size by taking all combinations of n – m
samples (usually m = 1)

Question 12

nonparametric tests

Answer 12

• Values from these tests are not expected to generalize to population parameters.
• That does not mean that
  the results or inferences cannot generalize
• These tests rely on non-continuous data, ordinal or nominal. Continuous data that violates an assumption
  is transformed to non-continuous data

Question 13

Mann-Whitney-U

Answer 13

• equivalent of the independent samples t test, using ordinal-rankings
• Where continuous data is no
  longer valid/generalizable, it transforms data into categorical ranks
• This removes effects of outliers, nonhomogeneous variances, issues
  of normality, and other issues that can affect the spread of the data

Question 14

Wilcoxon Signed Rank

Answer 14

• equivalent of paired-sample t test, using ordinal-rankings
• works very similar to
  MWU. It involves ranking difference scores