structure mediation structural equation modeling

15
Structure Mediation Structural Equation Modeling

Upload: ariel-king

Post on 17-Dec-2015

217 views

Category:

Documents


1 download

TRANSCRIPT

Page 1: Structure Mediation Structural Equation Modeling

Structure

Mediation

Structural Equation Modeling

Page 2: Structure Mediation Structural Equation Modeling

Research Questions: (from Tabachnick & Fidell, Chapter 2)

• Degree of relationship amongst variables• Correlation• Linear Regression

• Prediction of group membership• Logistic Regression

• Structure• Mediation• Structural Equation Modeling (SEM)

• Significance of group differences• 2 groups: t-test• 3+ groups: ANOVA

Page 3: Structure Mediation Structural Equation Modeling

Research Questions: (from Tabachnick & Fidell, Chapter 2)

• Degree of relationship amongst variables• Correlation• Linear Regression

• Prediction of group membership• Logistic Regression

• Structure• Mediation• Structural Equation Modeling (SEM)

• Significance of group differences• 2 groups: t-test• 3+ groups: ANOVA

Page 4: Structure Mediation Structural Equation Modeling

Overview of “Structure”

• Defined: Testing interrelationships amongst variables

• Variables: Variables are continuous and/or categorical

(notice we are not talking about IVs and DVs)

• Relationship: Structure amongst variables

• Example: What is the relationship between provocation, anger, aggression, identifying with victim, perceiving outgroup as cohesive, etc

• Assumptions: If linear: Normality. Linearity. Multicollinearity If categorical:

Multicollinearity

Page 5: Structure Mediation Structural Equation Modeling

Relationship to correlation/regression/logistic (CRL)

• CRL involves:• 1 DV• 1+ IV

DV

IV 3

IV 2

IV 1

Page 6: Structure Mediation Structural Equation Modeling

Relationship to correlation/regression/logistic (CRL)

• CRL involves:• 1 DV• 1+ IV

• Structure• 2+ variables

IV3

IV 2

IV 1

IV 3IV 2IV 1 IV4

IV 3

IV 4IV 1 IV2

Just a few of the permutations:(any variable can go in any position)

Page 7: Structure Mediation Structural Equation Modeling

Relationship to correlation/regression/logistic (CRL)

• CRL involves:• 1 DV• 1+ IV

• Structure• 2+ variables

NOT CAUSATION(only correlation)

PSEUDO CAUSATION(“true” causation is experiments)

Page 8: Structure Mediation Structural Equation Modeling

How to test for structure:(1) Goal is to find best fitting model

(2) You find best fitting model by looking at converging evidence of various criteria

(3) Start with “confirmatory” analysis testing your hypothesis

(4) Then move to “exploratory” analysis in which you first disconfirm rival hypotheses, and then test for new hypotheses

(5) You have so many possible permutations of the variables that exploratory analysis is usually not comprehensive

Page 9: Structure Mediation Structural Equation Modeling

Mediation: Terminology• See my PsychWiki page --

http://www.psychwiki.com/wiki/Mediation

• Variables:• X is the predictor• Y is the outcome• M is the mediator

• Paths• C is the total effect• C’ is the direct effect • A-to-B is the indirect

Page 10: Structure Mediation Structural Equation Modeling

Mediation: Baron and Kenny• Most commonly used and most frequently cited test of

mediation, but also the most flawed.

• Four steps• X predicts Y (path c sig)• X predicts M (path a sig)• M predicts Y (path b sig)• X does NOT predict Y when controlling for M (path c’ NOT sig)

Page 11: Structure Mediation Structural Equation Modeling

Mediation: Sobel test• The Sobel test is superior to the Baron & Kenny

method in terms of all the limitations of the B&K method (e.g., power, Type I error, suppression effects, addressing the significance of the indirect effect).

• Math is complicated, but basically the Sobel test tests the significance of the relationship between c and c’

Page 12: Structure Mediation Structural Equation Modeling

Mediation: Example

• Baron and Kenny:• In the first step of the analysis, there was a significant relationship between

Provocation and Aggression ( = .20, p = .05). • In the second step of analysis, there was a significant relationship between

Provocation and Anger ( = .26, p = .01). • In the final step of the analysis, there was a significant relationship between

Anger and Aggression ( = .26, p = .03), while the relationship between Provocation and Aggression became non-significant ( = .10, p = .31).

• Sobel• There was a significant initial relationship between Provocation and

Aggression ( = .20, p = .05) that was non-significant after controlling for the mediator ( = .14, p = .31) which indicates Anger mediates the relationship between Provocation and Aggression.

Provocation Aggression

Anger

Page 13: Structure Mediation Structural Equation Modeling

SEM: Terminology• Exogenous variable: not caused by another• Endogenous variable: caused by another• Coefficients: strength of relationship• Path model: see below• Model fit: see next page

Identification

Entitativity

Retribution towards the Perpetrator

Retribution towards the

Group

Anger composite

.25** .45***

.23*

.20*

.31*** .56***

Page 14: Structure Mediation Structural Equation Modeling

SEM: Criteria• Theory:

(1) Evaluating multiple fit indices simultaneously is recommended…(2) because different indices assess different aspects of goodness-of-fit…(3) and there is not always agreement on what constitutes good fit…(4) so satisfactory models should show consistently good-fitting results on many different indices.

• Four recommended criteria:(1) Comparison: Chi-square: p < .05 (2) Parsimony: Ratio of x2/df < 3 (3) Absolute fit: SRMR < .08 (4) Relative fit: CFI > .95

Page 15: Structure Mediation Structural Equation Modeling

SEM: Example• Overall model fit was excellent:

X2=1.03, p =.794, x2/df =.34, SRMR =.03, CFI =1.00. • Alternative models achieved less satisfactory fit:

(1) Other models didn’t reach criteria from hypothesized model

(2) Nested models (subset of other) was sig chi-square test

(3) Un-nested models had lower AIC value

Identification

Entitativity

Retribution towards the Perpetrator

Retribution towards the

Group

Anger composite

.25** .45***

.23*

.20*

.31*** .56***