Incorporating Stability Information into Cross-Sectional Estimates.
stim
The stim package fits the Stability Informed Model which incorporates variable stability–how a variable correlates with future versions of itself–into cross-sectional estimates. Assuming the process is stationary, the model is specified correctly, and the stability values are correct, the Stability Informed Model can estimate parameters that are unbiased for cross-lagged (longitudinal) effects when only cross-sectional data are available.
For more information on the Stability Informed Model see https://psyarxiv.com/vg5as
This tutorial outlines how to estimate a Stability Informed Model using the stim package within an SEM framework.
Installation
You can install the development version of stim from GitHub with
devtools::install_github("https://github.com/AnnaWysocki/stim")
Example
Let’s create some data to use in our example.
library(stim)
S <- matrix(c(1, .3, .3,
.3, 1, .3,
.3, .3, 1), nrow = 3, ncol = 3)
example_data <- as.data.frame(MASS::mvrnorm(n = 300, mu = rep(0, 3), Sigma = S,
empirical = TRUE))
# Add column names to dataset
colnames(example_data) <- c("X", "Y", "Z")
stim Function Overview
Estimate a single or a set of Stability Informed Models using the stim()
function.
stim()
has five arguments
- data : A data frame with the measured variables. Not needed if S is provided.
- S: A covariance matrix for the measured variables. Not needed if data is provided.
- n: Number of observations in data set. Not needed if data is provided.
- model: An object with the cross-sectional model specified in lavaan syntax.
- stability: an object with the stability information for each variable in the model
More details on the model
and stability
arguments can be found below.
The model
Argument
Input an object with the cross-sectional model specified in lavaan syntax. The model syntax should be specified as a cross-sectional path model in lavaan (See https://lavaan.ugent.be/tutorial/tutorial.pdf for information on lavaan syntax).
This input determines what parameters/effects are estimated. Note, the Stability Informed Model can estimate a maximum of (p * (p -1))/2 parameters (where p is the number of measured variables). These parameters can be, for example, cross-lagged effects or residual covariances.
To estimate the effect of X on Y, I could create the following object
model <- 'Y ~ X' # outcome ~ predictor
More complex models can be specified as well.
model2 <- 'Y ~ X
Z ~ X + Y'
The default is to constrain all residual covariances to 0. But this constraint can be relaxed by specifying a residual covariance in the model syntax.
model2 <- 'Y ~ X
Z ~ X + Y
X ~~ Y' # Allows X and Y to have covarying residuals
The above model object specifies 4 estimated parameters, but, with 3 measured variables, the Stability Informed Model can only estimate 3 parameters. The remaining effects can either be fixed to 0 or fixed to a non-zero value.
model2 <- 'Y ~ .6 * X # fix effect of X on Y to .6
Z ~ X + Y
X ~~ Y'
Labels can be specified for the estimated parameters.
model2 <- 'Y ~ .6 * X
Z ~ Effect1 * X + Y # label the estimated effect of X on Z
X ~~ Y'
If no label is specified for a cross-lagged parameter, the default label is ‘CL’ and a subscript with the predictor name and the outcome name.
Residual covariances are labeled ‘RCov’ and a subscript with the names of the two variable whose residuals are covarying.
Inputs for the stability
Argument
Input a object with the stability information for each variable in the model.
To fit model2
, the stability input should have a stability value for X, Y, and Z.
stability <- c(X = .5, Y = .1, Z = .1)
The elements or columns in the stability object need to be named, and the names must match the variable names in the data
or S
input.
Multiple stability values can be specified for each variable. This results in multiple Stability Informed Models being estimated (one for each stability condition).
stability <- data.frame(X = c(.5, .55), Y = c(.1, .15), Z = c(.1, .2))
rownames(stability) <- c("Model 1", "Model 2")
stability
#> X Y Z
#> Model 1 0.50 0.10 0.1
#> Model 2 0.55 0.15 0.2
If this is the stability
input, two models will be estimated. One model where the stability values for X, Y, and Z are .5, .1, and .1, respectively, and one where the stability values for X, Y, and Z are .55, .15, and .2, respectively.
Estimate the Stability Informed Model
modelFit <- stim(data = example_data, model = model2, stability = stability)
#> StIM: Stability Informed Models
#> -------------------------------------
#> -------------------------------------
#>
#> Variables (p): 3
#> Sample Size (n): 300
#> Estimated Parameters (q): 3
#> Degrees of Freedom: 0
#> Number of Models Estimated: 2
#>
#> -------------------------------------
Instead of the data input, I could also use the covariance matrix and sample size inputs.
modelFit <- stim(S = cov(example_data), n = nrow(example_data), model = model2, stability = stability)
Some information about the model(s) is automatically printed out when the stim()
function is run. The output from this function is an object of type stim
. When the summary()
function is used on a stim
object, a summary of the estimated Stability Informed Models will be printed.
summary(modelFit)
#> StIM: Stability Informed Models
#> -------------------------------------
#> -------------------------------------
#>
#> Variables (p): 3
#> Sample Size (n): 300
#> Estimated Parameters (q): 3
#> Degrees of Freedom: 0
#>
#> -------------------------------------
#> Model 1
#>
#> Stability:
#> X Y Z
#> 0.5 0.1 0.1
#>
#> Autoregressive Effects:
#> ARX ARY ARZ
#> 0.50083541 -0.07953268 -0.24509203
#>
#> Cross Lagged Effects:
#> Effect Estimate Standard.Error P.Value
#> Effect1 0.466 0.213 0.029
#> CLYZ 0.687 1.127 0.542
#>
#> Residual Covariances:
#> Effect Estimate Standard.Error P.Value
#> RCovYX 0.011 0.068 0.873
#>
#> -------------------------------------
#> Model 2
#>
#> Stability:
#> X Y Z
#> 0.55 0.15 0.2
#>
#> Autoregressive Effects:
#> ARX ARY ARZ
#> 0.55091897 -0.02944912 -0.16022850
#>
#> Cross Lagged Effects:
#> Effect Estimate Standard.Error P.Value
#> Effect1 0.330 0.643 0.608
#> CLYZ 0.874 3.049 0.775
#>
#> Residual Covariances:
#> Effect Estimate Standard.Error P.Value
#> RCovYX -0.026 0.067 0.696
#>
#> -------------------------------------
#>
Eploring the stim
Output Object
The object modelFit
contains a list with information for the Stability Informed Model
- n: Sample size
- p: Number of measured variables used in the Stability Informed Model
- q: Number of estimated parameters
- df: Degrees of freedom
- stability : A table of the stability conditions
- CLEffectTable : A table with information on which cross-lagged effects are estimated
- CLMatrices: A list of matrices (1 for each Stability Informed Model that was estimated) with the estimated cross-lagged effects and their associated standard errors and p-values.
- RCovMatrices: A list of matrices (1 for each Stability Informed Model that was estimated) with the estimated residual covariances and their associated standard errors and p-values.
- ARVector: A list of vectors (1 for each Stability Informed Model that was estimated) with the values for the auto-regressive effects.
- lavaanObjects : A list of lavaan objects (1 for each Stability Informed Model that was estimated)
- NoWarnings : A vector with information on whether there were any errors or warnings for each of the estimated models
- CSModelSyntax : The user-specified model syntax (input for model argument)
- SIMSyntax : The syntax for the Stability Informed Model–model syntax for the lavaan function
- modelImpliedEquations : Model implied equations for the latent covariances and auto-regressive effects
stability
A table of the stability conditions. Each row contains the stability information for one Stability Informed Model.
modelFit$stability
#> X Y Z
#> Model 1 0.50 0.10 0.1
#> Model 2 0.55 0.15 0.2
CLEffectTable
A table with information on the cross-lagged paths. It has the predictor and outcome names, cross-lagged effect labels, and whether the cross-lagged path is estimated or constrained.
modelFit$CLEffectTable
#> predictor outcome name estimate
#> 1 X_0 Y 0.6 No
#> 2 X_0 Z Effect1 Yes
#> 3 Y_0 Z CLYZ Yes
CLMatrices
A list of matrices with the estimated cross-lagged effects and standard errors and p-values for each of the estimated cross-lagged effects. Each matrix corresponds to one of the estimated Stability Informed Models.
modelFit$CLMatrices
#> $Model1
#> Effect Estimate Standard.Error P.Value
#> 3 Effect1 0.466 0.213 0.029
#> 6 CLYZ 0.687 1.127 0.542
#>
#> $Model2
#> Effect Estimate Standard.Error P.Value
#> 3 Effect1 0.330 0.643 0.608
#> 6 CLYZ 0.874 3.049 0.775
RCovMatrices
A list of matrices with the estimated residual covariances and their standard errors and p-values. Each matrix corresponds to one of the estimated Stability Informed Models.
modelFit$RCovMatrices
#> $Model1
#> Effect Estimate Standard.Error P.Value
#> 16 RCovYX 0.011 0.068 0.873
#>
#> $Model2
#> Effect Estimate Standard.Error P.Value
#> 16 RCovYX -0.026 0.067 0.696
ARVector
A list of vectors with the values for each auto-regressive effect. Each vector corresponds to one of the estimated Stability Informed Models.
modelFit$ARVector
#> $Model1
#> ARX ARY ARZ
#> 0.50083541 -0.07953268 -0.24509203
#>
#> $Model2
#> ARX ARY ARZ
#> 0.55091897 -0.02944912 -0.16022850
lavaanObjects
A list of lavaan objects (1 for each Stability Informed Model)
To output the lavaan object easily, you can use the lavaanSummary() function
lavaanSummary(modelFit)
#> Model 1
#> lavaan 0.6-12 ended normally after 180 iterations
#>
#> Estimator ML
#> Optimization method NLMINB
#> Number of model parameters 12
#>
#> Number of observations 300
#>
#> Model Test User Model:
#>
#> Test statistic 0.000
#> Degrees of freedom 0
#>
#> Parameter Estimates:
#>
#> Standard errors Standard
#> Information Expected
#> Information saturated (h1) model Structured
#>
#> Latent Variables:
#> Estimate Std.Err z-value P(>|z|)
#> X_0 =~
#> X (ARX) 0.500 NA
#> Y 0.600
#> Z (Eff1) 0.466 0.213 2.184 0.029
#> Y_0 =~
#> X 0.000
#> Y (ARY) -0.079 0.036 -2.203 0.028
#> Z (CLYZ) 0.687 1.127 0.610 0.542
#> Z_0 =~
#> X 0.000
#> Y 0.000
#> Z (ARZ) -0.245 0.292 -0.838 0.402
#>
#> Covariances:
#> Estimate Std.Err z-value P(>|z|)
#> X_0 ~~
#> Y_0 (CvXY) 0.299 0.060 4.977 0.000
#> Z_0 (CvXZ) 0.299 0.060 4.977 0.000
#> Y_0 ~~
#> Z_0 (CvYZ) 0.299 0.060 4.977 0.000
#> .X ~~
#> .Y (RCYX) 0.011 0.068 0.160 0.873
#>
#> Variances:
#> Estimate Std.Err z-value P(>|z|)
#> X_0 (VarX) 1.000
#> Y_0 (VarY) 1.000
#> Z_0 (VarZ) 1.000
#> .X 0.747 0.081 9.175 0.000
#> .Y 0.659 0.087 7.535 0.000
#> .Z 0.225 1.430 0.157 0.875
#>
#> Constraints:
#> |Slack|
#> ARX - (0.5/VarX) 0.000
#> ARY - ((0.1-0.6*CovXY)/VarY) 0.000
#> ARZ-((0.1-(Effect1*CovXZ+CLYZ*CvYZ))/VrZ) 0.000
#> CovXY - ((0.6*VarX+ARY*CovXY)*ARX+RCovYX) 0.000
#> CvXZ-((Effct1*VrX+CLYZ*CvXY+ARZ*CXZ)*ARX) 0.000
#> CYZ-(0.6*(E1*VX+CLYZ*CXY+ARZ*CXZ)+(E1*CXY 0.000
#>
#> Model 2
#> lavaan 0.6-12 ended normally after 193 iterations
#>
#> Estimator ML
#> Optimization method NLMINB
#> Number of model parameters 12
#>
#> Number of observations 300
#>
#> Model Test User Model:
#>
#> Test statistic 0.000
#> Degrees of freedom 0
#>
#> Parameter Estimates:
#>
#> Standard errors Standard
#> Information Expected
#> Information saturated (h1) model Structured
#>
#> Latent Variables:
#> Estimate Std.Err z-value P(>|z|)
#> X_0 =~
#> X (ARX) 0.550 NA
#> Y 0.600
#> Z (Eff1) 0.330 0.643 0.513 0.608
#> Y_0 =~
#> X 0.000
#> Y (ARY) -0.029 0.036 -0.816 0.415
#> Z (CLYZ) 0.874 3.049 0.286 0.775
#> Z_0 =~
#> X 0.000
#> Y 0.000
#> Z (ARZ) -0.160 0.704 -0.227 0.820
#>
#> Covariances:
#> Estimate Std.Err z-value P(>|z|)
#> X_0 ~~
#> Y_0 (CvXY) 0.299 0.060 4.977 0.000
#> Z_0 (CvXZ) 0.299 0.060 4.977 0.000
#> Y_0 ~~
#> Z_0 (CvYZ) 0.299 0.060 4.977 0.000
#> .X ~~
#> .Y (RCYX) -0.026 0.067 -0.391 0.696
#>
#> Variances:
#> Estimate Std.Err z-value P(>|z|)
#> X_0 (VarX) 1.000
#> Y_0 (VarY) 1.000
#> Z_0 (VarZ) 1.000
#> .X 0.694 0.081 8.530 0.000
#> .Y 0.646 0.087 7.392 0.000
#> .Z 0.041 4.668 0.009 0.993
#>
#> Constraints:
#> |Slack|
#> ARX - (0.55/VarX) 0.000
#> ARY - ((0.15-0.6*CovXY)/VarY) 0.000
#> ARZ-((0.2-(Effect1*CovXZ+CLYZ*CvYZ))/VrZ) 0.000
#> CovXY - ((0.6*VarX+ARY*CovXY)*ARX+RCovYX) 0.000
#> CvXZ-((Effct1*VrX+CLYZ*CvXY+ARZ*CXZ)*ARX) 0.000
#> CYZ-(0.6*(E1*VX+CLYZ*CXY+ARZ*CXZ)+(E1*CXY 0.000
You can also print a subset of the lavaan objects by using the subset
argument.
lavaanSummary(modelFit, subset = 1)
#> lavaan 0.6-12 ended normally after 180 iterations
#>
#> Estimator ML
#> Optimization method NLMINB
#> Number of model parameters 12
#>
#> Number of observations 300
#>
#> Model Test User Model:
#>
#> Test statistic 0.000
#> Degrees of freedom 0
#>
#> Parameter Estimates:
#>
#> Standard errors Standard
#> Information Expected
#> Information saturated (h1) model Structured
#>
#> Latent Variables:
#> Estimate Std.Err z-value P(>|z|)
#> X_0 =~
#> X (ARX) 0.500 NA
#> Y 0.600
#> Z (Eff1) 0.466 0.213 2.184 0.029
#> Y_0 =~
#> X 0.000
#> Y (ARY) -0.079 0.036 -2.203 0.028
#> Z (CLYZ) 0.687 1.127 0.610 0.542
#> Z_0 =~
#> X 0.000
#> Y 0.000
#> Z (ARZ) -0.245 0.292 -0.838 0.402
#>
#> Covariances:
#> Estimate Std.Err z-value P(>|z|)
#> X_0 ~~
#> Y_0 (CvXY) 0.299 0.060 4.977 0.000
#> Z_0 (CvXZ) 0.299 0.060 4.977 0.000
#> Y_0 ~~
#> Z_0 (CvYZ) 0.299 0.060 4.977 0.000
#> .X ~~
#> .Y (RCYX) 0.011 0.068 0.160 0.873
#>
#> Variances:
#> Estimate Std.Err z-value P(>|z|)
#> X_0 (VarX) 1.000
#> Y_0 (VarY) 1.000
#> Z_0 (VarZ) 1.000
#> .X 0.747 0.081 9.175 0.000
#> .Y 0.659 0.087 7.535 0.000
#> .Z 0.225 1.430 0.157 0.875
#>
#> Constraints:
#> |Slack|
#> ARX - (0.5/VarX) 0.000
#> ARY - ((0.1-0.6*CovXY)/VarY) 0.000
#> ARZ-((0.1-(Effect1*CovXZ+CLYZ*CvYZ))/VrZ) 0.000
#> CovXY - ((0.6*VarX+ARY*CovXY)*ARX+RCovYX) 0.000
#> CvXZ-((Effct1*VrX+CLYZ*CvXY+ARZ*CXZ)*ARX) 0.000
#> CYZ-(0.6*(E1*VX+CLYZ*CXY+ARZ*CXZ)+(E1*CXY 0.000
NoWarnings
A vector with logical information on whether there were any errors or warnings for each of the estimated models.
TRUE means no warnings FALSE means warnings.
modelFit$NoWarnings # Means no warnings for both models
#> [1] TRUE TRUE
CSModelSyntax
The user-specified model syntax (input for model argument)
modelFit$CSModelSyntax
#> [1] "Y ~ .6 * X \n Z ~ Effect1 * X + Y # label the estimated effect of X on Z\n \n X ~~ Y"
SIMSyntax
The syntax for the Stability Informed Model–model syntax for the lavaan function. This contains the syntax to specify the structural part of the Stability Informed Model as well as the parameter constraints for the auto-regressive paths and the latent correlations
modelFit$SIMSyntax
#> [[1]]
#> [1] "X_0=~ARX*X+0.6*Y+Effect1*Z"
#> [2] "Y_0=~0*X+ARY*Y+CLYZ*Z"
#> [3] "Z_0=~0*X+0*Y+ARZ*Z"
#> [4] "X_0~~ CovXY*Y_0"
#> [5] "X_0~~ CovXZ*Z_0"
#> [6] "Y_0~~ CovYZ*Z_0"
#> [7] "X_0~~VarX*X_0"
#> [8] "Y_0~~VarY*Y_0"
#> [9] "Z_0~~VarZ*Z_0"
#> [10] "Y~~RCovYX*X"
#> [11] "ARX==0.5/VarX"
#> [12] "ARY==(0.1-0.6*CovXY)/VarY"
#> [13] "ARZ==(0.1-(Effect1*CovXZ+CLYZ*CovYZ))/VarZ"
#> [14] "CovXY==(0.6*VarX+ARY*CovXY)*ARX+RCovYX"
#> [15] "CovXZ==(Effect1*VarX+CLYZ*CovXY+ARZ*CovXZ)*ARX"
#> [16] "CovYZ==0.6*(Effect1*VarX+CLYZ*CovXY+ARZ*CovXZ)+(Effect1*CovXY+CLYZ*VarY+ARZ*CovYZ)*ARY"
#>
#> [[2]]
#> [1] "X_0=~ARX*X+0.6*Y+Effect1*Z"
#> [2] "Y_0=~0*X+ARY*Y+CLYZ*Z"
#> [3] "Z_0=~0*X+0*Y+ARZ*Z"
#> [4] "X_0~~ CovXY*Y_0"
#> [5] "X_0~~ CovXZ*Z_0"
#> [6] "Y_0~~ CovYZ*Z_0"
#> [7] "X_0~~VarX*X_0"
#> [8] "Y_0~~VarY*Y_0"
#> [9] "Z_0~~VarZ*Z_0"
#> [10] "Y~~RCovYX*X"
#> [11] "ARX==0.55/VarX"
#> [12] "ARY==(0.15-0.6*CovXY)/VarY"
#> [13] "ARZ==(0.2-(Effect1*CovXZ+CLYZ*CovYZ))/VarZ"
#> [14] "CovXY==(0.6*VarX+ARY*CovXY)*ARX+RCovYX"
#> [15] "CovXZ==(Effect1*VarX+CLYZ*CovXY+ARZ*CovXZ)*ARX"
#> [16] "CovYZ==0.6*(Effect1*VarX+CLYZ*CovXY+ARZ*CovXZ)+(Effect1*CovXY+CLYZ*VarY+ARZ*CovYZ)*ARY"
modelImpliedEquations
Model implied equations for the latent covariances and auto-regressive paths
modelFit$modelImpliedEquations
#> [[1]]
#> [1] "ARX==0.5/VarX"
#> [2] "ARY==(0.1-0.6*CovXY)/VarY"
#> [3] "ARZ==(0.1-(Effect1*CovXZ+CLYZ*CovYZ))/VarZ"
#> [4] "CovXY==(0.6*VarX+ARY*CovXY)*ARX+RCovYX"
#> [5] "CovXZ==(Effect1*VarX+CLYZ*CovXY+ARZ*CovXZ)*ARX"
#> [6] "CovYZ==0.6*(Effect1*VarX+CLYZ*CovXY+ARZ*CovXZ)+(Effect1*CovXY+CLYZ*VarY+ARZ*CovYZ)*ARY"
#>
#> [[2]]
#> [1] "ARX==0.55/VarX"
#> [2] "ARY==(0.15-0.6*CovXY)/VarY"
#> [3] "ARZ==(0.2-(Effect1*CovXZ+CLYZ*CovYZ))/VarZ"
#> [4] "CovXY==(0.6*VarX+ARY*CovXY)*ARX+RCovYX"
#> [5] "CovXZ==(Effect1*VarX+CLYZ*CovXY+ARZ*CovXZ)*ARX"
#> [6] "CovYZ==0.6*(Effect1*VarX+CLYZ*CovXY+ARZ*CovXZ)+(Effect1*CovXY+CLYZ*VarY+ARZ*CovYZ)*ARY"