an introduction to path analysis and structural modeling · pdf filetoday’s class •...
TRANSCRIPT
An Introduction to Path Analysis andStructural Equation ModelingMultivariate Methods in Education
ERSH 8350Lecture #13 (and final)November 30, 2011
ERSH 8350: Lecture 13
Today’s Class
• An introduction to path analysis
• An introduction to structural equation modeling
• Information about the take‐home final exam Drafts for revisions: Due 11:59pm Wednesday, December 7th
Final Drafts: Due 11:59pm, Wednesday December 14th
ERSH 8350: Lecture 13 2
AN INTRODUCTION TO PATH ANALYSIS
ERSH 8350: Lecture 13 3
Path Analysis
• Path analysis is a statistical technique used to estimate a set of simultaneous regression equations All variables are observed – assumed measured without error
In SEM, we can use a path analysis for latent variables
• Path models are useful for Mediation analyses: the effect of one variable on another is mediated by one or more variables
Simultaneous equations: multiple predictions of multiple variables in a model
ERSH 8350: Lecture 13 4
Data Example
• Today’s example data come from the UCLA Academic Technology Services website http://www.ats.ucla.edu/stat/mplus/seminars/introMplus_part2/path.htm
• The data file contains four variables collected from 200 respondents: High School GPA (hs) College GPA (col) GRE Score (gre)
Note: GRE scores are not measured perfectly – so this is likely to over‐power this analysis
Graduate School GPA (grad)
ERSH 8350: Lecture 13 5
Initial Setup: Determining What is Possible
• Much like the other analyses in this class, the number of “paths” in a path analysis is constrained by the number of covariances between the observed variables Here, a path represents a
regression coefficient
• Mplus can give you the descriptive statistics for an analysis with a fairly simple set of commands:
ERSH 8350: Lecture 13 6
Mplus Output: Descriptive Statistics
ERSH 8350: Lecture 13 7
Note: Estimates come from ML –with a MVN distribution
Covariances
Means
Variances
Mplus Output: Standardized Coefficients
ERSH 8350: Lecture 13 8
Correlations
Initial Analysis: A Just Identified Model
• For our first analysis, we will examine the following set of equations:
• GRE score depends on HS (?) and COL GPA• GRAD GPA depends on HS GPA (?), COL GPA, and GRE score
• Note: 6 coefficients We have 6 covariances – so this models is called “just identified” Just identified models give no meaningful fit statistics
ERSH 8350: Lecture 13 9
A Path Diagram of Path Analysis
ERSH 8350: Lecture 13 10
Mplus Syntax and Output for Path Model #1
ERSH 8350: Lecture 13 11
Path Analysis Unstandardized Parameter Output
ERSH 8350: Lecture 13 12
Path Analysis Standardized Parameter Output
ERSH 8350: Lecture 13 13
Types of Effects in Path Analysis
• Depending on how a model is set up, there are three types of effects in a path analysis: Direct effects – the influence of one variable directly on another Example: HS GPA on GRE Score
Indirect effects – the influence of one variable on another, as mediated by one or more variables Example: HS GPA on GRAD GPA, through GRE Score
Total effects – The aggregate effect of direct and indirect effects
ERSH 8350: Lecture 13 14
Mplus Syntax for Indirect and Total Effects
ERSH 8350: Lecture 13 15
Mplus Unstandardized Indirect Effects Output
ERSH 8350: Lecture 13 16
Mplus Standardized Indirect Effects Output
ERSH 8350: Lecture 13 17
Over‐Identified Models
• Over identified path models are models with fewer coefficients than covariances Tests of model fit are available Even if model fit is poor, information may be meaningful
• New model:
ERSH 8350: Lecture 13 18
Mplus Syntax
ERSH 8350: Lecture 13 19
Mplus Model Fit Output
ERSH 8350: Lecture 13 20
Mplus Unstandardized Output
ERSH 8350: Lecture 13 21
Mplus Standardized Output
ERSH 8350: Lecture 13 22
Mplus Unstandardized Indirect Effects Output
ERSH 8350: Lecture 13 23
Mplus Standardized Indirect Effects Output
ERSH 8350: Lecture 13 24
Path Analysis Wrap‐Up
• Path analysis is a powerful technique for testing multivariate regression models with direct and indirect effects
• The example shown today was meant to provide a basic overview of the technique Many details omitted
• For more information, come to the SEM course next semester
ERSH 8350: Lecture 13 25
AN INTRODUCTION TO STRUCTURAL EQUATION MODELING
ERSH 8350: Lecture 13 26
Structural Equation Modeling• Similar to confirmatory factor analysis and path analysis,
structural equation model seeks to describe relationships between latent variables Generally speaking, SEM is for latent variables…path analysis is for
observed variables
• SEM consists of two key elements: Measurement model – typically a confirmatory factor model
The link between the data and the latent variables Structural model – a path model or factor model for the
latent variables Generally uses the covariance matrix between latent factors
• As such, the basic SEM analysis consists of portions you already know
ERSH 8350: Lecture 13 27
SEM: Example Data
• To demonstrate the structural equation model, we will use data presented in the PCA/EFA lecture Gambling research instrument: 41 items measuring 10 factors
• We will use these data to demonstrate SEM through: A higher order factor model
A path model for the higher order factors
• Before we do either of these, we must first construct a measurement model A CFA model for the data
ERSH 8350: Lecture 13 28
Pathological Gambling: DSM Definition
• To be diagnosed as a pathological gambler, an individual must meet 5 of 10 defined criteria:
Pathological Gambling 29
1. Is preoccupied with gambling 2. Needs to gamble with increasing
amounts of money in order to achieve the desired excitement
3. Has repeated unsuccessful efforts to control, cut back, or stop gambling
4. Is restless or irritable when attempting to cut down or stop gambling
5. Gambles as a way of escaping from problems or relieving a dysphoricmood
6. After losing money gambling, often returns another day to get even
7. Lies to family members, therapist, or others to conceal the extent of involvement with gambling
8. Has committed illegal acts such as forgery, fraud, theft, or embezzlement to finance gambling
9. Has jeopardized or lost a significant relationship, job, educational, or career opportunity because of gambling
10. Relies on others to provide money to relieve a desperate financial situation caused by gambling
Research on Pathological Gambling
• The Gambling Research Instrument (Feasel, Henson, & Jones, 2002) was created with 41 Likert‐type items Items were developed to measure each criterion
• Example items (ratings: Strongly Disagree to Strongly Agree): I worry that I am spending too much money on gambling (C3) There are few things I would rather do than gamble (C1)
• The instrument was used on a sample of experienced gamblers from a riverboat casino in a Flat Midwestern State Casino patrons were solicited after playing roulette
Pathological Gambling 30
Measurement Model Mplus Syntax
ERSH 8350: Lecture 13 31
Measurement Model Output
ERSH 8350: Lecture 13 32
Structural Equation Model #1
• The first model we will estimate attempts to determine if the 10 gambling criteria are all related to a hierarchical factor of gambling:
ERSH 8350: Lecture 13 33
Gambling
C1 C2 C3 C4 C5 C6 C7 C8 C9 C10
Mplus Syntax
ERSH 8350: Lecture 13 34
Mplus Output
ERSH 8350: Lecture 13 35
Mplus Output
• Unstandardized:
• Standardized:
ERSH 8350: Lecture 13 36
Structural Equation Model #2
• The DSM suggests the gambling criteria are influenced by three factors: dependency, disruption, and loss of control:
ERSH 8350: Lecture 13 37
C1 C2 C3C4 C5 C6 C7C8 C9 C10
Dependency Loss of Control
Disruption
Mplus Syntax
ERSH 8350: Lecture 13 38
Mplus Output
ERSH 8350: Lecture 13 39
Mplus Output
• Unstandardized:
• Standardized:
ERSH 8350: Lecture 13 40
Structural Equation Model #3
• Instead of correlated factors, could we determine if dependency caused loss of control, which caused disruption?
ERSH 8350: Lecture 13 41
C1 C2 C3C4 C5 C6 C7C8 C9 C10
Dependency Loss of Control
Disruption
Mplus Syntax
ERSH 8350: Lecture 13 42
Mplus Output
ERSH 8350: Lecture 13 43
Mplus Output
• Unstandardized:
ERSH 8350: Lecture 13 44
Mplus Output
• Standardized:
ERSH 8350: Lecture 13 45
SEM Wrap Up
• Structural equation modeling is a procedure where relationships between latent variables is evaluated and tested statistically Based on covariances of latent variables
• The modeling technique is very powerful: it extends well beyond what was shown today
ERSH 8350: Lecture 13 46
TAKE HOME FINAL EXAM INFORMATION
ERSH 8350: Lecture 13 47
Take Home Final Exam Information
• The take home final exam is a two‐week test where you are asked to perform an intricate data analysis using techniques learned in this class
• The final will not be easy – but can be done I would suggest you do not procrastinate
• I am offering you the chance to receive feedback if you submit a draft of your final by next Wednesday Your draft must be an attempted answer, not simply syntax and output
ERSH 8350: Lecture 13 48
Course Evaluations
• Please complete the course evaluation at:
https://portal.coe.uga.edu/apps/authorize/
• Thank you for a wonderful semester I hope you found this course informative!
ERSH 8350: Lecture 13 49