Multivariate Correlational Research (Part 2)

1
Multivariate Correlational
Research (Part 2)
Dr. Stefanie Drew
[email protected]
2
Review: The three causal rules
• Is there covariance?
• Is there temporal precedence?
• Are there third variables that could explain the
relationship?
3
Advanced Correlational Techniques
Multivariate Techniques
• Longitudinal Designs
▫ Establishing temporal precedence
• Multiple Regression Design
▫ Ruling out third variable
• “Pattern and Parsimony” Approach
▫ Results of variety of correlational studies all
Note* In all techniques,
support causal theory
variables are measured not
manipulated
4
Multivariate Design
• Involves more than two measured variables
• Widely used, especially when experiments are
not an option
Variable 1
Variable 2
Variable 3
5
6
Multiple Regression
• Multiple Regression: statistical
technique used to test for the
influence of third variables
• Conducting multivariate
correlational study
Image adapted from pn.bmj.com/content/7/4/259.extract on March 19, 2014
7
Barros et al., 2009:
Examining Recess Time and Behavior
• Two key variables
▫ Recess time: number of
minutes of recess
▫ Behavior problems:
standardized measurement
teacher rates
• Not binary correlation, because
measuring additional variable
with intention to look at
interrelationships
▫ Number of economically
disadvantaged children
▫ Public or private classrooms
▫ Class size
8
What is the overall point of multiple
regression analysis?
• Test whether association of
•
interest still true when statistically
controlling for suspected third
variables
▫ Question: Does the relationship
between recess and behavior
problems stay within different
levels of income?
Control for: holding a potential
third variable steady while
investigating the association
between two other variables
9
The Process
• Researchers examine third variables
▫ E.g. economic disadvantage to see if it correlates with other two
variables of interest
• Researcher must examine the data when controlling for
other variable
▫ E.g. when they take into account relationship between behavior
and income, does recess still have an effect on behavior
• Similar to identifying subgroups
▫ E.g. does bivariate relationship remain across all subgroups
Look within an income bracket, and see if behavior and recess
correlated
Look at next income bracket, see if behavior and recess correlated.
10
Barros et al., 2009
Examining Recess Time and Behavior
• Find overall relationship is negative
11
• Overall relationship is negative
Possible Outcome
• ALSO…
▫ Classrooms with poorest children ($)
have more
behavior problems
and shorter
recess
▫ Classrooms with wealthiest children
($$$) have fewer
behavior
problems and longer
recess
• If only look at one subset, _do_
still see negative relationship between
recess and behavior
• Conclusion: relationship between
recess and behavior still there when
controlling for income
12
• Overall relationship is negative
• ALSO…
Alternative Outcome ▫
Classrooms with poorest children
behavior
($) have more
problems and shorter
recess
▫ Classrooms with wealthiest
children ($$$) have
fewer
behavior problems and longer
recess
• If only look at one subset, do not
still see negative relationship
between recess and behavior
• Conclusion: relationship between
recess and behavior not there
when controlling for income
▫ Income was the third variable
responsible for relationship
13
Using Regression
• Testing whether a key relationship remains
when statistically controlling suspected third
variable
• Which of the two possible outcomes best
describes the relationship between recess and
behavior?
vs.
14
Terminology
• Criterion variables (a.k.a. dependent variables): the
variable interested in understanding/predicting
▫ Example: behavior problems
• Predictor variables (a.k.a. independent variables):
the remaining variables measured in analysis
▫ Example: amount of recess and levels of income
15
Regression tables
• Criterion variable typically goes in top
row or title
• Column for Beta Values
▫ (beta, β or standardized beta)
16
Beta
• A.k.a. β , standardized beta
• One beta for each predictor variable
• Can compare (within same table) different
predictor variables to each other based on their
β values
17
Beta vs r
• Similar to “r” in terms of relationship direction
▫ When other predictor variables are controlled for
+β: positive relationship between predictor and criterion variables
-β: negative relationship between predictor and criterion variables
β = 0: no relationship
• Similar to “r” in terms of relationship strength
▫ The larger β, the stronger the relationship
▫ The smaller β, the weaker the relationship
• Different from “r”
▫ No clear indication on effect size. Why?
18
b values
• A.k.a. unstandardized coefficient
• Similar to β in terms of relationship direction
▫ When other predictors are controlled for
+b: positive relationship between
-b: negative relationship between
• Cannot compare two bs within same table. Why?
19
A second look at Regression tables
20
What about significance?
21
Additional Predictors
• β indicates relationship present
between predictor and criterion
variables when all other
predictors held constant
• What is the purpose of adding
additional predictors?
▫ Helps control “third” variables
at same time
▫ Can look at other predictor
variables and get an idea of
other contributing factors
22
Summary
Measurement
Association
r bivariate correlations + r: positive association
- r: negative association
r = 0: no association
β multiple regression
+ β: positive association
(when other predictors controlled for)
- β: negative association
(when other predictors controlled for)
β = 0: no association
(when other predictors controlled for)
b multiple regression
+ b: positive association
(when other predictors controlled for)
- b: negative association
(when other predictors controlled for)
b= 0: no association
(when other predictors controlled for)
Can compare
within table?
23
A note about popular press…
• Studies may report a simple relationship but use
phrases
▫ “have been controlled for”
▫ “taking into account”
▫ “correcting for”
• Phrases indicate using multiple regression
24
Multiple Regression
• Basically… is the association still present when
you break it down into subgroups?
• Problems
▫ Controls third variable problem, but does not
account for temporal precedence
▫ Can only control third variables researchers
measure (e.g. not unknown third variables)
25
26
Pattern and Parsimony
• Converging evidence from
several studies
▫ Variety of methods used
• Using most parsimonious
explanation for an entire
pattern of data
• Example: smoking and lung
cancer
▫ Multiple studies looking at
the toxicity of cigarettes
Study
2
Study
1
Study
3
Single Theory
27
Another note about popular press…
• Journalists may only report results of recent
study (don’t address full pattern)
• Selectively representing research (without giving
full context)
28
Multivariate Design and Four Validities
• Construct
▫ How well is each variable measured?
• External
▫ Do results generate to other populations?
▫ Were samples taken randomly?
• Statistical
▫ What is the effect size?
▫ Is it statistically significant?
▫ Are there subgroups, outliers, curvilinear relationships?
• Internal
▫ Get closer to causal statements
Longitudinal designs establish temporal precedence
Multiple regression designs help rule out third variables
Exit Ticket
EXIT