Seminar 4 Flashcards
List the parametric assumptions of the two-sample t-test.
DV needs to be numeric, IV needs to be a factor with two levels, the ones we test in this course are normality and homogeneity of variances.
For the scope of this course, are the parametric assumptions of ANOVA different to t-test?
No
What is an ANOVA, when do we use it?
The ANOVA is an omnibus test. We use it to check if there is an effect of an IV on a DV when the IV has more than two levels.
If we have three groups and a significant ANOVA result, does that mean we know where the differences in the groups are?
No. The omnibus test simply tells us that there is an effect, it does not tell us where that effect is. We need to follow this up with multiple comparisons (post-hoc tests).
What is the usual post-hoc test following a significant ANOVA?
TukeyHSD (by the way HSD stands for honest significant difference).
If the ANOVA is not significant, should we follow up with a post-hoc test anyway?
No. If there is no effect of IV on DV, we are not justified to check comparisons.
What methods can we use to check normality. When can we say it has been violated?
Shapiro-Wilk first. If p-value is < 0.05, that means normality assumption is violated. HOWEVER, if n > 60 you should follow up with QQPlots and check qualitatively as well.
What methods can we use to check homogeneity of variances. When can we say it has been violated?
Levene’s test. If p-value is > 0.05 that means the there are unequal variances. HOWEVER, if n > 60 you should check the box-plots as well.
If the parametric assumptions are violated, what test can we use instead of ANOVA?
Kruskal-Wallis is a good alternative and is a ranked test. This also means we can use it with ordinal data, which is already ranked.
What test should we really use as a post-hoc for the non-parametric alternative to ANOVA? What did I say I will accept for this course due to software limitations?
Dunn’s test is a good post-hoc. Unfortunately R-commander doesn’t support it, so I will accept TukeyHSD as long as you acknowledge that the most suitable test would have been Dunn’s.
Explain what a type I error is. How is it different to a type II error?
Type I (α) = probability of finding an effect when one doesn’t exist.
Type II (β) = probability of NOT finding an effect when one does in fact exist.
If we increase type I error rate, what happens to the type II error rate?
As α goes down, β goes up.
In medical research we often see confidence levels of 99%, rather than 95%, why do you think this is?
If you’re testing a new drug on patients, the last thing you want is a false positive (type I error). Imagine the lawsuit that would follow if you administer a drug that doesn’t actually work. It’s safer to make a type II error (false negative).
In the social sciences we sometimes see confidence levels of 90% rather than 95%. Why do you think this is?
Individual differences add a lot of noise to data. Because of this, it’s often hard to find effects and type II errors are more likely. Social scientists often decrease confidence levels (or used to anyway, it’s been discouraged now). To be fair, replication is social science experiments is abysmal, probably for this reason.
