Wk 11 - SEM Flashcards
What are latent variables? (x1)
Unobservable psych constructs
How does SEM relate to CFA? (x3)
An extension of:
* Theory-driven approach to hypothesis testing * Based on the same mathematical principles
In characterising relationships between observed and latent varaibles, how is SEM like (x1), and different (x1) to CFA?
Can model correlational associations
Can also be used to model causal relationships (‘causal’ may be a little strong…)
What assumptions are made by SEM (that it shares with CFA)? (x2)
Multivariate normality
Linear relationships among variables
What is the typical focus of SEM? (x1)
Testing hypotheses about relationships between latent variables
What methods are subsumed by SEM? (x3)
Which make it…(x1)
Despite… (x1)
CFA,
ANOVA,
Regression
Very powerful and flexible analytic technique
(but some restrictions on use)
What are ‘manifest’ or ‘indicator’ variables in SEM? (x1)
Observed/measured variables
How are manifest/indicator variables’ variance calculated in SEM? (x2)
Unique for each variable,
Estimated from data
What constitutes the variance in latent variables in SEM? (x1)
Variance shared by subsets of manifestt/indicator variables
In what ways might latent variables relate in SEM? (x2)
Correlations (bidirectinal arrows) Causal paths (unidirectional)
What are exogenous latent variables in SEM? (x2)
Those not caused by other FACTORS in the model
Summarise variance shared by subsets of OBSERVABLE variables
What are endogenous latent variables in SEM? (x2)
Those predicted by other FACTORS (latent variables) in the model
Summarise hypothesised causal relations between constructs summarised by the exogenous VARIABLES
What is ‘disturbance’ in SEM? (x1)
Error terms associated with endogenous latent variables
What do direct causal paths describe in SEM? (x2)
Effects of a predictor variable,
Controlling for effects of all other variables
What are the 2 stages of conducting an SEM?
Measurement model:
*CFA to specify factor structure of subsets of variables
Structural model:
*Examine relationships among latent variables ID’d by measurement model
When do you specify order of variables’ association in SEM? (X2)
For masurement model,
And then for structural
What model fit evaluations are required in SEM? (x3, x2)
Fit indices (same as CFA):
* Chi-square, plus * Relative indices
Comparison with other plausible models
*Can matter less for measurement models
Under what conditions might you refine the model in SEM? (x2)
Free up additional parameters if initial fit is poor
Fix parameters to zero if they don’t contribute to fit (i.e., are ≈ 0)
Considering fit as well as parsimony, parameters should only be free/estimated if they… (x2)
Clearly improve model fit
*Contribute to more accurate depiction of data
How are estimated parameters evaluated in SEM? (x2)
But they should also be… (x1)
Chi-square, or
Akaike Information Criteria
Theoretically justified, not just ad hoc data fitting
Considering fit as well as parsimony in SEM, we should ask ourselves? (x1)
Because… (x1)
Is the model working due to complexity, or a good fit?
SEM has buckets of free parameters, so lots of additional flexibility…
Why is the traditional CFA chi-square test not necessarily useful when comparing alternative SEM models? (x3)
Because models in CFA are nested
SEM might compare qualitatively different structures
*ie differences in order of variable associations
What issue arises when comparing non-nested models in SEM? (x2)
A comparison model with (eg) more free-parameters
Will generate a higher chi-square
Meaning that more parsimonious models are disadvantaged in analysis
What is the solution to issue of biased chi-squares generated by comparing non-nested models in SEM? (x1, plus explain x2)
AIC (Akaike Information Criterion)
= chi-square + 2k
*where k = number of free parameters
What does the AIC do in practice? (x2)
Introduces a penalty for model complexity
So that simpler models can compete
How should the AIC(Akaike Information Criteria) be interpreted? (x2)
It summarises the fit/flexibility trade-off
So choose model with lowest AIC
How should you plan sample size for SEM? (x3)
Same as CFA:
* Aim for 10+/parameter * 5+ may suffice (reduced power)
What are the 3 steps for conducting SEM? (x2, x2, x3)
Specify measurement and structural models
*ID fixed and free parameters
Estimate parameters of measurement model
*Max observed/predicted fit for variance/covariance matrices
Examine model fit/parameter estimates
*Refine till well fitting/parsimonious
*Test alternatives and compare fits
What issues might constrain the use of SEM?
Sample size too small - need big for reliable estimates
Insufficient number of indicator variables (measurement model)
Poor fit of measurement model to data
How many indicator variables are required for measurement models in SEM? (x2)
Single-factor: 3
2+ factors: 2+ for each latent variable
Why is a poor-fitting measurement model an issue for SEM? x1
Cannot successfully identify latent variables underlying the data set
What method could you use instead of SEM if you don’t quite meet all the requirements? (x1, plus describe x4)
Path analysis:
* Creates composite variables by averaging observed measures * These play role of latent variables * So determine order of association, and * Evaluate fit/alternative models (as per CFA/SEM
While asking similar questions to the structural models in SEM, the earlier technique of Path Analysis was limited by… (x1)
Because of… (x1)
Inability to explicitly model error/unique variance
Because of lack of latent variables to account for shared variance