dySEM overview

Why dySEM?

The dySEM Workflow at a Glance

dySEM is designed to be maximally useful if you are following some best-practices for reproducibility when using R. Namely, using a separate directory with an R Studio Project (.Rproj, see here if you are new to using projects) will allow dySEM to be more helpful, by creating sub-folders for your scripts and output where it will automatically save any scripts you create, or any tables and/or figures of output that you create. dySEM will still do these things without the use of an R Studio project, but all bets are off for where R Studio will attempt to save them.

A typical dySEM workflow is then as follows:

  1. Import and wrangle your Data to a dyad structure data set
  2. Scrape variables from your data frame
  3. Script your preferred model
  4. Fit and Inspect your scripted model using lavaan
  5. Output statistical table(s) and or visualization(s)

You might also use optional dySEM calculators after Step 3 to get some additional information.

These families of functions–Scrapers, Scripters, and Outputters, and Getters–are listed and described in the Reference

We now demonstrate a typical dySEM workflow, using the built-in DRES data (Raposo, Impett, & Muise, 2020), in order to perform dyadic confirmatory factor analysis (CFA). More elaborate and specific vignettes are forthcoming to provide didactic materials for conducting other sorts of dyadic data analyses via dySEM.

1. Import and Wrangle Data

We will use a subset of DRES, consisting of 121 dyadic couples’ ratings of relationship quality on 9 of the PRQC indicators (1 = not at all, 7 = extremely; all indicators positively keyed). Structural equation modeling (SEM) programs like lavaan require dyadic data to be in dyad structure dataset, whereby each row contains the data for one dyad, with separate columns for each observation (in this case, indicator variables of latent relationship quality) made for each member of the dyad. We may eventually build in data-transformation functions to go from various data structures to a dyad structure, but for now, we recommend tidyr::pivot_wider or the tools provided by Ledermann & Kenny (2014)

Like many real-world analytic contexts, DRES contains a number of other variables that we aren’t interested in modeling at this time (specifically, 5 indicators of sexual satisfaction from the GMSEX for each dyad member). This will not be a problem for dySEM.

Our data set therefore results in a tibble that is 121 (# of couples) x 28 ((9 PRQC items + 5 GMSEX items) x 2 (# of dyad members)):

DRES
#> # A tibble: 121 × 28
#>    PRQC_1.1 PRQC_2.1 PRQC_3.1 PRQC_4.1 PRQC_5.1 PRQC_6.1 PRQC_7.1 PRQC_8.1
#>       <int>    <int>    <int>    <int>    <int>    <int>    <int>    <int>
#>  1        7        7        7        7        7        7        7        5
#>  2        6        6        6        7        7        6        5        5
#>  3        7        7        7        7        7        7        7        6
#>  4        6        6        6        7        7        6        5        6
#>  5        7        7        7        7        7        6        7        6
#>  6        6        6        6        6        6        3        6        5
#>  7        7        6        7        6        6        6        5        6
#>  8        6        7        7        7        7        6        5        6
#>  9        7        7        7        7        7        6        6        6
#> 10        6        6        6        7        7        7        4        4
#> # ℹ 111 more rows
#> # ℹ 20 more variables: PRQC_9.1 <int>, PRQC_1.2 <int>, PRQC_2.2 <int>,
#> #   PRQC_3.2 <int>, PRQC_4.2 <int>, PRQC_5.2 <int>, PRQC_6.2 <int>,
#> #   PRQC_7.2 <int>, PRQC_8.2 <int>, PRQC_9.2 <int>, sexsat1.1 <int>,
#> #   sexsat2.1 <int>, sexsat3.1 <int>, sexsat4.1 <int>, sexsat5.1 <int>,
#> #   sexsat1.2 <int>, sexsat2.2 <int>, sexsat3.2 <int>, sexsat4.2 <int>,
#> #   sexsat5.2 <int>

2. Scrape

The first step in a typical dySEM workflow is to scrape the indicator variables that are to feature in your latent dyadic model. The scraping functions in dySEM accomplish this by making an important but reasonable (in most cases) assumptions about how the useR has named their indicator variables. Specifically:

Indicator variables of a latent variable will be named in a highly repetitious manner, distinguished by partner using two numbers or characters

Anatomy of a Repetitious Indicator Name

The dySEM scrapers consider appropriately repetitiously named indicators as consisting of at least three distinct elements: stem, item, and partner. For longitudinal designs, a fourth element–time is also considered to be part of the repetitious structure of variable names, but we cover longitudinal variable-scraping in a separate vignette. delimiter characters (e.g., “.”, “_“) are commonly–but not always–used to separate some/all of these elements.

TO DO: MAKE THIS SIMPLER AND START WITH VISUALS AT THIS POINT

  1. The indicator stem (i.e., the character(s) that captures to which scale/latent variable the indicators correspond, e.g., “PRQC”, “sexsat”, “BFI”, etc.). The contents of indicator stems will vary considerably both within and between data sets.
  2. The indicator item number (i.e, the number that captures which indicator–within a set of indicators of some n length–is located in a given column, e.g., 1-9 for our PRQC items)
  3. And the number or character partner capturing to which member of the dyad (the first or second) a given indicator corresponds (e.g., “A” or “B”, “M” or “F”, “1” or “2”). Note: this is only about variable selection; this has no bearing on whether a given dyadic model is specified to be (in)distinguishable (which is determined by the script).

dySEM scrapers largely function by asking you to specify in what order the elements of variable names are ordered. For example: * x_order = "sip" would scrape variable names according to a stem –> item –> partner order (e.g., PRQC)

Using dySEM Scrapers

The scrapeVarCross function is your dySEM scraper for cross-sectional dyadic data. It can accommodate scraping indicators for models featuring one latent variable (e.g., as in our dyadic CFA), as well as bivariate latent variable models, such as the Actor-Partner Interdependence Model (APIM). We cover scraping and scripting of these bivariate models in other vignettes.

We first supply our data frame, DRES. We want to extract our PRQC indicators, which have the following properties:

Feeding this information to scrapeVarCross is quite straightforward:

dvn <- scrapeVarCross(DRES, x_order = "sip", x_stem = "PRQC", x_delim1="_",x_delim2=".",  distinguish_1="1", distinguish_2="2")

Before looking at what scrapeVarCross returns, you may be wondering:

  1. Where is information about Item number for each indicator specified? And…
  2. What should you do if your indicator names don’t use two (or any) delimiting characters?

The answer to 1. is that Item number is automatically captured “behind the scenes” by scrapeVarCross. Specifically, scrapeVarCross searches for (and then captures) any variable names containing your stem and any digit(s) (using a regular expression).

The answer to 2. is that you would simply omit the x_delim1 and/or x_delim2 arguments–by default, scrapeVarCross will create variable names from S I and P without any separating delimiters, unless you declare a character in one/both delimiter arguments.

scrapeVarCross returns a generic list (which I refer to as a dvn for a list of “dyad variable names”) consisting of 6 (or 9, if scraping for a bivariate model) elements:

  1. a vector of indicator names for the first member of the dyad
  2. a vector of indicator names for the second member of the dyad
  3. a number capturing how many indicators per dyad member were stored
  4. the distinguishing character in names for indicators from the first member of the dyad
  5. the distinguishing character in names for indicators from the second member of the dyad
  6. the total number of indicators scraped

This might not seem like much, but the list returned by scrapeVarCross contains all the information needed to automate the scripting of lavaan syntax for virtually any dyadic SEM that you can imagine.

3. Script

The script...() TODO: CREATE/LINK to family in Reference on pkgdown site: family of functions in dySEM simplify the process of accurately and reproducibly scripting dyadic SEMs to a singleton line of R code.

Each Scripter function is a wrapper for a series of Helper functions (see scriptHelpers.R if you are interested) that snatch the information about the indicators they need from a saved dvn object and combine it with other text to write the lavaan syntax for a particular part of the measurement (e.g., factor loadings, item intercepts) or structural (e.g., regression slopes, factor means) portion of your model.

Scripter functions like scriptCFA typically require only three arguments to be specified:

  1. the dvn object (e.g., from scrapeVarCross) to be used to script the model
  2. a mostly arbitrary name for the latent variable(s) you are modeling (bivariate model scripting functions like scriptAPIM have you input two names)
  3. the kind of parameter equality constraints that you wish to be imposed (if any), such as those corresponding to particular levels of measurement invariance (e.g., “loading”), or even a fully “indistinguishable” model (i.e., in which all measurement and structural parameters are constrained to equality between partners)

If you plan on scripting multiple models, I recommend that you name the output of Scripters to include information about the latent variable’s name (from 2.) and model (from 3.). For example, if we were to use scriptCFA to generate scripts for an indistinguishable CFA (i.e., both imposing dyadic invariance and equality of latent variances and means between partners [the default options for the constr_dy_meas and constr_dy_struct arguments]) of the PRQC items we scraped, we could specify


qual.indist.script <- scriptCFA(dvn, lvname = "Quality")

scriptCFA returns to your environment an (ugly, to the human-eyes) character object consisting of the lavaan syntax corresponding to the model that matches the Scripter function (i.e., in this case a CFA) and input for the model argument (i.e., configurally invariant).

Meanwhile, behind the scenes, scriptCFA has created a folder in your working directory called “scripts”, and has stored a .txt file containing the (less-ugly, to the human eyes) lavaan syntax for your model. The file will be named as a combination of your lvname and model arguments.

We think this syntax-exporting-.txt feature serves three important purposes:

  1. It makes it easy to immediately share your analytic code (i.e., just drop the .txt file in your OSF project)
  2. It can be useful as an exemplar from which to learn how certain model features of dyadic SEMs are scripted in lavaan. For example, our Scripters manually specify (and label) the estimation of certain parameters that would already be estimated by default, but this way you can learn how differences in model specification (e.g., changing to a different level of invariance) impacts the lavaan syntax. And…
  3. Should you require a model that is more customizable (e.g., a particular pattern of partial dyadic invariance), the text in the .txt file can serve as useful starting point for scripting your model that should (hopefully) only require but a few “handmade” changes (e.g., keeping most item intercepts equated via model = "intercepts, and manually freeing only those that are appreciably different).

The modeling efficiency and accuracy gained via dySEM’s automated scripting may already be apparent, but becomes painfully obvious once you leverage dySEM to quickly script a sequence of competing models (e.g., from configural invariance CFA –> fully indistinguishable CFA)


qual.res.script <- scriptCFA(dvn, lvname = "Quality", constr_dy_meas = c("loadings", "intercepts", "residuals"), constr_dy_struct = c("none"))

qual.int.script <- scriptCFA(dvn, lvname = "Quality", constr_dy_meas = c("loadings", "intercepts"), constr_dy_struct = c("none"))

qual.load.script <- scriptCFA(dvn, lvname = "Quality", constr_dy_meas = c("loadings"), constr_dy_struct = c("none"))

qual.config.script <- scriptCFA(dvn, lvname = "Quality", constr_dy_meas = c("none"), constr_dy_struct = c("none"))

The scripting of longitudinal dyadic SEM models is not yet supported by dySEM, but we hope to develop this functionality over the Spring/Summer 2021.

4. Fit and Inspect

By design, we have attempted to avoid functionality pertaining to model-fitting and inspection in dySEM: lavaan does that perfectly well itself. We therefore strongly recommend that you cultivate a command of lavaan’s basic functionality before delving too far with dySEM–the package tutorial website is a very good place to get started.

You can immediately pass any script(s) returned from a dySEM scripter (e.g., scriptCFA) to your intended lavaan wrapper (we recommend cfa–be sure to disable any options that might fix parameters, as the scripter has already taken care of manually specifying which parameters to fix or estimate), with your preferred estimator and missing data treatment. For example, with dyadic invariance testing, we recommend starting with the most parsimonious model (an indistinguishable model), and gradually relaxing constraints on different groups of parameters:


#Fit fully indistinguishable model
qual.ind.fit <- lavaan::cfa(qual.indist.script, data = DRES, std.lv = FALSE, auto.fix.first= FALSE, meanstructure = TRUE)

#Fit residual invariance model
qual.res.fit <- lavaan::cfa(qual.res.script, data = DRES, std.lv = FALSE, auto.fix.first= FALSE, meanstructure = TRUE)

#Fit intercept invariance model
qual.int.fit <- lavaan::cfa(qual.int.script, data = DRES, std.lv = FALSE, auto.fix.first= FALSE, meanstructure = TRUE)

#Fit loading invariance model
qual.load.fit <- lavaan::cfa(qual.load.script, data = DRES, std.lv = FALSE, auto.fix.first= FALSE, meanstructure = TRUE)

#Fit configural invariance model
qual.config.fit <- lavaan::cfa(qual.config.script, data = DRES, std.lv = FALSE, auto.fix.first= FALSE, meanstructure = TRUE)

At this point, the full arsenal of lavaan model-inspecting tools are at your disposal. Two that you will almost certainly want to make use of are summary and anova.

summary will useful for printing model fit information as well as parameter estimates and tests to your console. For example:

summary(qual.config.fit, fit.measures = TRUE, standardized = TRUE, rsquare = TRUE)

anova, meanwhile, will enable you to perform comparisons of competing nested dyadic models. For example:

anova(qual.config.fit, qual.load.fit, qual.int.fit, qual.res.fit, qual.ind.fit)
#> 
#> Chi-Squared Difference Test
#> 
#>                  Df    AIC    BIC  Chisq Chisq diff    RMSEA Df diff Pr(>Chisq)
#> qual.config.fit 125 5102.8 5280.6 573.33                                       
#> qual.load.fit   133 5096.5 5252.2 583.09     9.7565 0.042954       8     0.2825
#> qual.int.fit    141 5089.1 5222.5 591.66     8.5669 0.024403       8     0.3801
#> qual.res.fit    150 5083.5 5191.9 604.04    12.3872 0.056237       9     0.1924
#> qual.ind.fit    152 5081.4 5184.2 605.97     1.9264 0.000000       2     0.3817

You can learn about what other kinds of detail you can extract from a fitted lavaan model here.

5. Output

dySEM also contains functionality to help you quickly, correctly, and reproducibly generate output from your fitted model(s), in the forms of path diagrams and/or tables of statistical values. Path diagram creation is supported via the semPlot package’s semPaths function, whereas table creation is currently supported by the sjPlot package’s tab_df function, though users should be aware that we are considering a move to the gt package for improved tabling capacity.

The outputModel function is currently dySEM’s all-purpose outputting function.

The useR must specify the dvn of scraped variables used to script the model and the type of model being outputted (e.g., “cfa”). UseRs can specify whether they only want a path diagram or some table(s) to be outputted (e.g., by setting either figure = FALSE or table = FALSE), but by default both are created. UseRs can specify a directory path to where they want their file(s) to be written and saved (e.g., setting writeTo = "." to save in the current working directory). UseRs can further specify what kind of path diagram (e.g., using standardized or unstandardized value) or tables (e.g., featuring measurement- or structural-model parameter, or both) are created.

Whatever options the useR specifies, outputModel is typically run without assigning its output to an object, as it’s mostly to facilitate statistical reporting for scientific articles:

outputModel(dvn, model = "cfa", fit = qual.config.fit, 
            table = TRUE, tabletype = "measurement", 
            figure = TRUE, figtype = "unstandardized",
            writeTo = tempdir(), fileName = "dCFA_config")