Introduction

The following guide is an appendix to our paper “Measuring Rural Stress: a case study on structural equation modeling in R”. In the next segments you will find a simplified step by step process for building your own structural equation model.

Disclaimer: structural equation modeling is a complex area of statistics. Our paper is not intended to be a complete guide on this topic, rather we hope it can serve as an introduction to the power these tools can afford the rural mental health researcher. Those interested can find further readings in our bibliography.

Import Lavaan

Lavaan is a catch-all package that lets you fit a variety of latent variable models, including confirmatory factor analysis, structural equation modeling and latent growth curve models. This is the only package we’ll call here, although later on we’ll call functions from other packages directly.

library(lavaan)

Import Dataset

We will import the real-world dataset hosted on data.world. You can access the dataset following the link below or directly import it into R using this code snippet. We notice that our stress variables are imported as type dbl. These variables are in a Likert scale, which means they are actually ordered data. In a future step we should factorize these variables.

httr::GET("https://query.data.world/s/fif6xttgj4kvir4ixvnlnxdmlqhiuw?dws=00000", httr::write_disk(tf <- tempfile(fileext = ".xlsx")))
## Response [https://download.data.world/file_download/janky22/measuring-rural-stress-dataset/Mid-Western%20State%205%20County%20Stress%20Data.xlsx?auth=eyJhbGciOiJIUzUxMiJ9.eyJzdWIiOiJwcm9kLXVzZXItY2xpZW50Omphbmt5MjIiLCJpc3MiOiJhZ2VudDpqYW5reTIyOjpjYzEzYjg3MC02MzhhLTRiMjMtYjJiNi0xYzQ3YWZlNGUyYWIiLCJpYXQiOjE3MDEyMTA4MTQsInJvbGUiOlsidXNlciIsInVzZXJfYXBpX2FkbWluIiwidXNlcl9hcGlfZW50ZXJwcmlzZV9hZG1pbiIsInVzZXJfYXBpX3JlYWQiLCJ1c2VyX2FwaV93cml0ZSJdLCJnZW5lcmFsLXB1cnBvc2UiOmZhbHNlLCJ1cmwiOiJhNDI4MmU0ZDdiMzZlMjg0NGJjZjY3Yjk5NTg1ZDQ3ZThiYWMyZTlhIn0.r5WNYktKSxdfQRpD0p8A1Akha7v2oc_GVQK0fARnnfvJgTuCN2GQAqTN8hzJWYetVSkPh2jxr5hRj5l8qsYqpA]
##   Date: 2023-11-29 17:19
##   Status: 200
##   Content-Type: application/vnd.openxmlformats-officedocument.spreadsheetml.sheet
##   Size: 19.3 kB
## <ON DISK>  C:\Users\janky\AppData\Local\Temp\RtmpmAxPIl\file456c722f3424.xlsx
data <- readxl::read_excel(tf)
dplyr::glimpse(data)
## Rows: 251
## Columns: 11
## $ hispanic          <chr> "Yes", "No", "Yes", "No", "No", "No", "No", "No", "N…
## $ stress_finance    <dbl> 1, 1, 3, 4, 3, 3, 4, 1, 1, 3, 3, 3, 3, 3, 1, 1, 5, 1…
## $ stress_employment <dbl> 1, 1, 1, 4, 1, 1, 4, 1, 1, 3, 1, 1, 1, 1, 1, 1, 1, 1…
## $ stress_home       <dbl> 1, 4, 3, 4, 3, 2, 2, 1, 1, 3, 2, 1, 4, 1, 1, 1, 3, 1…
## $ stress            <dbl> 1, 4, 3, 4, 4, 2, 4, 1, 1, 3, 3, 2, 4, 2, 1, 1, 5, 1…
## $ income            <chr> "Lower than last year at this time.", "About the sam…
## $ income_about      <dbl> 0, 1, 0, 0, 1, 0, 1, 1, 0, 0, 1, 0, 1, 1, 1, 1, 0, 0…
## $ income_higher     <dbl> 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1…
## $ income_lower      <dbl> 1, 0, 0, 1, 0, 0, 0, 0, 1, 1, 0, 1, 0, 0, 0, 0, 1, 0…
## $ rural             <dbl> 0, 1, 0, 0, 1, 0, 1, 0, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1…
## $ countycode        <chr> "1", "2", "1", "1", "4", "1", "4", "1", "2", "2", "2…

SEM Assumptions

As with most statistical techniques, for SEM estimates to be accurate and reliable, our data needs to fit several assumptions.

We should check for:

  • Multivariate Normality
    • SEM assumes that the observed variables are multivariate normally distributed.
  • Homoscedasticity
    • The errors of the model should have constant variance.
  • Sample Size
    • SEM generally requires a larger sample size than traditional statistical methods. A common rule of thumb is to have at least 200 cases, although this can vary depending on the complexity of the model and the estimation method.
  • Linearity
    • SEM assumes linear relationships between variables.
  • Multicollinearity
    • Avoid high multicollinearity between predictors.
  • Normality of Residuals
    • The residuals of the model should be normally distributed.
  • Independence of Residuals
    • The residuals should be independent.

For this study case, we have made sure that our data fits all of these assumptions. We can safely continue to our next step.

Factorize Likert-type items

Likert and Liker-type items are ordered items. This needs to be accounted for in our model as estimates can vary greatly between ordered and unordered data. Treating ordered factors as continuous variables can lead to biased estimates. Our code snippet below will factorize our stress variables to account for their ordered nature in our model. For more information on this subject be sure to read:

for (i in 2:5){
data[,i] = factor(dplyr::pull(data[,i]),ordered = TRUE)
}

Build structual and measurement equations

Here we build our structural and measurement models. Our measurement model will relate our stress variables to our latent overall stress variable. Our structural model will define our regression equation with overall stress as our dependent variable and our chosen predictors as our independent variables. Finally, we add a correlated error term since stress and home stress were highly correlated. Without this term we could see negative variance errors in our output.

m <- '
  # latent variable - measurement model
  overall_stress =~ stress + stress_finance + stress_home + stress_employment
  
  # structural path - structural model
  overall_stress ~ rural + hispanic + income_higher + income_lower
  
  #correlated error
  stress~~stress_home
'

Create SEM object

Here we call our lavaan function sem() to create a sem object with our model and the data.

m.fit <- lavaan::sem(m, data = data)

Output summary of SEM

Finally, we can see how well our model did by calling a summary() function on our object. For a full report of measures of fitness be sure to include fit.measures = TRUE. You can also include R-Square measures by adding the argument rsquare = True. We evaluate our model fit using the CFI, TLI, chi-square and RMSEA measures and find that it fits our data well. Therefore, we can reliably draw inferences from our estimates.

If you are interested in further practice, you can also build another model including the county code to further see how stress is distributed between counties. You will need to hard code dummy variables for each county code, setting 1 as your reference level, i.e. county code 1 should be set to 0 for each dummy variable. Remember to evaluate your model fit using the CFI, TLI, chi-square and RMSEA measures.

HINT: The fastDummies package has handy functions for hard coding dummy variables.

lavaan::summary(m.fit, standardized = T, fit.measures = T)
## lavaan 0.6.16 ended normally after 22 iterations
## 
##   Estimator                                       DWLS
##   Optimization method                           NLMINB
##   Number of model parameters                        25
## 
##   Number of observations                           251
## 
## Model Test User Model:
##                                               Standard      Scaled
##   Test Statistic                                10.517      21.397
##   Degrees of freedom                                13          13
##   P-value (Chi-square)                           0.651       0.065
##   Scaling correction factor                                  0.543
##   Shift parameter                                            2.037
##     simple second-order correction                                
## 
## Model Test Baseline Model:
## 
##   Test statistic                              1895.905    1383.589
##   Degrees of freedom                                 6           6
##   P-value                                        0.000       0.000
##   Scaling correction factor                                  1.372
## 
## User Model versus Baseline Model:
## 
##   Comparative Fit Index (CFI)                    1.000       0.994
##   Tucker-Lewis Index (TLI)                       1.001       0.997
##                                                                   
##   Robust Comparative Fit Index (CFI)                            NA
##   Robust Tucker-Lewis Index (TLI)                               NA
## 
## Root Mean Square Error of Approximation:
## 
##   RMSEA                                          0.000       0.051
##   90 Percent confidence interval - lower         0.000       0.000
##   90 Percent confidence interval - upper         0.052       0.088
##   P-value H_0: RMSEA <= 0.050                    0.942       0.441
##   P-value H_0: RMSEA >= 0.080                    0.002       0.106
##                                                                   
##   Robust RMSEA                                                  NA
##   90 Percent confidence interval - lower                        NA
##   90 Percent confidence interval - upper                        NA
##   P-value H_0: Robust RMSEA <= 0.050                            NA
##   P-value H_0: Robust RMSEA >= 0.080                            NA
## 
## Standardized Root Mean Square Residual:
## 
##   SRMR                                           0.020       0.020
## 
## Parameter Estimates:
## 
##   Standard errors                           Robust.sem
##   Information                                 Expected
##   Information saturated (h1) model        Unstructured
## 
## Latent Variables:
##                     Estimate  Std.Err  z-value  P(>|z|)   Std.lv  Std.all
##   overall_stress =~                                                      
##     stress             1.000                               0.962    0.899
##     stress_finance     0.961    0.081   11.805    0.000    0.924    0.868
##     stress_home        0.690    0.049   14.004    0.000    0.663    0.641
##     stress_mplymnt     0.688    0.060   11.496    0.000    0.661    0.640
## 
## Regressions:
##                    Estimate  Std.Err  z-value  P(>|z|)   Std.lv  Std.all
##   overall_stress ~                                                      
##     rural             0.438    0.140    3.136    0.002    0.456    0.218
##     hispanic         -0.471    0.197   -2.386    0.017   -0.490   -0.174
##     income_higher     0.130    0.173    0.753    0.452    0.135    0.055
##     income_lower      0.686    0.156    4.400    0.000    0.714    0.333
## 
## Covariances:
##                    Estimate  Std.Err  z-value  P(>|z|)   Std.lv  Std.all
##  .stress ~~                                                             
##    .stress_home       0.245    0.060    4.077    0.000    0.245    0.658
## 
## Intercepts:
##                    Estimate  Std.Err  z-value  P(>|z|)   Std.lv  Std.all
##    .stress            0.000                               0.000    0.000
##    .stress_finance    0.000                               0.000    0.000
##    .stress_home       0.000                               0.000    0.000
##    .stress_mplymnt    0.000                               0.000    0.000
##    .overall_stress    0.000                               0.000    0.000
## 
## Thresholds:
##                    Estimate  Std.Err  z-value  P(>|z|)   Std.lv  Std.all
##     stress|t1        -0.831    0.276   -3.015    0.003   -0.831   -0.776
##     stress|t2        -0.417    0.273   -1.530    0.126   -0.417   -0.390
##     stress|t3         0.566    0.270    2.094    0.036    0.566    0.529
##     stress|t4         1.452    0.296    4.910    0.000    1.452    1.357
##     stress_fnnc|t1   -0.831    0.257   -3.237    0.001   -0.831   -0.780
##     stress_fnnc|t2   -0.586    0.250   -2.344    0.019   -0.586   -0.550
##     stress_fnnc|t3    0.393    0.244    1.609    0.108    0.393    0.369
##     stress_fnnc|t4    0.957    0.248    3.863    0.000    0.957    0.899
##     stress_home|t1   -0.342    0.251   -1.365    0.172   -0.342   -0.331
##     stress_home|t2    0.028    0.251    0.110    0.912    0.028    0.027
##     stress_home|t3    1.035    0.254    4.071    0.000    1.035    1.001
##     stress_home|t4    1.696    0.298    5.689    0.000    1.696    1.640
##     strss_mplymn|1   -0.298    0.265   -1.127    0.260   -0.298   -0.288
##     strss_mplymn|2    0.098    0.262    0.375    0.708    0.098    0.095
##     strss_mplymn|3    0.910    0.265    3.431    0.001    0.910    0.880
##     strss_mplymn|4    1.379    0.272    5.068    0.000    1.379    1.334
## 
## Variances:
##                    Estimate  Std.Err  z-value  P(>|z|)   Std.lv  Std.all
##    .stress            0.221                               0.221    0.193
##    .stress_finance    0.280                               0.280    0.247
##    .stress_home       0.629                               0.629    0.589
##    .stress_mplymnt    0.631                               0.631    0.591
##    .overall_stress    0.779    0.073   10.720    0.000    0.843    0.843
## 
## Scales y*:
##                    Estimate  Std.Err  z-value  P(>|z|)   Std.lv  Std.all
##     stress            1.000                               1.000    1.000
##     stress_finance    1.000                               1.000    1.000
##     stress_home       1.000                               1.000    1.000
##     stress_mplymnt    1.000                               1.000    1.000