1/13
Looks like no tags are added yet.
Effect sizes
indicates the proportion of the variance explained (what the effect is, or how big it is)
measures of effect size
R² (correlations and regression), cohen’s d (t-tests), eta squared, partial eta squared
Cohen’s d
d is the difference in means (scaled by the SD), is the difference between the largest and smallest group means scaled by SD and homogeneity of variance assumption is the same
Eta squared
used for one-way ANOVA, same as R², proportion of variance explained by your experiment
partial eta squared
for use with factorial ANOVA, proportion of variance uniquely explained by each IV
reporting effect sizes
you can have a sig effect but a very small eta squared
‘power’ of a statistical test
ability to detect an effect when it’s actually there, ability of a test to correctly reject the null hypothesis
power depends on
sample size (n), effect size (d), critera for significance (alpha level); want to make sure we have enough power to detect an effect
dangers of underpowered studies
too few participants (lack of power to detect effect/type 2 error or increased chance of type 1 error), explains failure to replicate
power as a function of sample size
sample size incresases so does power, more participants=inc chance of finding sig effect if there is one (power of 0.8 is typically good)
ways to estimate effect size
guess (not very approximate), pilot study (run pilot with fewer participants to estimate effect size, can use to project how many participants you should test), find previous studies (use their results to work out expected effect size for your experiment, can’t predict the future but better than guessing), meta-analysis
problem with power analysis
making up numbers isn’t meaningful, estimates not exact, for really complex designs it’s unlikely you’ll have sufficient precision in your estimates of effect size
types of power analyses
calculate how many participants are required for a study, calculate min req effect size detectable, calculate observed power, in G*Power
observed power, what if the results of a study are a false negative?
can run post-hoc power analyses