Upload
others
View
18
Download
0
Embed Size (px)
Citation preview
An Introduction to Latent Variable Modeling
Karen Bandeen-Roche Qian-Li Xue
Johns Hopkins Departments of
Biostatistics and Medicine
October 27, 2016
LATENT VARIABLES:
TRUTH, LIES, AND EVERYTHING BETWEEN
Karen Bandeen-Roche Department of Biostatistics Johns Hopkins University
ABACUS Seminar Series November 28, 2007
Objectives • What is a latent variable (LV)? • What are some common LV models?
• What are major features of LV modeling? – Hierarchical: structural and measurement components – Fitting – Evaluating fit – Predictions – Identifiability
• Why should I consider using—or decide against using—LV models?
Part I: Overview
“LATENT”? 1. Present or potential but not evident or active: latent talent. 2. Pathology. In dormant or hidden stage: a latent infection. 3. Biology. Undeveloped, but capable of normal growth under the
proper conditions: a latent bud. 4. Psychology. Present and accessible in the unconscious mind,
but not consciously expressed.
The American Heritage Dictionary of English Language, Fourth Edition, 2000
“existing in hidden or dormant form but usually capable of being
brought to light” Merriam-Webster’s Dictionary of Law, 1996
“LATENT” • “…concepts in their purest form… unobserved or unmeasured … hypothetical”
Bollen KA, Structural Equations with Latent Variables, p. 11, 1989
• “…in principle or practice, cannot be observed”
Bartholomew DJ, The Statistical Approach to Social Measurement, p. 12
• “Underlying: not directly measurable. Existing in hidden form but usually capable of being measured indirectly by observables.”
Bandeen-Roche K, Synthesis, 2006
“LATENT VARIABLES”?
• Ordinary linear regression model: Yi = outcome (measured) Xi = covariate vector (measured) εi = residual (unobserved)
Yi = XiT β+ εi
Ordinary Linear Regression Residual as Latent Variable
X
.
.
. . . .
.
. .
. . . .
.
Y . ε Y X ε
Boxes denote observables Ovals denote “unobserved” Straight arrows are causal Curved arrows denote association
1
Yi = XiT β+ εi
Mixed effect / Multi-level models Random effects as Latent Variables
time
vital
non-vital
Yij = β0 + β1 xi + β2 tij + β3 xi·tij + eij
.
.
. .
. .
.
.
.
. .
. .
.
.
.
. .
. .
.
.
.
. .
. .
.
0
β0 + β1
β0
β2
β2 + β3
Mixed effect / Multi-level models Random effects as Latent Variables
• b0i = random intercept b2i = random slope (could define more)
• Population heterogeneity captured by spread in intercepts, slopes
time
vital
non-vital
Yij = β0 + b0i + β1 xi + β2 tij + b2i tij + β3 xi·tij + eij
.
.
. .
. .
.
.
.
. .
. .
.
.
.
. .
. .
.
.
.
. .
. .
.
0
β0 + β1
β0
β2
β2 + β3
+ b0i slope: - |b2i|
Mixed effect / Multi-level models Random effects as Latent Variables
time
vital
non-vital
Yij = β0 + b0i + β1 xi + β2 tij + b2i tij + β3 xi·tij + eij
.
.
. .
. .
.
.
.
. .
. .
.
.
.
. .
. .
.
.
.
. .
. .
.
0
β0 + β1
β0
β2
β2 + β3
+ b0i slope: - |b2i|
Y X ε
t b
1
1
Latent variable model
Inflammation Mobility
X1
Xp
…
Y1
YM
…
δp
δ1 ε1
εM
ξ η
ζ1
1
1
1
1
1
Linear structural equations model with latent variables (LISREL):
Yij = outcome (jth measurement per “person” i) xij = covariate vector (corresponds to jth measurement, person i) λy
j = outcome “loading” (relates outcome LV to Y measurement) ηi = latent outcome=random coefficient vector, person i λx
j= covariate "loading" (relates covariate LV to jth x measurement) ξi = latent covariate = random coefficient vector, person i εij = observed response residual δij = observed covariate residual ςi = latent response residual vector (specified distribution)
Yij = λy
jTηi + εij
Xij = λjXTξi + δij
ηi = Bηi + Γξi + ςi
> My sense: It’s the unknown λj that distinguishes above as a
“latent variable model” in most minds
"LATENT VARIABLES”?
Latent Variables: What? Integrands in a hierarchical model
• Observed variables (i=1,…,n): Yi=M-variate; xi=P-variate • Focus: response (Y) distribution = GYx(y/x) ; x-dependence • Model:
– Yi generated from latent (underlying) Ui: (Measurement)
– Focus on distribution, regression re Ui:
(Structural)
• Overall, hierarchical model:
);( βxuF xU
);,)(, πxuUyF xUY =
∫ == )(),()( , xudFxuUyFxyF xUxUYxY
)( xyG xY
Latent variable model
Inflammation Mobility
X1
Xp
…
Y1
YM
…
δp
δ1 ε1
εM
ξ η
Measurement Measurement
ζ1
Structural
Well-used latent variable models
Latent variable scale
Observed variable scale
Continuous Discrete
Continuous Factor analysis LISREL
Discrete FA IRT (item response)
Discrete Latent profile Growth mixture
Latent class analysis, regression
General software: MPlus, Latent Gold, WinBugs (Bayesian), NLMIXED (SAS) gllamm (Stata)
Why do people use latent variable models?
• The complexity of my problem demands it • NIH wants me to be sophisticated • Reveal underlying truth (e.g. “discover”
latent types) • Operationalize and test theory • Sensitivity analyses • Acknowledge, study issues with
measurement; correct attenuation; etc.
Latent Variable Models: Philosophy • Why?
– To operationalize / test theory – To learn about measurement errors, differential reporting – They summarize multiple measures parsimoniously – To describe population heterogeneity – Popperian learning
• Why not? – Their modeling assumptions may determine scientific conclusions – Their interpretation may be ambiguous
• Nature of latent variables? • Uniqueness (identifiability) • What if very different models fit comparably? (estimability) • Seeing is believing
• Import: They are widely used
Part II: Major elements of
latent variable modeling
1. Model choice
Example Pro-inflammation in Older Adults
• Inflammation: central in cellular repair • Hypothesis: dysregulation=key in accel. aging
– Muscle wasting (Ferrucci et al., JAGS 50:1947-54; Cappola et al, J Clin Endocrinol Metab 88:2019-25)
– Receptor inhibition: erythropoetin production / anemia (Ershler, JAGS 51:S18-21)
Stimulus (e.g. muscle damage)
IL-1# TNF-α IL-6 CRP
inhibition
up-regulation
# Difficult to measure. IL-1RA = proxy
Example Pro-inflammation in Older Adults
Inflam.
regulation Adverse outcomes
Y1
Yp
…
Determinants
e1
ep
Theory informs
relations (arrows)
ς λ1
λp
Measurement
Structural
Pro-inflammation in Older Adults InCHIANTI data (Ferrucci et al., JAGS, 48:1618-25) • LV method: factor analysis model
– Continuous indicators, latent variables – Two distinct underlying variables – Down-regulation IL-1RA path=0 – (Conditional independence)
Inflammation 2
Down-reg.
IL-6
TNFα
CRP IL-1RA
IL-18
Inflammation 1
Up-reg.
“LATENT VARIABLES”? Linear structural equations model with latent variables (LISREL):
Yij = outcome (jth measurement per “person” i) xij = covariate vector (corresponds to jth measurement, person i) λy
j = outcome “loading” (relates outcome LV to Y measurement) ηi = latent outcome=random coefficient vector, person i λx
j= covariate "loading" (relates covariate LV to jth x measurement) ξi = latent covariate = random coefficient vector, person i εij = observed response residual δij = observed covariate residual ςi = latent response residual vector (specified distribution)
Yij = λy
jTηi + εij
Xij = λjXTξi + δij
ηi = Bηi + Γξi + ςi
Latent variable models Factor Analysis Measurement Model
X=Λxξ+δ
Φ=Var(ξ); Θδ=Var(δ)
111111 ... iippxixi δξλξλ +++=x
221212 ... iippxixi δξλξλ +++=x
imipxMpixMiM δξλξλ +++= ...11x
Latent variable models Factor Analysis Measurement Model
X=Λxξ+δ Φ=Var(ξ); Θδ=Var(δ)
• Assumptions • Most frequently: (ξ, δ) ~ multivariate normal • ξ δ • Constraints on ϕ, Θδ (“theory”)
• Ex: Θδ diagonal – indicators uncorrelated given LVs i.e. factor model; conditional independence
π
2. Fitting
Estimation Overview • Most common: Likelihood-‐based approaches
– Primary challenge: the integral
• Approxima=on (Laplace) • Numerical integra=on • Stochas=c integra=on
– Gradient methods – E-‐M algorithm
• Bayesian approaches (MCMC)
• Least squares or analogs
€
FY |X (y | x) = FY |u,x∫ (y | u,x)dFu|x(u | x)
ML Estimation Factor model
• Likelihood has closed form (MVN)
~
1
~2 ||)2()|( 1
1
ii xxM
M
mexf
−Σʹ′−−−
=ΣΠΠ=θ
Σ=Θ+ΛΘΛ= δξ')( xxxVar
δξ +Λ= xX
Pro-inflammation in Older Adults (Bandeen-Roche et al., Rejuv Res)
• LV method: factor analysis model
Inflammation 2
Down-reg.
IL-6
TNFα
CRP IL-1RA
IL-18
Inflammation 1
Up-reg.
.36
. 59 . 45 . 31
. 31
-.59
-.40
.20
3. Evaluating fit
Methods Global measures
• Goodness of fit testing
– Hypothesis: H0: GY|X(y|x) = FY|X(y|x;π,β) for some (π,β) ε Θ
– Method: Deviance goodness of fit testing, analogs – Usual issues for quality of asymptotic distribution approximation – Inflammatory analysis: Deviance goodness of fit pvalue > 0.5
• Global fit indices: “Hundreds” of them
Methods Residual checking
• Per-item: Observed – expected
• Residuals with respect to association structure – Continuous Y: Covariance or correlation matrix residuals S-
– Categorical Y: • Odds ratio matrix residuals: Q- Implied, Q has elements [ad/bc]ij from
cross-tabulation of items i & j • O-E cell counts for the full cross-tabulation of items (I1xI2x…xIM cells, where
Ij denotes the number of categories for item j)
• All cases: normalized residuals most useful
Σ
ψ
Example Residual checking
• NFΚB-gated systemic inflammation
Methods Other
• Posterior predictive checking Gelman et al., Statistica Sinica, 1996
• Pseudo-value analysis: More to come
4. Prediction
Latent variable scoring Overview
• Task: Estimate persons’ underlying status – “fill in” values for the Ui
• Fundamental tool: Posterior distribution
Latent variable scoring Posterior mean estimation
• Posterior mean = most common method – Typically: Empirical Bayes (filling in estimates
for parameters) – Minimizes expected posterior quadratic loss
• Linear case (LISREL): Yields Best Unbiased Linear Predictor (BLUP)
Latent variable scoring LISREL (factor) measurement model
• Posterior mean is closed form linear
,
• “Regression method”
xTX
1ˆˆˆˆ −ΣΛΦ=ξ Σ= Var(X)
Latent variable scoring LISREL (factor) measurement model
• Alternative method: “Bartlett” scores – Paradigm: treat ξi as fixed parameters per i;
estimate these via weighted least squares
δi ~ N(0, )
• Which is better? – Depends on analytic purpose
,ˆiii δξ +Λ= XX Σ̂
Latent variable scoring Frequent purpose: “multi-stage” regression
• Step 1: Fit full latent variable measurement model(s) (Y,X) ,
• Step 2: Obtain predictions Oi given Yi,
and/or Xi, • Step 3: Obtain via regression of Oi on Xi
or Yi on Oi, as case may be
ΛY ΛX
ΛYΛX
B
Latent variable scoring Frequent purpose: “multi-stage” regression
• Result: In the fully linear model, provided that estimators in Step 1 are consistent: – (a) When the covariate is being predicted, employing
the regression method in Steps 2-3 consistently estimates B
– (b) When the outcome is being predicted, employing the least squares method in Steps 2-3 consistently estimates B
• Brief rationale for (a): method is analogous to regression calibration with replicates Carroll & Stefanski, JASA, 1990
5. Identifiability
One last issue Identifiability
• Models can be too big / complex • A model is non-identifiable if distinct
parameterizations lead to identical data distributions – i.e. analysis not grounded in data
• Weak identifiability is common too: – Analysis only indirectly grounded in data (via
the model)
Identifiability
data (ground)
model
analysis
strong
Identifiability
data (ground)
model
analysis
weak
Identifiability
data (ground)
model
analysis
non
Objectives • What is a latent variable (LV)? • What are some common LV models?
• What are major features of LV modeling? – Hierarchical: structural and measurement components – Fitting – Evaluating fit – Predictions – Identifiability
• Why should I consider using—or decide against using—LV models?