M4 - Path Analysis Flashcards
Part A - Question: SEM is… (pick all that apply)
Used as a statistics-driven approach to determine variable relations when you are unsure which variables are related to each other.
Useful for statistically comparing competing models.
A way to test the inter-relationships among a range of variables.
Completely different from multiple regression in the output it will provide
Useful for statistically comparing competing models.
and
A way to test the inter-relationships among a range of variables.
Part B - Question 1: Tick all statements that apply for the model shown here.
Model shows 3 rectangular boxes 1 - Dietary restraint 2 - Body dissatisfaction 3 - Binge Eating There is a bidirectional arrow connecting boxes 2 and 3
Binge eating is unrelated to dietary restraint.
Dietary restraint, body dissatisfaction, and binge eating are latent variables.
Body dissatisfaction is a predictor of binge eating
Binge eating is unrelated to dietary restraint.
How are latent variables, measured variables and residuals depicted in path model diagrams?
latent - circle or ellipse
measured - square or rectangular
residuals - circles coming off each predicted (dependent/endogenous variable
How are causal and correlational relationships depicted in path model diagrams?
causal has arrow going from one variable to another - unidirectional
correlation has a bidirectional arrow between two variables
What are purely endogenous, partially endogenous and exogenous variables?
Exogenous is unaffected by other variables. It has arrows going from them (exo = out of) to another variable
Partially endogenous in a variable that is partially affected by other variables but also has an exogenous component. It will have an arrow going from it on one side and an arrow going to it on the other side
Purely endogenous is a variable that is wholly affected by other variables. It will only have arrows going into it, no arrows going out of it
What is structural equation modelling (SEM)?
SEM is a way to analyse associations between variables
It provides a way to conceptualise relationships between variables differently and then test them for goodness of fit to the data
It is a general umbrella term for
- Path Analysis - structured approach - how are the variables related?
- Factor Analysis - measurement approach - how the variables are measured
- Path and Factor Analysis combined
- and others beyond 4th year level….
FA helps with reliability between constructs modelled in PA
How does path analysis fit within SEM?
Path Analysis refers to the structural component of SEM.
Path Analysis deals with measured variables and their error terms and arrows of modelled relationships
What do you need to do as a researcher in terms of prep, data and model statistics to assess SEM?
Minimum of one model
- ->Consultation of the literature for possibilities around how the variables relate to each other
- ->Knowledge of how to construct model in stats software like SPSS AMOS
Some data
- ->good to have an idea of the model before data collection
- ->ensure there is sufficient data collected
- ->consider model complexity
Statistics
- ->Sig tests, R2
- ->Goodness of fit, model comparison
How is path analysis different from MR?
Path Analysis
- structural model that requires specification of how variables are related
- requires sophisticated thinking of how things occur in reality
- variables can be both IVs or DV
- Multiple DVs can be tested at a time
MR allows
- only one DV can be tested in a model
- less critical thinking, the calculations work out relationships
- variable cannot be both IV and DV
What are the advantages of path analysis?
- allows variables to be both an IV and a DV in the same model
- will theoretically provide a more sophisticated, nuanced and better fitting model of how the variables relate in real life
Explain how to evaluate your Path Analysis model (using the chi-square statistic)
Global Fit - Primary Statistic
Chi2 (x2)- a discrepancy measure between actual relationships between variables in the data and implied relationships between variable of a given model
- takes all the variables in the model. Runs a correlational matrix and compares the actual correlation in our data to the correlations implied by the model
- high Chi2 means there is a large difference between our data and the implied associations in the model
- chi2 of 0 means perfect fit
- negative chi2 not possible as its a squared statistic
- significant is P .05 because this means the model is not sig. diff from the data
Can also look at goodness of fit from a local perspective (beta weights, R2 variance explained)
Where a x2 indicates a poor fit, look to local to see where that might be coming from
What is the general approach to SEM as a researcher?
- Read the literature
- Devise a model
- Collect the data
- Test the model against your data.
NOTE: SEM will not work without collection of data!!!
What are the basic building blocks for SEM?
Circles - latent variables and error terms
Squares /rectangles - measured variable
Arrows - single headed (direction of prediction), double headed (two variables are related) , zero arrow (no relationship between variables)
error term refers to all the unexplained variance in the DV, ie the variance that is not explained by the predictor variable
Can SEM provide evidence of a causal relationship?
Like with MR to confirm a causal relationship the study must be longitudinal and not cross-sectional.
In a cross-sectional design we say that our results are consistent with a model where IV influence DV, but if cross-sectional then can’t establish true causality
However, SEM can at least provide some evidence that supports the notion of a causal relationship.
Criteria for causality
- the two are related (correlational analysis)
- the IV preceded the DV in time (established through longitudinal design)
- changes in the IV lead to changes in the DV - (established through experimental design)
How does failing to draw paths between related variables affect the path analysis model results?
This can adversely affect our ability to accurately predict other pathways across the model, not just where the relationship was failed to be identified
