Minor - SEM

Question 1

Little and Rubin (2002)

Answer 1

3 Types of missingness: MCAR, MAR, MNAR

Question 2

Centering predictors is recommended to aid in interpretability

Answer 2

Aiken and West (1991)

Question 3

What are the benefits of SEM?

Answer 3

​to estimate and remove error; triangulate multiple measures and separate unique from measurement error; propose more complicated causal models and then test those in a confirmatory nature (model testing/fitting approach)

Question 4

What are rotations? Difference between orthogonal/oblique rotation?

Answer 4

Used in EFA and is intended to enhance the interpretability of retained factors and
yielding a simple structure where each factor explains as much variance as possible in non overlapping sets of indicators.

Orthogonal rotation (e.g., varimax rotation) - factors are all uncorrelated

Oblique rotation (e.g., promax rotation) allows for correlated factors.

oblique is preferred as some constructs may have correlated factors. Oblique also provides factor correlations.

Question 5

What is covariance?

Answer 5

The basic unit of SEM; The strength of a linear relationship between X and Y and their variabilities with a single number; unstandardized form of a correlation

Question 6

What is disturbance?

Answer 6

1 - R squared. Represents proportion of unexplained variance and is made up of measurement error and omitted causes.

Question 7

Difference between EFA/CFA?

Answer 7

EFA is a factor analysis where one seeks a simple latent variable structure without causal arrows between variables. CFA on the other hand, is described as taking a specific hypothesized structure and assessing how well it accounts for the observed relationships found within the data. EFA (unrestricted model) doesn’t require a priori specification of number of factors whereas CFA does.

CFA can do anything that EFA does and doesn’t require as many decision points (PCA vs factor analysis; rotation method; number of factors to retain) which may bias your model.

Question 8

What is PCA and when to use it?

Answer 8

Great for data reduction purposes, finding components that explain the most variance

Question 9

What is latent growth curve modeling?

Answer 9

Newsom (2015) - method that investigates level change in a variable over time and while allowing for the investigation of linear and nonlinear trends and individual differences within these trends. LGCM’s involve two latent variables—one representing the intercept and the other representing the slope​.

Question 10

What is SEM?

Answer 10

A causal inference method that takes 3 inputs (a set of theoretically-sound hypotheses; a set of questions about causal relationships; data from studies) and produces 3 outputs (estimates of model parameters; a set of logical implications of the model; the degree to which the model are supported by the data)

Question 11

What is the goal of SEM?

Answer 11

to discover a model that makes theoretical sense, is reasonably parsimonious, and has acceptably close correspondence to the data. Essentially you are establishing if your proposed model fits the data provided.

Question 12

How to determine minimum sample size?

Answer 12

N:q Rule. N represent cases needed and q represents model parameters. A ratio of 20:1 is ideal, 10:1 is acceptable. A general rule is over 200.

Question 13

Tabachnick and Fidell (2013)

Answer 13

Provides intro to matrix algebra. Covers adding/subtracting, multiplication, division of matrix algebra. Trace; Scalar; Transpose

Question 14

Loehlin (2004)

Answer 14

Covers Wright’s tracing rules for path analysis: no loops, no going forward then backward, only on curved arrow per path.

Question 15

What are the goals for model identification? 2 general rules for identifying a model?

Answer 15

to be able to express every free model parameter as a unique function of elements in the population covariance matrix

The models’ df must be at least 0 (counting rule); every latent variable including disturbance/error terms must be assigned to a scale

Question 16

2 kinds of path model?

Answer 16

Recursive and Non-recursive. Recursive is straightforward. All causal effects unidirectional.

Question 17

Kenny (1979) - Minimum number of indicators? Causal Inference Criteria?

Answer 17

3-5 per factor.

Temporal precedence (x precedes y in time)
Relationship between variables (variables are orthogonal)
nonspuriousness (ruling out confounds)

Question 18

Preacher and MacCallum (2003)

Answer 18

Critiques common missteps/assumptions often made in the SEM literature. The authors also provide 3 general recommendations regarding the use of exploratory techniques: avoid PCA unless the researcher is interested in data reduction, use a combination of criteria (like scree plot) to determine an appropriate number of factors to retain instead of using an arbitrary cutoff (eigenvalues cutoff of 1), and avoid the use of orthogonal varimax rotation as factors are rarely uncorrelated in empirical studies and oblique rotation provides factor correlations.

Question 19

Best Model Fit citation? What is model fit? Best indices to use? Indicators of multicollinearity?

Answer 19

West, Taylor, and Wu (2012)

entails attempting to match the variances, covariances, and means of observed data in your model.

Chi square: NS
CFI > .95
RMSEA

Question 20

Finney and DiStefano (2013)

Answer 20

analyzing non normal/categorical data

5 ordinal categories = continuous

nonnormal data assessed by skewness (>2) and kurtosis (>7)

Satorra Bentler (MLM) robust standard errors are for non-normal data
Yuan Bentler (MLR) robust standard errors for nonnorma/missing data

Question 21

Newsom (2015)

Answer 21

Longitudinal Models: stability is absence of change and change is absence of stability. cross-lagged models intended to examine stability over time and causal precedence to determine the causal direction between variables and whether one variable causes another

Question 22

What is Factor Invariance Testing? What are the 4 types?

Answer 22

investigates measurement properties of a factor across groups in order to determine if the measure translates well across groups.

Weak factorial invariance: loadings = across groups
Strong: loadings & measurement intercepts = “”
Strict: loadings intercepts, measure residual
variances “”
Structural: loadings, intercepts, residual variances, factor variances, and factor means “”

Question 23

Estimators for binary/ordinal variables

Answer 23

DWLS/WLS (Diagonally weighted least squares)