Escolar Documentos
Profissional Documentos
Cultura Documentos
Introduction to Path Analysis and Structural Equation Modelling with AMOS Week 4
Daniel Stahl Biostatistics and Computing
Estimation methods: ML Checking assumptions of ML with AMOS (Kurtosis, Skewness, Mardia's statistic, Mahalanobis distance) Bootstrapping Goodness of fit measures - Goodness of fit measures that penalizes for lack of parsimony (overfitting) Model selection using AIC criteria Modification of the model (modification indices suggest which e.g. paths are missing Confirmatory factor analysis
Goodness of fit measures based on predicted and observed covariance matrix but penalizes for lack of parsimony (overfitting)
Goodness of fit measure Root mean square residual (RMR) The smaller the better 0.05 Measure based on difference based on covariance residuals Discrepancy from chi2 distribution per df Effected by range of scales , standardise RMR would be better (not in AMOS) Least affected by sample size, penalizes for lack of parsimony, compare different models, robust
Today
Confirmatory factor analysis Simple SEM Multigroup comparisons/factorial invariance
Root mean square error of approximation (RMSE, RMSEA) Parsimony ratio (PRATIO) Parsimony Index Parsimonious NFI (PNFI) Parsimonious CFI (PCFI)
Df model/df independence >0.9 PRATIO * BBI (pratio=df model/df default model) PRATIO * NNFI PRATIO * CFI Not in AMOS (?), Smaller is better Smaller is better
Confirmatory Factor Analysis Confirmatory factor analysis is used: to confirm a hypothesized factor structure. as a validity procedure in measurement research. CFA confirms a specific relationship between the items and the factors
Certain items are hypothesized to go to given factors.
I1 Factor I
I2
I2
E
E
I3 Factor II
I3 Factor II
I4
I4
pain
1
e1
W ell-being
depress
depress function 1 .337** -.455** .337** 1 -.421** -.455** -.421** 1 **. Correlation is significant at the 0.01 level (2-tailed).
pain
e2
e2
1
e3
1
Predictor variable
function
e3
Dependent variables
Well-being
Explained variance
e1
x1 = 1 * f + 1 x 2 = 2 * f + 2 x 3 = 3 * f + 3
W ell-being
depress
e2
function
e3
with Var (observed variable xi ) = 1 (= Total variance of observed variable xi ) Var ( i ) = i (= unexplained variance of xi , measurement error)
Because the variables are standardis ed (we use the correlatio n matrix) the coefficient i is the correlatio n coefficien t ri between latent tra it f and x i . explained variance : ri2 Explained variance of x i by latent factor f = 2 i Unexplaine d variance of x i = i Var(x i ) = 2 + i = 1 i
1
1
e1
Var(x i ) = i2 + i Cov(xi , x j ) = i * j
0.34 0.45 1 R = 0.34 1 0.42 0.45 0.42 1
2 1 = 1 + 1
W ell-being
2 depress
1 * 2
1 = 2 + 2 2 1 = 2 + 3 3
e2
function
cov(x i , x j ) = i * j
e3
Output
SEM analysis AMOS:
Explained variance (=1-error variance)
e1 e2 e3
Communalities pain depress function Initial .233 .204 .288 Extraction .365 .312 .568
Factor loadings
2 1
= 0.36
2 = 0.31 2 2 = .57 3
Communalities (explained or extracted variance)
.36
.31
.57
pain
Factor Matrixa Factor 1 .604 .558 -.754
depress
.60 .56 -.75
function
Well-being
visperc
err_v
x1 = 11 * f1 + 0 * f 2 + 1 x2 = 12 * f1 + 0 * f 2 + 2 x3 = 13 * f1 + 0 * f 2 + 3 x4 = 0 * f1 + 24 * f 2 + 4 x5 = 0 * f1 + 25 * f 2 + 5 x6 = 0 * f1 + 26 * f 2 + 6 Corr( f1 , f 2 ) =
spatial
cubes
err_c
lozenges
err_l
paragrap
err_p
verbal
sentence
err_s
wordmean
err_w
Exercise: CFA
Use the data set grnt.fem.sav and do a confirmatory factor analysis: Six observed variables: visual perception, cubes, lozenges, paragraph comprehension, sentence completion and word meaning. Model 1: The first three measures are hypothesized to measure spatial ability, the second three measures were believed to measure verbal ability. It was assumed that the latent variables spatial and verbal ability are related to a more general ability factor and should therefore covary. Draw a path diagram for the hypothesized model and estimate the model using AMOS. Are the data in agreement with the model? How good is the model fit?
Exercise - part 2
Model 2: Is the correlation between the two latent trait necessary? (Hint: make a second model and constraint the correlation to 0) Model 3: Another researchers claims that Lozenges is influenced by both latent factors. Can you support this hypothesis? Model 4: Another researchers claims that verbal ability influences spatial ability. How would you model this hypothesis (using the loadings of the items of the first model)? This is now a true SEM. Which models are equivalent structural models?
Model 1
1
Model 1 Model 2
1
Model 3 Model 4
visperc
err_v
spatial
cubes
err_c
lozenges
err_l
paragrap
err_p
verbal
sentence
err_s
wordmean
err_w
Exercise 2
See handout Victimization Chi2 test Bollen Stein Bootstrap CFI RMR RMSEA AIC BIC
Boys:
visperc
1
err_v
visperc
err_v
spatial
cubes
err_c
spatial
cubes
err_c
lozenges
err_l
lozenges
err_l
paragrap
err_p
paragrap
err_p
verbal
sentence
err_s
verbal
sentence
err_s
wordmean
err_w
wordmean
err_w
Comparison of regression weight w1 between males and females. If >2 or <-2 than significant different (but be aware of multiple testing!
\date Todays date in short format. \df Degrees of freedom \filename Name of the current AMW file.
Use \longfilename to display the complete path to the current AMW file.
\format Format name (e.g. standardised results) \gfi Goodness of fit index (GFI) \group Group name \ifi Incremental fit index (IFI) \model Model name \npar Number of distinct parameters \p p value associated with discrepancy function (test of perfect fit) \rmsea Root mean square error of approximation (RMSEA) \rmseahi Upper bound of 90% confidence interval on RMSEA \rmsealo Lower bound of 90% confidence interval on RMSEA
Constraining parameters
e2
-.24 .27
e2
-.43 .15
depress
function
depress
function
e2
-.34 .21
depress
-.46
e2
-.34 .21
depress
function
function
Measurement invariance
An observed score is said to be measurement invariant if a persons probability of an observed score does not depend on his/her group membership, conditional on the true score. That is, respondents from different groups, but with the same true score, will have the same observed score. More formally, given a persons true score, knowing a persons group membership does not alter the persons probability of getting a specific observed score.
Model 1 Model 2 Chi2 test Bollen Stein Bootstrap CFI RMR RMSEA AIC BIC
Model 3 Model 4
Course: SEM and path analysis Using AMOS to do path analysis and SEM Model specification , identification( ), and estimation Underidentified: number of model parameters < number of observed parameter: Just identified=# of model parameter=number of observed parameter df=0 Overidentified:=# of model parameter > number of observed parameter df>0 model testing is possible Evaluating model fit Interpreting parameter estimates SEM and causality