8-14

Question 1

If two way anova AxB isn’t sig, should you re run and remove the interaction?

Answer 1

In a balanced design this won’t be a differenc, in unbalanced it can

Question 2

A:B and A*B in R…

Answer 2

A:B is a by b interaction

A*B is short for A, B, A:B

Question 3

How to run one factor repeated measures ANOVA? (With no replication within subject)

Answer 3

Two factors, treatment and subject

Looking at one way ANOVA will give too many df and is an example of pseudoreplication

Question 4

How to run one factor repeated ANOVA with replication within subject?

Answer 4

Look at treatment, subject and the interaction

Y~ treatment + error (subject/treatment)

Question 5

Fixed vs random effects

Answer 5

Fixed effect - interested in differences between the levels of the factor

Random - no interested in doffs between levels but them as a random sample of a population of possible levels - eg subject

Question 6

3 solutions if data correlated within subjects?

Answer 6

Summarise
Multivariate
Multi level modelling

It can adjust df in ANOVA to alllw for correlation

Question 7

How do you decide when to go multivariate?

Answer 7

Maucley’s test lf sphercity

Question 8

How does multi level model work?

Answer 8

The variances at each level in hierarchy and the correlation between them are used to estimate how much data needs to be pooled
(Averaged, so reducing df to 1 for each level)
Degree of pooling known as shrinkage

If data are no more correlated within levels than between levels there isn’t much shrinkage

Question 9

What test uses deviance as a model fit?

Answer 9

Logistic regression

Question 10

Null model vs Saturated

Answer 10

X has no effect, so proportions same for all values of x

Saturdated - model fits perfectly, proportions are allowed to vary independently for every value of x

Question 11

Generlalized linear models - testing effects via the change in deviance between a model with and without a predictor using chi sqr is more robust than using Z score for coefficient … true or false?

Answer 11

True

Question 12

WhT needs doing when checking models in parametric test

Answer 12

Check error distribution (after analysis, should be straight line- 4 graphs on R
Check homogeneity of variance
Check for independence of data points
Check model fit

Question 13

How does normal probability plot work?

Answer 13

Sorts data from lowest to highest and calculated cumulative percentage of data of each value

Question 14

What are the 4 error distribution graphs?

Answer 14

Residuals vs fitted
Normal quartile plot (QQplot)
Standardised residuals vs fitted
Residuals vs leverage

Question 15

When do you not need to do model simplification?

Answer 15

If it’s a balanced design as effects are orthogonal
Or
If it’s an experiment

Question 16

Two main methods for model simplification

Answer 16

Start with all effects and remove non sig

And

AIC

Question 17

What are information theoretic approaches?

Answer 17

Based around maximum likelihood and finding parameters that best fit the data
Don’t reject models - just weights of evidence
AIC is hard to assess support for models

Best models are the once with the right balance of predictive power and simplicity

Question 18

Advantages and disadvantages of information theoretic approaches

Answer 18

Can compare models with same number of parameters (whereas sig testing relies on adding or removing parameters)

Can present a set of models which are equally parsimonious

But .. can find best model among a load of rubbish