SEM Flashcards
Structural Equation Modeling
○ Regression with latent (unobservable) variable that we can’t directly measure
○ Allows for multiple predictors and outcomes/dependent variables
○ Measurement model: identifies # of latent constructs via confirmatory factor analysis
○ Structural model: identifies causal relationships via regression paths
SEM fits the hypothesized model to the observed data: fit indices evaluate how likely it is that a given model gave rise to the observed data
Types of Constructs
- Reflective: construct is cause of the measure, is reflected by indicators (observable measures)
- Formative: measures cause the construct (Ex: SES)
Reflective Construct
□ Indicators are called “effect indicators” because reflective construct influences indicators
-Must be correlated
□ Items should show strong internal consistency reliability (correlated, because all reflection of underlying construct)
□ Factor loading: represented by lambda (y) in diagrams
-FL = relative weight that the item is given in the estimation of the latent factor
-Need not include all facets of a construct, as long as its unidimensional; can be interchanged which allows for parallel-forms
□ If a person’s level on the construct changes, their score on each measure should change accordingly
Identified Model
- # datapoints = # parameters
- df = 0
- Remember, df = #datapoints - #parameters
Under-identified Model
- # datapoints < # parameters
- df<0
- Remember, df = #datapoints - #parameters
- Requires you to add assumptions & variances
Over-identified model
- # datapoints > # parameters
- df > 0
- Remember, df = #datapoints - #parameters
Latent Class Model
- SEM for categorical models
- Ex: diagnosis
Mixture Model
-SEM for combo of categorical constructs and continuous constructs
Benefits of SEM
○ Allows for correlated errors (unlike CTT)
○ Handles multiple dependent variables simultaneously (unlike multiple regression)
○ Uses all available info/data using Full Info Maximum Likelihood (FIML) if pts are missing data (instead of deleting entire pt)
○ Can be used to account for different forms of measurement error (method bias)