Path diagrams provide visually appealing representations of the interrelationships among variables in structural equation models. However, it is no easy task to determine the “best” layout of the variables in a path diagram. The CALIS procedure provide three algorithms for laying out goodquality path diagrams. Understanding the underlying principles of these algorithms helps you select the right algorithm and finetune the features of a path diagram.
The section The ProcessFlow, GroupedFlow, and GRIP Layout Algorithms describes the principles of the algorithms. The sections Handling Destroyers in Path Diagrams and Postediting of Path Diagrams with the ODS Graphics Editor show how you can improve the quality of path diagrams in advance or retroactively. The section Default Path Diagram Settings details the default settings for path diagram elements and the corresponding options. By using these options, you can customize path diagrams to meet your own requirements. Finally, the sections Expanding the Definition of the Structural Variables and Showing or Emphasizing the Structural Components show examples of customizations that you can make to the structural components of models.
The most important factor in choosing an appropriate algorithm is the nature of the interrelationships among variables in the model. Depending on the pattern of interrelationships, the CALIS procedure provides three basic algorithms: the processflow algorithm, groupedflow algorithm, and GRIP algorithm. You can request these algorithms by specifying the ARRANGE= option in the PATHDIAGRAM statement. This section describes these algorithms and the situations in which they work well. The section also describes how the CALIS procedure selects the “best” algorithm by default (by specifying the ARRANGE=AUTOMATIC option).
The processflow algorithm (which you select by specifying ARRANGE=FLOW) works well when the interrelationships among the variables in the model (not including the error variables) follow an ideal processflow pattern. In such a pattern, each variable can be placed at a unique level so that all paths between variables exist only between adjacent levels and are aligned in the same direction. Figure 29.5 shows an example of a linear ordering of observed variables based on the directions of the paths. This model clearly exhibits an ideal processflow pattern: variables can be uniquely assigned to five levels, and the paths can occur only between adjacent levels and in the same direction.
Figure 29.6 shows a hierarchical ordering of the observed variables. The pattern of interrelationships of variables in this model is much like an organizational chart. Clearly, this path diagram also depicts an ideal processflow pattern.
Latent variables are often used in structural equation modeling. Some classes of latent variable models have ideal processflow patterns. For example, for confirmatory factor models in their “pure” form, the processflow algorithm can be used to show the hierarchical relationships of factors and variables. Figure 29.7 shows an example of a confirmatory factor model.
Other examples are higherorder factor models and hierarchical factor models, as shown in Figure 29.8 and Figure 29.9, respectively.
The groupedflow algorithm works well when the interrelationships of the latent factors follow an ideal processflow pattern. Such a pattern requires that each latent factor in the model be placed at a unique level so that all paths between the latent factors exist only between adjacent levels and in the same direction. This pattern is called an ideal groupedflow pattern because each latent factor in the model is associated with a cluster of observed variables. In other words, each cluster of observed variables and a latent factor forms a group. You can specify the ARRANGE=GROUPEDFLOW option in the PATHDIAGRAM statement to request the groupedflow algorithm.
Figure 29.10 illustrates a linear ordering of latent factors, and Figure 29.11 illustrates a hierarchical ordering of latent factors. Both path diagrams illustrate ideal groupedflow patterns.
A model that exhibits an ideal groupedflow pattern for factors can also exhibit an ideal processflow pattern for nonerror variables. The question is then which algorithm is better and under what situations. For example, you can characterize the interrelationships among the variables that are shown in Figure 29.10 by an ideal processflow pattern of all nonerror variables in the model. Figure 29.12 is the diagram that results if you use the ARRANGE=FLOW option for this model.
Unfortunately, the latent variables and observed variables in Figure 29.12 are now mixed within the two middle levels, burying the important clusters of observed variables and latent factors that are clearly shown in the groupedflow representation in Figure 29.10. Hence, the groupedflow algorithm is more preferable here. For this reason, PROC CALIS adds an extra condition for applying the processflow algorithm when the default automatic layout method is specified. That is, in order for the processflow algorithm to be used, variables within each level must also be of the same type (observed or latent), and an ideal processflow pattern is required in the nonerror variables.
The GRIP (Graph dRrawing with Intelligent Placement) algorithm is a general layout method for graphs that display nodes and links (or variables and paths in the context of path diagrams). Unlike the processflow and groupedflow algorithms, the GRIP algorithm is not designed to display specific patterns of variable relationships. Rather, the GRIP algorithm uses a graphtheoretic approach to determine an intelligent initial placement of the nodes (variables). Then the algorithm goes through refinement stages in rendering the final path diagram. Essentially, the GRIP algorithm provides a general layout algorithm when variables or factors in the model cannot be ordered uniquely according to the path directions. The algorithm balances the placement of the variables and groups related variables together.
PROC CALIS uses a modified version of the GRIP algorithm that is more suitable for structural equation models. Basically, error variables are ignored in the layout process (but are displayed in path diagrams if requested), and the lengths of the paths are adjusted according to the types of variables that are being connected by the paths. These modifications ensure that clusters of observed variables and factors are distinct from each other, as illustrated in Figure 29.13.
In Figure 29.13, Factor1
is the only exogenous variable in the model. This means that no directional path points to Factor1
, which is considered to be a level1 variable. At the next level are the variables that have directional paths from Factor1
. Therefore, y1
, y2
, y3
, Factor2
, and Factor3
are considered to be at level 2. However, because Factor2
has a directional path to Factor3
, Factor3
can also be considered to be at level 3. Hence, this model does not show either an ideal processflow pattern or an ideal
groupedflow pattern. However, the more general GRIP algorithm can still show three recognizable clusters of variables.
By default, the CALIS procedure uses an automatic method to determine the layout algorithm for path diagrams. This automatic method is equivalent to specifying the ARRANGE=AUTOMATIC option. Actually, the AUTOMATIC option is not associated with a specific layout algorithm. What it does is determine automatically which algorithm—the processflow, groupedflow, or GRIP algorithm—is the best for a given model.
The following steps describe how PROC CALIS automatically determines the best layout algorithm:
PROC CALIS checks whether the interrelationships among all nonerror variables in the model exhibit an ideal processflow pattern. In this ideal pattern, all nonerror variables in the model can be uniquely ordered according to the directional paths. PROC CALIS then checks whether variables at each level are of the same type (observed variables or latent factors). If they are, PROC CALIS uses the processflow algorithm as if the ARRANGE=FLOW option were specified. Otherwise, it continues to the next step.
PROC CALIS checks whether the interrelationships among all factors in the model exhibit an ideal processflow pattern. In this ideal pattern, all latent factors in the model can be uniquely ordered according to the directional paths. If they are, PROC CALIS uses the groupedflow algorithm as if the ARRANGE=GROUPEDFLOW option were specified. Otherwise, it continues to the next step.
PROC CALIS uses the GRIP algorithm as if the ARRANGE=GRIP option were specified.
In summary, when the default ARRANGE=AUTOMATIC option is used, PROC CALIS chooses a suitable layout algorithm by examining the interrelationships of variables. The procedure detects ideal processflow and groupedflow patterns when they exist in models.
No layout algorithm for path diagrams works perfectly in all situations. Chances are that some paths in your model would violate the ideal pattern that each of the basic layout algorithms assumes. In some cases, even if the violations are minor—for example, they occur in just one or two paths—they are sufficient to throw off the layout of the diagram.
There are two ways for you to alleviate the problems that are caused by these minor violations. One way is to proactively identify the paths in your model that violate the ideal pattern that the intended layout algorithm assumes. If these minor violations are ignored during the layout process, then the layout algorithm can work well for the majority of patterns in the path diagram. To accomplish this, the CALIS procedure provides the DESTROYER= option for you to specify these minor violations. This section presents examples to illustrate the use of this option.
The other way to alleviate these problems is to use the ODS Graphics Editor to improve the graphical quality of the path diagram that PROC CALIS produces. The ODS Graphics Editor provides several functionalities to improve your path diagrams manually. The section Postediting of Path Diagrams with the ODS Graphics Editor describes these functionalities.
To illustrate the use of the DESTROYER= option, consider a higherorder factor model that is specified by the following statements:
proc calis; path FactorG ===> Factor1Factor3, Factor1 ===> y1y3 , Factor2 ===> y4y6 , Factor3 ===> y7y9 , FactorG ===> y2 y7 ; run;
In this model, Factor1
, Factor2
, Factor3
, and FactorG
are all latent factors. All other variables are observed variables. Which would be the best layout algorithm to draw the
path diagram for this model?
Does this model exhibit an ideal processflow pattern? In this model, FactorG
is the only exogenous nonerror variable and is considered to be at level 1. Factor1
, Factor2
, Factor3
, y2
, and y7
all have directional paths from FactorG
, and therefore they are considered to be at level 2. However, because y2
and y7
also have directional paths from Factor1
and Factor3
, respectively, they can also be considered to be at level 3. Hence, this model does not exhibit an ideal processflow pattern.
Does this model exhibit an ideal groupedflow pattern? If you consider only the factors in the model, FactorG
is at level 1 and all other latent factors are at level 2 without ambiguities. Hence, this model exhibits an ideal processflow
pattern in factors or an ideal groupedflow pattern in all nonerror variables. You expect the groupedflow algorithm to be
used to draw the path diagram for this model if you specify the following statements:
proc calis; path FactorG ===> Factor1Factor3, Factor1 ===> y1y3 , Factor2 ===> y4y6 , Factor3 ===> y7y9 , FactorG ===> y2 y7 = destroyer1 destroyer2; pathdiagram diagram=initial notitle; run;
Figure 29.14 shows the output diagram. Although the path diagram still shows the hierarchical ordering of the latent factors, the two
paths, “FactorG ===> y2” and “FactorG ===> y7,” weaken the display of the factorvariable clusters for Factor1
and Factor3
. Therefore, these two paths can be called destroyer paths, or simply destroyers. For this reason, the effect parameters of
these two paths are labeled as “destroyer1” and “destroyer2” in the PATH statement to show their disruptive characteristics.
What if you ignore these destroyer paths in the layout process of the diagram? To do this, you can use the DESTROYER= option in the PATHDIAGRAM statement, as follows:
pathdiagram diagram=initial notitle destroyer=[FactorG ===> y2 y7];
Because the two destroyer paths are ignored when PROC CALIS determines the layout algorithm, an ideal processflow pattern is recognized, and the processflow algorithm is used to lay out all variables in the model. Then, all paths, including the destroyer paths, are simply put back into the picture to produce the final path diagram, as shown in Figure 29.15.
The diagram in Figure 29.15 is preferable to the one in Figure 29.14 because it clearly shows the three levels of variables in the higherorder factor model.
Destroyer paths do not necessarily occur only in ideal processflow or groupedflow patterns. Even when the GRIP algorithm is used for a general pattern of variable interrelationships, you might be able to identify potential destroyers in the path diagram. Consider the model that is specified by the following statements:
proc calis; path Factor1 ===> y1 y2, Factor2 ===> y3 y4, Factor3 ===> y5 y6, Factor1 ===> Factor2 Factor3, Factor2 ===> Factor3, y2 ===> y5 = destroyer; pathdiagram diagram=initial notitle; run;
This model exhibits neither an ideal processflow pattern nor an ideal groupedflow pattern. As a result, PROC CALIS uses the GRIP algorithm automatically to draw the path diagram. Figure 29.16 shows the output diagram.
Because of the presence of the destroyer path “y2 ===> y5,” the two factor clusters for Factor1
and Factor3
are drawn closer to each other in the path diagram, thus destroying their distinctive identities that are associated with
the related measured variables.
You can use the DESTROYER= option to fix this problem, as shown in the following modified PATHDIAGRAM statement:
pathdiagram diagram=initial notitle destroyer=[y2 ===> y5];
Figure 29.17 now shows a more preferable path diagram. The factor clusters in this diagram are more distinctive than those in Figure 29.16, in which the algorithm does not handle the destroyer path.
You can edit the output path diagram in PROC CALIS by using the ODS Graphics Editor. The ODS Graphics Editor provides the following useful functionalities for postediting path diagrams:
moves the variables around without losing the associated paths
aligns the selected variables vertically or horizontally
straightens or curves paths
reshapes or reroutes paths
relocates the labels of paths
To use the ODS Graphics Editor, you must first enable the creation of editable graphs. For information about the enabling procedures, see the section Enabling and Disabling ODS Graphics in Chapter 21: Statistical Graphics Using ODS. For complete documentation about the ODS Graphics Editor, see the SAS 9.4 ODS Graphics Editor: User's Guide. For information about editing nodes and links in output path diagrams, see Chapter 8 of the SAS 9.4 ODS Graphics Editor: User's Guide. Note that different terminology is used in the SAS 9.4 ODS Graphics Editor: User's Guide. Variables and paths in the current context are referred to as nodes and links in the User’s Guide.
Taking the conventions and clarity of graphical representations into consideration, the CALIS procedure employs a set of default settings for producing path diagrams. These default settings define the style or scheme for representing the path diagram elements graphically. For example, PROC CALIS displays the error variances as doubleheaded paths that are attached to the associated variables. In this way, there is no need to display the error variables so that more space is available to present other important graphical elements in path diagrams.
However, researchers do not always agree on the best style or scheme for representing the path diagram elements. Even the same researcher might want to use different representation style or scheme for his or her path diagrams in different situations. In this regard, you can override most of the default settings to customize path diagrams to meet your own requirements. Table 29.6 summarizes the default graphical and nongraphical settings of the path diagram elements. Related options are listed next to these default settings.
Table 29.6: Default Path Diagram Settings and the Overriding or Modifying Options
Graphical Element or Property 
Default Settings 
Options 

General Graphical Properties 

Layout method 
Automatically determined 

Models with path diagram output 
All models 

Path diagrams for structural models 
Not shown 

Paths that affect the layout 
All paths 

Solution type of path diagrams 
Unstandardized solution 

Structural components 
Not emphasized 

Paths 

Covariances between error variables 
Shown as doubleheaded paths 

Covariances between nonerror and error variables 
Shown as doubleheaded paths 

Covariances between nonerror variables 
Shown as doubleheaded paths in models specified using the FACTOR statement; not shown in other types of models 

Directional paths 
Shown as singleheaded paths 

Error variances 
Shown as doubleheaded paths in covariance models; shown as labels in mean and covariance models 

Means and intercepts 
Not shown as paths; shown as labels in mean and covariance models 

Variances of nonerror variables 
Shown as doubleheaded paths in covariance models; shown as labels in mean and covariance models 

Variables 

Error variables 
Not shown 

Relative sizes 
Observed variables:Factors:Errors = 1:1.5:0.5 

Variable labels 
Original variable names used 

Parameter Estimates 

Covariances between error variables 
Shown in all types of models 

Covariances between nonerror variables 
Shown in models specified using the FACTOR statement; not shown in other types of models 

Initial parameter specifications 
Shown with fixed values and userspecified parameter names 

Means and intercepts 
Shown in models with mean structures 

Variances of error variables 
Shown in all types of models 

Variances of nonerror variables 
Shown in all types of models 

Formats of Parameter Estimates 

Decimal places 
Two decimal places used 

Numerical values 
Shown in unstandardized and standardized solutions 

Parameter names 
Not shown 

Significance flags 
Shown in unstandardized and standardized solutions 

Graph Title and Fit Summary 

Decimal places in fit summary 
Two decimal places used 

Fit table 
Shown in unstandardized and standardized solutions 

Fit table contents 
A default set of available information and statistics 

Title 
Default title that reflects the identity and solution type of the model 
One oftenused customization in structural equation modeling is to display only the structural component (or structural model) of the full model, which consists of the measurement and the structural components. Traditionally, the structural component refers to the latent factors and their interrelationships, whereas the measurement component refers to observed indicator variables and their relationships with the latent factors.
For example, consider the full model that is displayed as a path diagram in Figure 29.18. The latent factors are Factor1
, Factor2
, and Factor3
. The remaining variables, y1
–y9
, are observed variables.
The structural component of this full model includes the three latent factors, their directional relationships (represented by the three directed paths), and their variance parameters (represented by the three doubleheaded arrows that are attached to the factors).
Showing only the structural component is useful when your research is focused mainly on the interrelationships among the latent factors. The measurement component, which usually contains a relatively large number of observed variables for reflecting or defining the latent factors, is of secondary interest. Eliminating the measurement component leads to a clearer path diagram to represent your model.
To request a path diagram that shows the structural component, you can specify the STRUCTURAL option in the PATHDIAGRAM statement, as in the following statements:
proc calis; path Factor1 ===> y1y3 , Factor2 ===> y4y6 , Factor3 ===> y7y9 , Factor1 ===> Factor2 Factor3, Factor2 ===> Factor3; pathdiagram diagram=initial structural; run;
Figure 29.19 shows the path diagram for the structural model. For simplicity in presentation, the DIAGRAM=INITIAL option is used to show the initial model specifications only.
The preceding specification produces the path diagram for the structural component, in addition to that for the full model. To output the path diagram for the structural component only, use the STRUCTURAL(ONLY) option in the PATHDIAGRAM statement, as follows:
pathdiagram diagram=initial structural(only);
Instead of eliminating the measurement component from the path diagram, PROC CALIS provides the EMPHSTRUCT option to highlight the structural component in the full model. For example, the following PATHDIAGRAM statement produces the path diagram in Figure 29.19:
pathdiagram diagram=initial emphstruct;
Figure 29.20 shows a complete model, with emphasis on the structural component. The path diagram labels only the latent factors. Observed variables are represented by small rectangles and are not labeled.
Certainly, the EMPHSTRUCT option is not limited to the initial path diagram, such as the one shown in Figure 29.20. For the unstandardized and standardized solutions, the EMPHSTRUCT option puts similar emphasis on the structural variables. In addition, the diagram displays only the parameter estimates of the structural components. Figure 29.21 shows the path diagram of an unstandardized solution that emphasizes the structural components.
Figure 29.21: Emphasizing the Structural Component in the Path Diagram for the Unstandardized Solution
The traditional definition of “structural model” or “structural component” might be too restrictive, in the sense that only latent factors and their interrelationships are included. In contemporary structural equation modeling, sometimes observed variables can take the role of “factors” in the structural component. For example, observed variables that have negligible measurement errors and serve as exogenous or predictor variables in a system of functional equations naturally belong to the structural component. For this reason or other reasons, researchers might want to expand the definition of “structural component” to include some designated observed variables. The STRUCTADD= option enables you to do that.
For example, the following statements specify a model that has two latent factors, Factor1
and Factor2
:
proc calis; path Factor1 ===> y1y3, Factor2 ===> y4y6, y7 ===> Factor1 Factor2; pathdiagram diagram=initial notitle structadd=[y7] label=[y7="y7 is Exogenous"] ; pathdiagram diagram=initial notitle structadd=[y7] struct(only); label=[y7="y7 is Exogenous"] ; pathdiagram diagram=initial notitle structadd=[y7] emphstruct; label=[y7="y7 is Exogenous"] ; run;
By default, only the latent factors Factor1
and Factor2
are treated as structural variables. However, because the observed variable y7
is exogenous and is a predictor of the two latent variables, it is not unreasonable to also treat it as a structural variable
when you are producing path diagrams. Hence, all three PATHDIAGRAM statements use the STRUCTADD= option to include y7
in the set of structural variables. For illustration purposes, variable y7
is labeled as “y7 is Exogenous” in the path diagrams. Although all three PATHDIAGRAM statements produce diagrams for the initial solution, they each display
the structural component in a different way.
The first PATHDIAGRAM statement produces a diagram for the full model in Figure 29.22. The structural component is not specifically emphasized. Because y7
is treated as a structural variable, it is larger than the other observed variables, and its size is comparable to that of
the default structural variables Factor1
and Factor2
.
The second PATHDIAGRAM statement produces a diagram that emphasizes the structural component of the full model in Figure 29.23. Again, because y7
is treated as a structural variable, it is much larger than the other observed variables, and it is emphasized in the same
way as the default structural variables Factor1
and Factor2
.
Figure 29.23: Including an Observed Variable as a Structural Variable: Emphasizing the Structural Component
The third PATHDIAGRAM statement produces a diagram for the structural component in Figure 29.24. Because y7
is treated as a structural variable, it is also shown in this path diagram, and its size is comparable to that of Factor1
and Factor2
.
If you do not include y7
as a structural variable in this PATHDIAGRAM statement, Figure 29.24 would display Factor1
and Factor2
as isolated variables that do not have any directional paths that connect them, making it an uninteresting path diagram display.