lava REFERENCE CARD

Linear Latent Variable Models in R

M ODEL BUILDING N ON - LINEAR CONSTRAINTS

• Initialize model (empty or multivarite regr. model) m <- lvm(); m <- lvm(c(y1,y2)~x+z) • Non-linear parameter constraints constrain(m,psi~beta+gamma) <- f unct
• Initialize from list of regression models m <- lvm(list(y~x,y~z,...)) • Non-linear regression (covariate x) constrain(m,psi~beta+x) <- f unct
• Add extra regression associations (slopes) regression(m) <- c(y1,y3)~u • Add extra parameters to model parameter(m) <- ~beta+gamma
• Add correlation between residual terms covariance(m) <- y1~y2+y3 • Add predictor/exogenous variable to model exogenous(m) <- ~x1+x2
• Remove associations between variables cancel(m) <- ~y1+y2 • Random slopes (x name of covariate) regression(m) <- y1~f(eta,x)
• Add variables addvar(m) <- ~y1+y2 • Print non-linear constraints constrain(m)
• Remove variables kill(m) <- ~y1+y2
• Code as latent (reverse with arg. clear=TRUE) latent(m) <- ~y1+y2 M ODEL INSPECTION
• Binary variables; library(lava.tobit) binary(m) <- ~y1+y2
• Examine parameter constraints intercept,covariance,regression,constrain
E QUALITY CONSTRAINTS • Extract variable names exogenous,endogenous,latent,manifest,vars
• Submodel (see also measurement) subset(m,~y1+y2+eta+x)
• Intercepts • List parameter names coef(m,mean=TRUE,labels=FALSE,...)
◦ Constrain intercepts to be identical intercept(m) <- c(y1,y2)~f(a) • Parents and children of nodes (union)) parents(m,~y1+y2);children(m,~x1+x2)
◦ Simultaneously fix several intercepts intercept(m,~y1+y2) <- list("a",2) • Extract (directed) pathways between variables path(m,y~x)

• Variance/covariance parameters P LOTTING

◦ Fix variance term and covariance between resid- covariance(m) <- y1~f(y1,v1)+f(y2,1)
ual to v1 resp. 1 • Plot method (lvm and lvmfit) plot(m,labels=TRUE,...)
◦ Fix multiple variance parameters covariance(m,~y1+y2) <- list("a",2) • Change appearance of nodes nodecolor(m,~y1+x,labcol=c("red","blue"),
◦ Simultaneously fix several covariance parameters covariance(m,c(y1,y2)~y2+y3) <- border,lwd=2, ...) <- c("blue","red")
list(2,"a","b",1) • Change label and appearance of edges edgelabels(m,y~x+z,col,...) <-
expression(rho)
• Slope/regression parameters
• Change labels of nodes (e.g. math expressions) labels(m) <- c(eta=expression(eta))
◦ y1 = x + az + · · · regression(m,c(y1,y2)~x+z) <-
y2 = x + bx + · · · list(1,"a",2,"b") • Extract graphNEL object (library(Rgraphviz)) Graph(m)

◦ y1 = x + · · · , y2 = az + · · · regression(m,c(y1,y2)~x+z) <- list(1,"a") S TATISTICAL I NFERENCE

◦ yi = ax + · · · regression(m) <- c(y1,y2,y3)~f(x,a)
• Estimate parameters (default MLE) e <- estimate(m,data,estimator,...)
• Fix parameters defined by index (see coef) parfix(m,c(3,4,12)) <- list(1,"a",2)
• Estimate multigroup model (default MLE) estimate(list(m,...),list(data,...),...)
• Label all free parameters (see multigroup) m <- baptize(m)
• Estimate under MAR assumption estimate(m,data,missing=TRUE,...)
• Remove linear constraints by fixing to NA (applies regression(m) <- c(y1,y2)~f(x,NA)
• Likelihood ratio tests vs. saturated model compare(e)
also to the intercept and covariance methods)
• Likelihood ratio tests compare(e1,e2,e3,...)
• Bracket notation. Define intercept and variance of regression(m) <- y[0:v]~f(x[a,1],b)
residual of y to 0 and ’v’ and of x to ’a’ and 1. And • Model indices based on score tests (or Wald tests) modelsearch(e,...)
define E(y|x) = b · x. • Identify empirical equivalent models equivalence(e,y~x,k=1,...)
• Calculate indirect and total effects of x on y effects(e, y~x)
S IMULATION • Non-linear constraints and approx. std.errors constraints(e)
• Mixtures of LLVM; library(lava.mixture) mixture(list(m1,m2),data)
• Simulate 100 observations from model m sim(m,100,...)
• Extract various likelihood summaries coef,score,information,logLik,AIC,gof
• Simulate with the slope-parameter of x on y set to sim(m,100,p=c("y"=1, "y<-x"=-2),...)
−2, and intercept of y to 1 (see coef) • Clustered correlated data estimate(m,data,cluster="id",...)
• Define conditional distribution distribution(m,~y1+y2) <- • Robust standard errors coef(e,type="robust")
function(n,mu,var,...) ... • Test for linearity; library(gof) cumres(e,...)
◦ Predefined distributions binomial.lvm,,uniform.lvm,normal.lvm,poisson.lvm,weibull.lvm(shape,scale,...),...
• Non-parametric bootstrap bootstrap(e,R=100,...)
• Define functional form of predictor on response functional(m,y~x) <- f unct • Likelihood-based confidence limits confint(e,idx,profile=TRUE,...)

c 2011 Klaus Kähler Holst

 E XAMPLES
Linear Latent Variable Models in R

S TRUCTURAL E QUATION M ODEL S IMULATION

MIMIC model Weibull with exponential distributed censoring

y1 0.7939 y2 1.208 y3 1.238

> m <- lvm(c(y1,y2,y3)~u) > m <- lvm(y~x1+x2+x3)

> regression(m) <- u ~ x1+x2
1 0.9925 1.009

> distribution(m,~y) <- weibull.lvm(shape=0.5,cens=rexp)

> latent(m) <- ~ u u 1.349
> distribution(m,~x3) <- binomial.lvm()
> d <- sim(m,100) > d <- sim(m,100)
> e <- estimate(m,d) 1.105 0.9037

> plot(e) x1 x2
M ULTIVARIATE PROBIT

R ANDOM REGRESSION
> m <- lvm(c(y1,y2)~f(x,b)+f(z,b))
> binary(m) <- ~y1+y2
Random slopes allowing for unbalanced designs
> covariance(m) <- y1~y2
> estimate(m,data,control=list(trace=1))
> m <- lvm(c(y1,y2,y3)~f(eta,1))
> regression(m,c(y1,y2,y3)~u) <- list("x1","x2","x3") M ULTIGROUP ANALYSIS I
> intercept(m,~y1+y2+y3) <- list("mu")
> covariance(m,~y1+y2+y3) <- list("v","v","v") P
log L(θ|d) = i log Li (θ|di )
> latent(m) <- ~u+eta
> estimate(m,data,missing=TRUE)
> estimate(list(m1,m2,m3),list(d1,d2,d3))
G RAPHICS I I NDIRECT EFFECTS . TOBIT /P ROBIT MODEL
y u E(y | x, z) = a(x + z)

> m <- lvm(list(y~z+x,z~x))

w v b > d <- transform(sim(m,100),z=factor(z>0),y=Surv(ifelse(y<1,y,1),y<1))
> m <- lvm(list(y ~ b+v+w, c(b,w) ~ x+z, u ~ b)) > e <- estimate(m,d)
> latent(m) <- ~b > effects(e,y~x)
> plot(m) x z
N ON - LINEAR REGRESSION
G RAPHICS II Bi-variate non-linear regression with random intercept. Estimated via Fischer scoring

> m0 <- lvm(c(y1,y2)~f(x,0)+f(eta,1))

> m <- lvm(y~x+z+u) u > latent(m0) <- ~eta
x
> labels(m) <- c(y=expression(psi), z=expression(zeta)) > covariance(m0) <- c(y1,y2) ~ f(0.01)
> nodecolor(m,~y+z+x,border=c("white","white","black"), > covariance(m0) <- c(eta) ~ f(0.01)
+ labcol="white", lwd=c(0,0,5)) <- > d <- sim(m0,50)[,manifest(m0)]
ρ
+ c("orange","indianred","lightgreen") > d <- transform(d,
> edgelabels(m,y~z+x, cex=c(2,3), col=c("orange","black"), y1=y1+pnorm(2*x),
+ labcol="darkblue",lwd=c(3,1)) <- expression(phi,rho) y2=y2+pnorm(2*x))
> plot(m,layoutType="circo") ψ ζ
φ
> m <- m0
> parameter(m) <- ~ nu+alpha+xi
> intercept(m) <- c(y1,y2) ~ f(mu)
I NSTRUMENTAL VARIABLE > covariance(m) <- c(y1,y2) ~ f(v)
> covariance(m) <- eta ~ f(zeta)
IV estimator (not available with e.g. non-recursive structures) > constrain(m, mu ~ x + alpha + nu + xi) <- function(x) pnorm(x[1]*x[2]+x[4]) + x[3]
> e <- estimate(m,d,control=list(trace=1,method="NR",gamma=0.99))
> estimate(m,data,estimator="IV")

c 2011 Klaus Kähler Holst