Introduction to Multivariate Relationships
Overview of Multivariate Relationships — Clean Study Format
1. Transition from Bivariate to Multivariate Analysis
Chapters 7–9: Focused on two‑variable (bivariate) associations.
Real social science research requires multivariate methods to:
Add additional variables.
Test whether an observed association is real.
Determine whether a third variable explains or modifies the relationship.
Purpose of Multivariate Analysis
Researchers add variables to determine whether:
The association is genuine.
The association is spurious (explained by a third variable).
The association changes under certain conditions (interaction).
Example: Gender & Political Party
Bivariate finding: Men are more likely to be Republican than women.
Multivariate step: Control for political ideology.
If gender differences disappear within liberals or conservatives → ideology explains the association.
Other possible controls: income, education, religion.
10.1 Association and Causality
Scientific Goal
Science aims to identify causal relationships.
Social causes (e.g., poverty → delinquency) are harder to isolate than biological ones.
Causal Direction
Causality is asymmetric:
X→Y
X = explanatory variable
Y = response variable
Three Criteria for Causality
Association
X and Y must vary together.
Tested with chi‑square, t‑tests, regression slopes.
Time Order
X must occur before Y.
Some variables (race, age, gender) are inherently prior.
Elimination of Alternative Explanations
Must rule out third variables (Z).
Experiments vs. Observational Studies
Experiments: Time order is controlled by the researcher.
Observational studies: Variables may move together due to shared trends (e.g., divorce and crime both rising over time).
Association ≠ Causation
A third variable (Z) may cause both X and Y → spurious association.
Anecdotes Don’t Disprove Causality
Causal claims are probabilistic, not absolute.
10.2 Controlling for Other Variables
Definition
A variable is controlled when its influence is removed from the X–Y association.
Types of Control
Experimental control: Hold conditions constant in a lab.
Statistical control: Group cases by similar values of control variables.
Example 10.1: Height & Math Achievement
Bivariate correlation: 0.81.
Lurking variable: Age.
After controlling for grade (age), correlation drops to near zero → spurious.
Partial Tables
For categorical variables, create separate tables for each level of the control variable.
Example 10.2: Scouting & Delinquency
Bivariate: Scouts = 9% delinquent; Non‑scouts = 15%.
Control: Church attendance (High, Medium, Low).
Within each level, delinquency rates are identical → original association was spurious.
Sampling Variation
Associations rarely drop to exactly zero due to sampling error.
Lurking Variables
Unmeasured variables that distort associations.
10.3 Types of Multivariate Relationships
1. Spurious Associations
X₁–Y association disappears when controlling for X₂.
Examples: Height–math; scouting–delinquency.
Example 10.4: Vacations & Heart Attack Risk
After controlling for SES, the association remains, suggesting it may not be spurious.
2. Chain Relationships
Structure:
X1→X2→Y
X₁ is an indirect cause; X₂ is the mediator.
Example 10.5: Education → Income → Longevity
3. Multiple Causes
Y has more than one cause (X₁ and X₂).
Causes may be independent or correlated.
Controlling for X₂ reduces but does not eliminate X₁’s effect.
4. Suppressor Variables
No association appears until a third variable is controlled.
Example 10.6: Education & Income
Age masks the positive relationship.
5. Statistical Interaction
The effect of one predictor depends on the level of another.
Example 10.7: Education × Gender on Income
Men: +$4000 per year of education
Women: +$2000 per year
Example 10.8: SES × Age on Health
SES gap widens with age.
Simpson’s Paradox
Partial tables reverse the direction of the bivariate association.
Confounding
Two explanatory variables both affect Y and are associated with each other.
10.4 Inferential Issues in Statistical Control
Small Sample Sizes in Partials
Controlling for X₂ reduces sample size →
Larger standard errors
Wider confidence intervals
Smaller chi‑square values
Categorization Issues
Quantitative controls shouldn’t be categorized too crudely.
3–4 levels is typical; too many levels → tiny sample sizes.
Comparing & Pooling Measures
Compare parameters across levels of the control variable.
Standard Error for Differences
se=se12+se22
95% CI for Difference
(Estimate2−Estimate1)±1.96×se
Example 10.9: Happiness & Marital Happiness
Male gamma: 0.674 (se = 0.0614)
Female gamma: 0.689 (se = 0.0599)
Difference: 0.015
CI: (–0.153, 0.183) → includes 0 → no significant difference.