Correlations Flashcards
What crucial characteristic must two tests show in order to test for correlation between them? (x1)
Why? (x1)
What is the practical consequence? (x1)
Decent variability (variance/SD) Because the variation is all that's being measured when seeing how one score is relative to another Need to design measures to yield a decent spread of scores
What happens to the correlation magnitude if score range is truncated? (x2)
May be reduced - ie give the impression the relationship is smaller than in reality
Are absolute values of scores relevant to correlations? (x1)
Why? (x2)
No
Because it effectively standardises measures to z - discarding original mean and SD
What effect does applying linear transformations to raw scores have on correlations? (x2)
None - the relationship with other variables will be unaffected
Why should you always check the scatterplots of data when running correlations? (x2)
Because correlations only represent linear relationships
But same coefficient hypothetically possible for e.g. curves, single outliers, vertical stacking of data points
What is the SPSS significance calculation for correlations? (x1)
Which is important because… (x1)
A t-test that is checking whether the correlation coefficient is significantly different from zero
We’re only ESTIMATING correlation via sampling - ie a tiny portion of the proposed population
Why is the correlation shown by SPSS only an estimate? (x1)
And why is this still ok? (x1)
But what s one side effect? (x1)
Because it’s from only a small sample
Relatively small numbers have shown to be pretty accurate of general population
Margin of error will be larger with smaller samples
What does the p value in correlation significance tell us? (x3)
HOW LIKELY it is that you will get the correlation coefficient you did, IF the “real” population correlation was actually zero
e.g. if you get .05 with correlation coefficient of 0.80, it means that there’s only 5% chance the population correlation you’re trying to estimate is actually zero rather than something bigger
(where this error is a result of sampling error – i.e. the fact you haven’t tested everyone).
What effect does the size of a correlation have on significance, and thus on required sample sizes? (x4)
Large correlation = easier to detect as larger than zero -
So can get away with smaller sample to show significance
Small correlation = need for more sensitivity to show its significance -
ie need a larger sample
What does it mean if our correlation margin of error contains zero? (x2)
It is not significantly greater than zero
So we can’t conclude that the population correlation is likely to be greater than zero
What are the consequences for the effects of sample size on correlation significance? (x1)
And how might you then decide on sample size? (x1)
Choice of sample size for given study depends on expected size of correlation (ie from prior research)
Power analysis - tells you exactly how many you need for given correlation to be significant
What are the sample sizes needed for significance for large, medium and weak correlations? (x3)
.50 needs 28 people
.30 needs 85
.10 needs 783
Does a statistically significant correlation guarantee practical importance? (x1)
Why? (x2)
No
Because if you test enough people, even a tiny difference (e.g. slightly effective intervention) can be significant
While bringing little REAL benefit
What is ‘clinical significance’? (x4)
The practical importance of any effect found through correlation
ie whether its a strong effect
As distinct from statistical significance,
Which may reflect large sample but tiny correlation
When can we use Pearson’s r? (x2)
Why is this best? (x1)
When calculating correlations of parametric variables (normal distribution, interval/ration scale)
More sensitive to variations in the data
