Transcript Feb. 27-- Experimental design





Graduate symposium deadline Friday
CSBS student conference April 25 (deadline
April 10)
Thesis defenses
Outline due on Friday



What is internal validity? How can it be
increased?
What are the relationships between internal,
external, and construct validity?
What is needed to establish causality?







History
Maturation
Testing
Instrumentation
Mortality
Regression to the mean
What are each of these (examples) and how
can they be decreased/controlled?



Why does it happen?
When will it happen more?
How can you estimate it?








Selection
Selection-history
Selection-maturation
Selection-testing
Selection-instrumentation
Selection-mortality
Selection-regression
What are each of these (examples) and how
can they be decreased/controlled?





Diffusion of treatment
Compensatory rivalry
Resentful demoralization
Compensatory equalization of treatment
What are each of these (examples) and how
can they be decreased/controlled?


X O
R O X O
R O
O
N O X O
N O
O
N
O




vs. random selection
What is the strongest design in terms of
internal validity threats?
Why would you want to add a pretest? Or
not?
Are there times that you wouldn’t want to
randomly assign?




Independent vs. dependent variables
Within vs. between-participants designs:
advantages?
What are control variables?
What can you do to reduce confounds in
research?



What are they and why do we care?
What are levels vs. variables?
What kind of FD is this? How many main effects
are possible? How many interactions? How
many groups? How many people at n=10?
 IVs: Presence of an audience, difficulty of task (easy,
medium, or hard). DV = nervousness
 IVs: Gender, Drug vs. CBT vs. control, previous
experience with treatment (yes or no). DV=reduction
in symptoms


People who were alone helped more often
than those who were with a confederate.
Seminary students also helped more often
than business students.
Men and women helped equally overall, but
men were more likely to help women than to
help men.








Null results
Main effect
Interaction
Block design
Covariate
Solomon 4-group
Switching design
ANOVA vs. ANCOVA vs. regression



If a factorial design is good, is a bigger one
better?
When would you want to include a variable as
an IV vs. a covariate vs. something you match
on vs. a control variable?
How do you decide which variables to
include?


What is a moderator?
How is a moderator different from a/an






IV?
DV?
Confound?
Mediator?
Covariate?
How can you design a study to test a
moderator (various ways)?


What is a mediator?
How is a mediator different from a/an






IV?
DV?
Confound?
Moderator?
Covariate?
How can you design a study to test a
mediator (various ways)?

A study finds that convicted criminals are more likely than noncriminals to score
low (negatively) on the Attitudes toward Women Scale. Further research shows,
however, that this is only true for violent criminals. There is no relationship
between nonviolent criminal activity and negative attitudes toward women.
Identify the mediator or moderator.

2. A researcher finds that by increasing self-focus in children, she can decrease
their likelihood of cheating on a test. By decreasing cheating, in turn, she finds
that academic self-efficacy increases. Identify the mediator or moderator.

3. Dr. Laylor finds a relationship between physical attractiveness and selfconfidence. He later determines that the primary cause of this relationship is the
positive feedback physically attractive people receive from others. Identify the
mediator or moderator.

4. Boys who are popular with the same-sex tend to also be popular with the
opposite-sex, whereas girls who are popular with the same-sex tend to be less
popular with the opposite-sex. Identify the mediator or moderator.

5. Identify and explain at least two potential mediators and two potential
moderators for the following relationship: School size and academic
achievement.

Table 1—why do some areas of psychology
use more mediators than others?

Case 1: Cat IV, Cat Mod:
 2 x 2 ANOVA

Case 2: Cont IV, Cat Mod:
 Could do correlations sep and compare BUT
 Better to do regression and compare
unstandardized Betas
 Or SEM software

Case 3: Cat IV, Cont Mod:
 Figure out how level of moderator affects IV-DV
relationship
 If linear, do hierarchical regression, showing that
XZ adds to the effects of X and Z on Y

Case 4: Both Cont:
 Use Case 2 if step function or
 Use Case 3 if linear, quadratic

How do you know which one is the
moderator vs. IV?

Causal steps (Baron & Kenny)





IV related to M
M = i + aX + e
IV related to DV
Y = i + c1X + e
M and IV related to DV Y = i + c2X + bM + e
C1 greater than C2 (look at size and sig)
Limitations:




Not good for multilevel, probit, logistic, survival
Need to test for whether C1 > C2 (Sobel test)
Low power esp when IV and DV aren’t related
Overestimates effect of IV on DV if error in M



Used to test whether C1 > C2
Good for sample sizes of 50+ with 1 M
Or 100-200 for >1 M


Use multiple measures of M and use SEM
Distribution of the product
 PRODCLIN
 Better Type I error rates, higher power

Computer-intensive methods
 Aka resampling
 Fewer assumptions





Residuals are independent in equations 2 and
3
No XM interaction in equation 3
Direction is correct (DV doesn’t cause M)
Measurement is perfect, esp. in M
No unmeasured variables that cause X, Y, or
M
 IV related to M
M = i + aX + e
 IV related to DV
Y = i + c1X + e
 M and IV related to DV Y = i + c2X + bM + e

Complete vs. partial mediation
 Use hierarchical regression
 Test C2 significance


Inconsistent mediation
Multilevel mediation
 Can increase T1 error if you ignore

Categorical DV mediation
 Use logistic or probit regression

Multiple mediators




Longitudinal mediation
Moderated mediation
Mediated moderation
Mediated baseline by treatment moderation
How do you know which variable is the
mediator?
 How can a moderator lead you to a mediator?
 How can a mediator lead you to a moderator?
 Can a variable be both a mediator and a
moderator at the same time?
 Are there variables that are always going to be
one or the other?
 Does a mediator have to be correlational?
 Remember that one study does not a
mediator/moderator make





Think of sections as headings in your thesis
Keep headings parallel
Always have at least 2 in each group
Think about the best logical order for things,
and keep the order consistent throughout the
paper


Quasi-experiments
2 book chapters plus other chapter