2023 Lecture6 Part1

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 48

Repeated-measures designs and

factor analysis: Part 1

Pre-Master course Business Research Methods


Dr. Kristin Kronenberg
Agenda
• Repeated-measures designs
• Factor analysis
Agenda
• Repeated-measures designs
• Factor analysis
Why conduct a repeated-measures ANOVA?
• Same entities contributing to the different means
- Students' performance at different points in time
- A group of travelers individually rating all accomodations
they visited together
Why conduct a repeated-measures ANOVA?
Why conduct a repeated-measures ANOVA?
• Same entities contributing to the different means
- Students' performance at different points in time
- A group of travelers individually rating all accomodations
they visited together

• Also known as
- Within-participants design
- Within-subjects design
- Related design
Looking back: the t-test
• Different entities
- Independent t-test (independent-measures t-test;
independent-means t-test)
- Participants who received actual medication vs. those who
received a placebo
- Same logic: independent design
• Same or related entities
- Paired-samples t-test (dependent t-test)
- Students' knowledge before and after this lecture
- Same logic: repeated-measures design
Looking back: the t-test
• Different entities
- Independent t-test (independent-measures t-test;
independent-means t-test)
- Participants who received actual medication vs. those who
received a placebo
- Same logic: independent design
• Same or related entities
- Paired-samples t-test (dependent t-test)
- Students' knowledge before and after this lecture
- Same logic: repeated-measures design
Example: somewhere in the jungle
• I'm a celebrity, get me out of here
• 8 contestants
• 4 animals (or parts)
• Outcome: time to retch
The data
Stick Kangaroo Fish Witchetty
Celebrity Mean s2
Insect Testicle Eye Grub
1 8 7 1 6 5.50 9.67
2 9 5 2 5 5.25 8.25
3 6 2 3 8 4.75 7.58
4 5 3 1 9 4.50 11.67
5 8 4 5 8 6.25 4.25
6 7 5 6 7 6.25 0.92
7 10 2 7 2 5.25 15.58
8 12 6 8 1 6.75 20.92
Mean 8.13 4.25 4.13 5.75
The ANOVA approach to repeated-measures designs

• Basically, a linear model


• Residuals are (obviously) not independent, but affected by
- Between-participant factors (should be independent)
- Within-participant factors (not independent)
• Solution: apply additional assumptions to 'allow for' the
application and interpretation of the model
Problems with the ANOVA approach
• Same participants in all conditions
- Scores across conditions correlate
- Violates assumption of independence
• Assumption of sphericity
- Crudely put: the correlation across conditions should be the
same
- If not, adjust degrees of freedom and 'make it harder' to get
a significant F-statistic
The assumption of sphericity

Sphericity: a measure of how closely


the shape of an object resembles that
of a perfect sphere
The assumption of sphericity
• In statistics: sphericity means that the correlation between
treatment levels is the same
• This means that the variances of differences between
conditions are equal
• Tested using Mauchly's test
• Estimated and adjusted for using the
- Greenhouse-Geisser estimate
- Huynh-Feldt estimate
Assessing sphericity

Sphericity is assumed if the variances are roughly equal


Assessing sphericity
• Mauchly's test
- Are variances of the differences between conditions equal?
- If p < 0.05, sphericity is violated
- If p > 0.05, sphericity is met
- Depends upon sample size, thus questionable results
What to do if you cannot assume sphericity?

• Measures
- Greenhouse-Geisser estimate
- Huynh-Feldt estimate
What to do if you cannot assume sphericity?

• Measures
- Greenhouse-Geisser estimate
- Huynh-Feldt estimate
• Multiply df by these estimates to correct for the degree
of non-sphericity
• G-G is conservative, and H-F liberal (always check both)
The F-statistic for repeated-measures designs
• In a repeated-measures design, the effect of the experiment
(the independent variable) is shown up in the within-
participant variance (rather than in the between-group
variance)
• Experimental manipulation on same entities
• Within-participant variance made up of
- Individual differences
- Effect of experiment
The F-statistic for repeated-measures designs
• In a repeated-measures design, the effect of the experiment
(the independent variable) is shown up in the within-
participant variance (rather than in the between-group
variance)
• Experimental manipulation on same entities
• Within-participant variance made up of
- Individual differences
- Effect of experiment
The ANOVA approach to repeated-measures designs

In a repeated-measures design, both 𝑺𝑺𝑴 and 𝑺𝑺𝑹 are


parts of the within-participant variance
The total sum of squares (𝑺𝑺𝑻 )
• Same logic as before
• The grand variance is the variance of all scores, whichever
group they came from

𝑺𝑺𝑻 = 𝑠 2𝑔𝑟𝑎𝑛𝑑 𝑁 − 1

• Here: N = 32 (8 participants, 4 animals)

𝑺𝑺𝑻 = 𝑠 2𝑔𝑟𝑎𝑛𝑑 𝑁 − 1 = 8.19 32 − 1 = 𝟐𝟓𝟑. 𝟖𝟗

• df = N - 1 = 31
The within-participant sum of squares (𝑺𝑺𝑾 )
• This is what makes the repeated-measures design different
from an independent design
• 𝑺𝑺𝑾 represents individual differences within participants
• In independent designs, these individual differences were
quantified with 𝑆𝑆𝑅
𝑘 𝑛 𝑘

𝑺𝑺𝑹 = ෍ ෍(𝑥𝑖𝑔 − 𝑥𝑔ҧ )2 = ෍ 𝑠 2𝑔 𝑛𝑔 − 1


𝑔=1 𝑖=1 𝑔=1

• Different participants in each condition, therefore calculate


𝑆𝑆𝑅 in each condition and add up to total
The within-participant sum of squares (𝑺𝑺𝑾 )
• In repeated-measures designs, we want to know about the
variation within an entity (not within a condition)
• Same equation, but adapted to look at entities

𝑺𝑺𝑾 = 𝑠 2 𝑒𝑛𝑡𝑖𝑡𝑦 1 𝑛1 − 1 + 𝑠 2 𝑒𝑛𝑡𝑖𝑡𝑦 2 𝑛2 − 1 + ⋯ + 𝑠 2 𝑒𝑛𝑡𝑖𝑡𝑦 𝑛 𝑛𝑛 − 1

• Calculate variation in each individual's scores, then add up


these variances for all entities, here:

𝑆𝑆𝑊 = 236.50

• df = N – 1 =3 for each entity, and df = 3 * 8 = 24 in total


The model sum of squares (𝑺𝑺𝑴 )
• Which part of 𝑆𝑆𝑊 can be attributed to our experimental
manipulation? (This is why we conducted the experiment.)
• We look at the differences between group (animals) means
and the overall mean (as in an independent design)
𝑘

𝑺𝑺𝑴 = ෍ 𝑛𝑔 (𝑥𝑔ҧ − 𝑥𝑔𝑟𝑎𝑛𝑑


ҧ )2
𝑔=1

• 𝑆𝑆𝑀 = 83.13
• df = k - 1 = 4 (conditions, here: animals) - 1 = 3
The residual sum of squares (𝑺𝑺𝑹 )
• 253.58 units of variation to be explained
• Variation across conditions accounts for 236.50 units
• Experimental manipulation explains 83.13 units
• How much variation cannot be explained by the model
(and is therefore outside experimental control)?

𝑺𝑺𝑹 = 𝑆𝑆𝑊 − 𝑆𝑆𝑀 = 236.50 − 83.13 = 𝟏𝟓𝟑. 𝟑𝟕

• 𝑑𝑓𝑅 = 𝑑𝑓𝑊 − 𝑑𝑓𝑀 = 24 − 3 = 21


The mean squares

𝑆𝑆𝑀 83.13
𝑀𝑆𝑀 = = = 27.71
𝑑𝑓𝑀 3

𝑆𝑆𝑅 153.37
𝑀𝑆𝑅 = = = 7.30
𝑑𝑓𝑅 21
The F-statistic
• Remember: ratio of variation explained by the model and
unexplained variation

𝑀𝑆𝑀 27.71
𝐹= = = 3.79
𝑀𝑆𝑅 7.30

• Compare against critical value (see Appendix)


• Thank you, SPSS ☺
And what about 𝑺𝑺𝑩 ?
• Between-participant sum of squares

𝑺𝑺𝑩 = 𝑆𝑆𝑇 − 𝑆𝑆𝑊 = 253.89 − 236.50 = 𝟏𝟕. 𝟑𝟗

• Individual differences between cases/participants


Assumptions in repeated-measures designs
• Linear model with all ist assumptions
• Assumption of independence
• Sphericity
What does a suitable dataset look like?
What does a suitable dataset look like?
Repeated-measures ANOVA in SPSS
Repeated-measures ANOVA in SPSS
Repeated-measures ANOVA in SPSS
Repeated-measures ANOVA in SPSS
The IBM knowledge center
Repeated-measures ANOVA in SPSS
Repeated-measures ANOVA in SPSS
Descriptives
Correcting for sphericity
Correcting for sphericity

• If data are perfectly spherical, measures will be 1


• They are not
• Correct for non-sphericity
• Correct df for F-statistic
• Increases p-value
Correcting for sphericity

• df for the model sum of squares: 3


• df for the residual sum of squares: 21
• Adjust these values by multiplying them by the
estimate of sphericity (e.g. 0.533)
• F-statistic unchanged, but p-value increases
The F-statistic
Contrasts
Post hoc-tests
Reporting
• Same as for independent designs
• Additionally: pay attention to reporting the corrected df
• The Greenhouse-Geisser estimate of the departure from
sphericity was 𝜀 = 0.53. The time to retch was not
significantly affected by the type of animal eaten, F(1.60,
11.19) = 3.79, p = 0.063.
• The Huynh-Feldt estimate of the departure from sphericity
was 𝜀 = 0.67. The time to retch was significantly affected
by the type of animal eaten, F(2, 13.98) = 3.79, p = 0.048.
…have a short break…

You might also like