PROFILE ANALYSIS
Profile Analysis Main Point: Repeated measures multivariate analysis One/Several DVs all measured on the same scale
Profile Analysis Main Point: Most commonly used as a time series design Measured several times on the same DV
Profile Analysis Main Point: Doubly multivariate – several different DVs are measured over time “Doubly” because there are double layers, or multiple DVs measured a couple times
Research questions: Mainly: Do people have different “profiles” on a set of measures
One Issue Measures much have the same range of scores with the values having the same meaning Because test of profiles measure the differences in adjacent DVs for that “time” measurement Difference scores are called segments
Profile Parts Parallelism profiles Do the different groups have different parallel profiles ANOVA comparison = interaction
Profile Parts Levels: Overall group differences – regardless of parallelism, does one group on average have a higher score on the collected set of measures? Between subjects ANOVA analysis
Profile Parts Flatness – similarity of responses on the DV independent of group Do all the DVs (or times of the DV) elicit the same average response?
Profile Parts Contrasts after profile – if you get differences then you have to follow up with a type of contrast analysis
Examples and Follow Ups Example data – I have a class I’ve taught a couple times Class 1 Quiz 1 Quiz 2 Quiz 3 Quiz 4 Quiz 5 Quiz 6 Quiz 7 Class 2 Quiz 1 Quiz 2 Quiz 3 Quiz 4 Quiz 5 Quiz 6 Quiz 7
Example
Class 1 Quiz 1 Quiz 2 Quiz 3 Quiz 4 Quiz 5 Quiz 6 Quiz 7 Class 2 Quiz 1 Quiz 2 Quiz 3 Quiz 4 Quiz 5 Quiz 6 Quiz 7 Parallelism = interaction – do the lines cross? Levels – between these two are they different? Flatness – are these the same over time?
Limitations - Theoretical Choice of DV Limited to scales that are the same Easy to use when you are repeated the same scale over and over
Limitations - Theoretical Choice of DV If the units are not the same you can convert to z- score Differences in profiles attributed to the differences in group treatments Causal if you have manipulated them.
Limitations – Practical Sample size – use a between subjects anova analysis if you don’t have a program that will run multivariate program More people in the smallest group than there are DVs Rule of thumb is 10 cases to 1 on DVs
Limitations - Practical Repeated measures ANOVA has more power Collecting more data points from the same people, so that reduces error Error is controlled with in person, instead of with in group Still need more people than a univariate analysis
Power Usually a little stronger – you have to deal less with Sphericity With g*power – you can do this as a regular repeated measures – but you will need to run more people than regular repeated measures with very small effect sizes
Limitation - Practical Unequal N isn’t a big deal Also harder to have because you measure people several times, ends up being missing instead of unequal
Missing Data Special imputation because it’s missing See page 345 Basically involves summing and averaging the scores that you do have for the person, and then averaging the other scores from everyone else Or you can do a HLM (hierarchical linear model) if imputing scores is not a good idea (cancer study)
Normality Robust! Check! Unless there are fewer cases in a cell than there are DVs
Outliers All DVs get outlier analysis Could do it for each time segment
Homogeneity If sample sizes are equal, homogeneity of variance is not necessary since all scores came from the same person Box’s M still is applicable p<.001
Linearity For parallelism and flatness, you are assuming linearity since you are checking if the lines are flat or cross You use bivariate charts to get combos of the DV
Multicollinearity – Singularity But we want our DVs to be correlated because they are all measured from the same people?! Statistically will not run when R2 value research.999
Issues Univariate versus multivariate Sphericity – the correlation between each time measurement must be the same With a multivariate test you will never meet this assumption With only two levels of the IV, not a big deal
Fixes Greenhouse-Geisser or Huynh-Feldt – are adjustments given automatically for violations Adjusts the significance values to be more conservative Or you could lower your alpha rate (so you need a lower p value) but then you lose power
Issues Univariate versus Multivariate Do both! If they give you same result, then report univariate (much easier!) Trend analyses – do this instead if it makes sense with your data