Study Notes on Multiple Regression
Overview of the Unit
Covered Module One
Looked into correlations
Currently at Multiple Regression
Transition to Multiple Regression
Multiple regression is significantly foundational in research.
Personal preference for statistical methods:
- If restricted to one method: Multiple Regression
- In reality, choice would be Structural Equation Modeling (SEM)Comparison between SEM and Multiple Regression:
- SEM akin to riding a unicycle: more complex
- Multiple Regression feels more accessible, like riding a bicycle
Lecture Objectives
Conceptual overview of multiple regression
Worked example related to car purchase using multiple regression
Step-by-step breakdown of research steps:
- Power calculations
- Outlier assessment
- Initial assumption testing (moved to a separate lecture for clarity)Interpretation of output
Summary of multiple regression in three slides
Brief insight into the mechanics of regression
Definition of Multiple Regression
Involves two or more independent variables (IVs) to predict a single dependent variable (DV).
Comparison to Simple Regression:
- Simple Regression: only one IV
- Multiple Regression: multiple IVs (continuous or categorical)
Independent & Dependent Variables
Predictors (IVs):
- Primarily continuous, can also be dichotomous (binary)
- Use of dummy coding to include categorical variablesDependent Variable (DV):
- Must be continuous or a scaled variable
Methods and Related Concepts
Current focus: Multiple Regression
Future topics: Assumptions of Regression, Hierarchical Regression, Moderation, and Mediation
Predicting Child IQ Example
DV: Child IQ at age 10 (continuous variable)
IVs:
- Mother’s smoking
- Months of breastfeedingTransformation from correlation to multiple regression:
- Involves multiple r instead of a single r
- Squared multiple r gives percentage of variance explained in DV by IVs combined
Understanding the Output
Introduction of Beta Weights:
- Each IV has a beta weight that indicates the relationship with the DV
- Multiple r captures combined effect of beta weights
Nature of Multiple r
Properties of Multiple r:
- Can only be positive (unlike r which can be both)
- Reflects the cumulative effect of positive and negative beta weightsLimitations of Multiple r:
- Potential bias in favor of positive association
- Use of squared multiple r to determine explained variance
Application Example: Car Pricing
Objective: Determine how IVs (kilometers, year model, sale type) predict car price
- Required analysis steps include:
- Research question formulation
- Statistical approach selection (Multiple Regression)
- Hypothesis development
- Power computation (post hoc analysis with pre-set alpha levels)Sample Size Considerations:
- Use rules of thumb for power calculations based on number of predictors (k)
- Sample size of around 74 for overall multiple r, 107 for individual predictors
Data Cleaning and Outlier Examination
Essential to check for outliers in multiple regression
- Multivariate outliers important due to multiple IVsMahalanobis distance for multivariate outlier detection:
- Calculation involves distance from group centroid
- Assessment through Chi-squared distribution with a critical alpha of 0.001Decision-making with outliers:
- Remove entire cases with problematic outliers
- Conduct analyses with and without outlier cases to compare results
Regression Output Overview
Snapshot of regression output tables
Key elements include:
- Model summary with multiple r
- ANOVA table indicating significance of the modelInterpretation of beta weights alongside significance testing
- Unstandardized and standardized beta weights bring clarity to predictor contributions
Beta Weights Interpretation
Unstandardized Beta Example:
- Unit change in year results in increase in price
- Flipping perspective for correct interpretation is necessaryStandardized Beta Weights:
- Allow comparison of predictor strength across different units
- Interpretation in standard deviation units for consistency
Correlation Types in Regression
Types of correlations obtained in multiple regression:
- Zero-order correlation
- Partial correlation
- Semi-partial correlation (most relevant)Importance of distinguishing between correlation types for thorough analysis
Regression Equation Formulation
Understanding the regression equation: y hat representation
Calculation of residuals:
- Difference between observed score and predicted score
- Residual as a measure of error in model predictions
Practical Application in Car Purchase
Data analysis application: predict value of potential car purchases using regression equation
- Application involves evaluating multiple characteristics of carsDecision-making based on calculated predictors reflecting market conditions
Generalizing Beyond Sample Data
Importance of generalizing models to broader populations
- Example scenario of negotiating with a dealership using predictor weights
- Comparison to actual market value based on calculated predictions
Inner Workings of Multiple Regression
Ordinary Least Squares (OLS) criterion for optimization
- Minimizes sum of squared residuals through calculated predictionsFinal insights and reflections on regression's capabilities
Conclusion
Wrap up of lecture contents and reinforcement of understanding multiple regression.
Call to apply concepts in practical situations beyond the classroom setting.