Introduction to Multivariate Relationships
Absolutely — here is a clean, professionally formatted, study‑ready version of your multivariate‑relationships outline. I kept your SOC 110 style in mind: clear hierarchy, minimal clutter, strong structure, and easy scanning for exams.
📘 Introduction to Multivariate Relationships — Clean Outline Notes
(With Stata commands, diagrams, and practice problems)
1. Context of Multivariate Analysis
Multivariate analysis extends bivariate analysis by adding additional variables to:
test whether an association persists, weakens, or disappears
evaluate causal claims
identify confounding, mediation, and interaction
Example:
Gender → Party ID
But is this due to ideology, income, education, or religion?
2. Association and Causality
2.1 What Is Causality?
A causal relationship must satisfy all three:
1. Association
X and Y must vary together.
Categorical:
chi2Quantitative: regression slope, t-test
Stata
tabulate scout delinquent, chi2
regress math_score height
2. Time Order
Cause must precede effect.
Clear for race, age, gender
Ambiguous for scouting ↔ delinquency
Experiments fix time order; observational studies often cannot
3. Elimination of Alternative Explanations
Must rule out third variables (Z).
This is the hardest criterion.
Association ≠ Causation
Anecdotal Evidence
Anecdotes do not disprove causality unless they violate one of the three criteria.
3. Statistical Control
3.1 What Is Control?
A variable is “controlled” when its influence is removed by:
holding it constant
analyzing associations within levels of the control variable
3.2 Laboratory vs. Statistical Control
Lab control: physically hold variables constant
Statistical control: group observations by similar values of Z
3.3 Example 10.1 — Height & Math Achievement
Bivariate correlation: 0.81
Lurking variable: Age
Within grade levels (2, 5, 8), correlation ≈ 0
Stata
pcorr math_score height age
regress math_score height age
3.4 Example 10.2 — Scouting, Delinquency, Church Attendance
Bivariate: Scouts = 9% delinquent; Non-scouts = 15%
After controlling for church attendance → no difference
Church attendance is the common cause
Stata
by church_attendance: tabulate scout delinquent
3.5 Lurking Variables
Unmeasured variables that distort associations.
Example: “Hard work” → high GPA & high income.
4. Types of Multivariate Relationships
4.1 Spurious Association
X₁–Y disappears when controlling X₂.
X₂ causes both X₁ and Y.
Diagram
X2 → X1
X2 → Y
Stata
regress Y X1 X2
4.2 Chain (Mediation) Relationship
X₁ → X₂ → Y
X₁ is an indirect cause.
Diagram
X1 → X2 → Y
Stata (SEM mediation)
sem (X2 <- X1) (Y <- X2)
4.3 Multiple Causes
Y has several causes (independent or correlated).
Diagram
X1 → Y ← X2
Stata
regress delinquency gender race
4.4 Suppressor Variables
Association appears only after controlling Z.
Stata
regress income education age
4.5 Interaction (Effect Modification)
Effect of X₁ on Y changes across levels of X₂.
Diagram
X1 × X2 → Y
Stata
regress income education gender c.education#c.gender
4.6 Simpson’s Paradox
Direction of association reverses when controlling for Z.
5. Inferential Issues in Statistical Control
5.1 Sample Size Reduction
Controlling Z → smaller partial tables →
larger SE
wider CI
smaller chi-square values
5.2 Categorization Pitfalls
Too few categories → poor control
Too many → tiny sample sizes
5.3 Comparing Measures Across Groups
Standard error for difference: [ SE = \sqrt{se_1^2 + se_2^2} ]
Confidence interval: [ (\hat{\theta}_2 - \hat{\theta}_1) \pm z \cdot SE ]
5.4 Example 10.9 — Gamma for Men vs. Women
Men: 0.674
Women: 0.689
CI for difference: (–0.153, 0.183) → not significant
Stata
display (0.689 - 0.674) - 1.96*sqrt((0.0614)^2 + (0.0599)^2)
display (0.689 - 0.674) + 1.96*sqrt((0.0614)^2 + (0.0599)^2)
6. Practice Problems (SOC 110 Style)
1. Firefighters & Damage Costs
Identify the lurking variable.
→ Size of fire
2. Scouting & Delinquency
Create partial tables by church attendance.
Stata
by church_attendance: tabulate scout delinquent, row
3. Mediation
Draw a path diagram for:
Education → Income → Health
4. Interaction
Plot income vs. education by gender.
Stata
twoway (scatter income education if gender==1) ///
(scatter income education if gender==0)
5. Simpson’s Paradox
Analyze death penalty data by victim race.