Contrast Coding

Question 1

Why use contrasts?

Answer 1

Dummy coding may not reflect our hypothesis - override what dummy coding gives you

Control error rates at 5% - significance tests are not independent of each other so error rate escalates

Question 2

What are the options for exploring differences between means?

Answer 2

Orthogonal contrasts/contrast coding:
hypothesis driven, planned what you are doing prior

Post hoc tests:
not planned, no hypothesises, compare all means, multiple t tests with adjusted p values

Trend analysis - only useful for ordered means

Question 3

What is the basic idea of planned contrasts?

Answer 3

The variability explained by the model is due to ppts being assigned to diff groups, this variability can be broken down further to test specific hypotheses - break down variance according to hypothesis made before experiment - break down SSM more, dividing up a cake

Question 4

What are the assumptions?

Answer 4

Independent - to control type 1 they must be independent contrasts, testing unique hypothesis, only use a group once. e.g. if you cut a piece of cake off, can’t re stick it

Only 2 chunks - each contrast should only compare two chunks as can make clear interpretations about the findings

K-1 - should end up with one less contrast than the number of groups you started with

Question 5

Choosing contrasts

Answer 5

Usually, the first contrast will compare any control conditions to any experimental ones

Question 6

Rules of contrast coding

Answer 6

1 - groups coded with positive weights compared to groups coded with negative weights

2 - choose the magnitude, weight assigned to the group should be equal to number of groups in opposite chunk

3 - the sum of weights for a comparison should be 0

4 - if a group isn’t involved, assign it a weight of 0

Question 7

What line do you read from in SPSS?

Answer 7

Do not assume equal variances - corrects for the amount of heteroscedasticity

Question 8

When do we use post hoc tests?

Answer 8

In the absence of specific hypothesis - comparing every mean against each other

Question 9

What are the problem with post hoc tests?

Answer 9

Not very scientific, better to be theory driven
Inflates the type 1 error - we want to only make mistakes 5% of the time but with every test, it mounts up increasing the rate. 5 groups = 10 tests = 40% error

Question 10

What is the solution for post hoc tests?

Answer 10

Be more conservative - adjust the alpha level
a = alpha level divided by number of tests

but losing power to detect differences