Structural equation modeling (SEM) lets researchers examine how multiple variables influence other variables. This might sound familiar given your introduction to statistics knowledge. In fact, SEM borrows heavily from regression. However, while regression allows a researcher to test if one continuous independent variable (IV) predicts another continuous dependent variable (DV) via a linear relationship, SEM allows for complicated networks with multiple IVs influencing multiple DVs.
A bit of a visual representation helps here.
Here is a visal representation of an SEM model, call a Path Diagram
However, at the core of SEM is the idea that statisticians are often interested in understanding concepts that can not be perfectly and directly measured.
Two important new vocabulary terms to introduce before getting any further into this explanation of SEM are latent and observed variables. Latent variables the abstract ideas researchers want to study, and observed variables are the concrete pieces of information that a research collects to define and build out the observed variables.
For example, a psychologist may be interested in intelligence, which cannot be measured directly. Scores on an IQ test, standardized tests, or school performance could all serve as observed variables that tap into and estimate the latent variable.
As a consumer of research, be aware that the researcher will evaluate the overall model, and then discuss the relationship between just two variables.
Industrial/Organizational Psychology: How does employee loyalty affect an employee's intention to stay at a job?
Latent variables: Loyalty (IV) and intention to stay (DV)
Observed variables that could be used to measure loyalty:
Participant responses to survey questions about loyalty.
Employee's length of service with the company.
Participants' responses to a survey about their attitude and feelings of loyalty towards their direct supervisor
Participants' responses to a survey about their attitude and feelings of loyalty towards their co-workers.
Observed variables that correspond to intention to stay:
Participants' resposnes to survey questions on whether or not they see themselves in their current poisiton a year from now.
Ask your participants for the number of other jobs they have applied for in the last year.
Ask participants to list the number of job search-related behaviors they have engaged in in the last year, including but not limited to a) checking job postings with other employers, b) discussing leaving the current job with friends and family, c) revising their resume.
Developmental Psychology: How do parents' attitudes about reading affect a child's attitude toward reading?
Latent variables: Parents' attitudes (IV) and the child's attitude (DV).
Observed variables that could measure parents' attitudes towards reading:
Parent responses to survey questions about reading
The number of books the parent read in the past year.
The number of adult books in the home.
Observed variables that correspond to the child's attitude toward reading:
Contact the child's school for their score on reading aptitude tests.
Child score on a reading-attitudes survey
Ask the child's teacher for the number of times the child voluntarily reads during free time at school.
Clincal Psychology: How does family support affect military members' mental health during deployment?
Latent Variables: Family support (IV) and military members' mental health during deployment (DV).
Observed variables that could measure family support:
Family members' responses to a support survey,
Family members' participation in available support groups,
The stability of their local support network.
Observed variables for mental health:
Scores on depression and anxiety inventories
Commanding officer's assessment of the service member's apparent well-being.
Schmidt and colleagues (2010) studied how 200 Italian high school students appraised and coped with an important final exam that they needed to pass for their diploma.
The researchers proposed a model connecting students' appraisals of the exam, their emotions surrounding the exam, and the coping strategies they used. In other words, they asked whether students' thoughts and feelings about an important exam related to how they coped with it. To test this, they had to give the participants a bunch of questionnaires to ask them about their appraisals, emotions, and coping strategies.
Thedy then fed participant responses into a Confirmatory Factor Analysis, which is the first step of a Structural Equation Model. You can read more about Factor Analysis here.
Several parts of the model were latent variables. These included three types of emotional reactions to exam stress: Frustration and Powerlessness, Anxiety and Fear, and Positive Emotions. The researchers also examined several types of coping, including suppression, reappraisal, social support, distancing, problem-focused coping, and drug use.
Note that the paper provides extensive data to support proper interpretation of the SEM. This often includes a) conducting correlation analyses among all of the latent variables, and b) a factor analysis (which is also explained by this tool) is sometimes used to create latent variables based upon the scales used in an experiment.
For example, based on a factor analysis, the authors created latent variables. During the experiment, participants filled out multiple questionnaires, answering questions about coping strategies, emotions they experience, and their personal appraisal of the importance of the upcoming high school exit exam. Using these data, the researchers conducted a factor analysis, drawing items from the surveys to create latent variables titled frustration/powerlessness, anxiety/fear, and positive emotion.
When an author describes their final SEM, they will generally present their findings in this format. NOTE: The chi-square is in there. Essentially, all of these measures I'm about to describe look at the difference between the data you have versus the way the data would look if your hypothesis is true. So, not unlike chi-square assessment of the difference between observed and expected values. The numbers will vary:
There are two steps to interpreting SEM:
How well does the overall hypothesized model fit the actual data?
How well does any one variable predict any other variable within the whole model?
To read a research paper, it is probably more important that you understand step 2, but this guide will still explain the first step of evaluating the overal model.
Step 1: Reading about how well the overall model fits the actual data.
Below is an excerpt from the data we've been discussing. This data, presented this way, is commonplace in an SEM article.
"This global model had the following fit indexes: χ 2(1342, N = 610) = 3880.71, p < 0.001; SRMR = 0.078; RMSEA = 0.056; CFI = 0.90."
Short explanation:
All of these numbers test how well your proposed model fits the data that was gathered and analyzed. Hopefully, you remember the idea of "fit" from learning about regression. Again, the backbone of SEM is regression, which also tests how well a regression line fits the actual data. While linear regression tests how well one variable predicts another variable, these numbers indicate, simultaneously, how well all of the variables predict all of the other variables. Typically, you report this statistic, followed by a sentence stating whether the data is a good fit.
Long explanation:
If you would like to learn a bit more about the way fit is determined, here is a long explanation:
χ 2(1342, N = 610) = 3880.71: This is describing a chi-square. 1342 is the degree of freedom, followed by the sample size of 610. The test statisitcs was 3880.71, and it was statistically significant since the p value is less than .001
SRMR= 0.078: Standardized Root Mean Square Residual
RMSEA = 0.056: The Root Mean Square Error of Approximation is the absolute measure of how well a proposed model fits the actual data. A small value indicates a good fit. Specifically, values less than or equal to .05 are considered a good fit, .05-.08 a reasonable fit, .08-.10 a mediocre fit, and >1.0 a poor fit.
CFI = 0.90: The Comparative Fit Index measures the goodness of fit between the actual data and the proposed model. It can range from 0-1, with 1 being a perfect fit. Typically, anything over .90 is considered a good fit. This is 0.90, so it is a good fit.
However, the more pertinent interpretation is almost always listed in a figure. The figure will likely be accompanied by text in which the researchers will draw attention to relationships in the figure that support or fail to support the research hypotheses.
Next, we will discuss the second step of interpretation, in which we understand the relationships between specific variables. And that interpretation in very visual.
How do you interpret this table? Let's start with one of the latent dependent variables, Positive Emotions (highlighted in red). The standardized path coefficient from Positive Emotions to Reappraisal (an observed variables that corresponds to positive emotion, and highlighted in purple) was .73 (in yellow).
The important thing to notice is that these coefficients tell us about the direction and strength of relationships in the model. The .73 represents a strong positive relationship between Positive Emotions and Reappraisal. In everyday terms, students who reported more positive emotions also tended to use reappraisal more often as a coping strategy.
This makes sense given how the study defined reappraisal. It included efforts to learn from the experience and see its positive aspects. When you see .73 beside an arrow in the SEM diagram, you can translate it into a statement about how strongly those two variables were related in the researchers' model.
You can interpret the other beta this way, noting that a negative beta (like between Anxiety/Fear and Distancing can be negative, indicating that as one value increases, the other value decreases.
Schmidt, S., Tinti, C., Levine, L. J., & Testa, S. (2010). Appraisals, emotions and emotion regulation: An integrative approach. Motivation and Emotion, 34(1), 63–72. https://doi.org/10.1007/s11031-010-9155-z