How To Use Graphviz for SEM Models and Path Diagrams

Graphviz is a free graph visualisation program that produces structural diagrams. While its primary purpose is visualization of algorithms and flow charts, it can also be used to draw path diagrams and structural equation models. The syntax resembles somewhat that of R's sem package: it uses the same idea of -> signs to represent the arrows. The variables of the model are nodes in Graphviz slang, and the relations are edges.

Example 1: Holzinger-Swineford CFA model

Holzinger and Swineford data and model is one of the favorite toy data sets in confirmatory factor analysis popularized by Joreskog (1969) and used as an example in Yuan and Bentler (2007). The data set itself is available here with SPSS syntax, here as Stata dta file, as well as in R package MBESS.

The path diagram associated with this model can be produced in Graphviz as follows:

Here's the code that produces this graph using Graphviz' dot engine. 


digraph HSCFA {

   vis -> x1 [weight=1000];

   vis -> x2 [weight=1000];

   vis -> x3 [weight=1000];

   text -> x4 [weight=1000];

   text -> x5 [weight=1000];

   text -> x6 [weight=1000];

   math -> x7 [weight=1000];

   math -> x8 [weight=1000];

   math -> x9 [weight=1000];

   vis -> math [dir=both];

   vis -> text [dir=both];

   text -> math [dir=both];

  

   x1 [shape=box,group="obsvar"];

   x2 [shape=box,group="obsvar"];

   x3 [shape=box,group="obsvar"];

   x4 [shape=box,group="obsvar"];

   x5 [shape=box,group="obsvar"];

   x6 [shape=box,group="obsvar"];

   x7 [shape=box,group="obsvar"];

   x8 [shape=box,group="obsvar"];

   x9 [shape=box,group="obsvar"];



   { rank = same; x1; x2; x3; x4; x5; x6; x7; x8; x9 }

   { rank = same; vis; math; }

   { rank = max; d1; d2; d3; d4; d5; d6; d7; d8; d9 }

 

   d1 -> x1;

   d1 [shape=plaintext,label=""];

   d2 -> x2;

   d2 [shape=plaintext,label=""];

   d3 -> x3;

   d3 [shape=plaintext,label=""];

   d4 -> x4;

   d4 [shape=plaintext,label=""];

   d5 -> x5;

   d5 [shape=plaintext,label=""];

   d6 -> x6;

   d6 [shape=plaintext,label=""];

   d7 -> x7;

   d7 [shape=plaintext,label=""];

   d8 -> x8;

   d8 [shape=plaintext,label=""];

   d9 -> x9;

   d9 [shape=plaintext,label=""];

  

}
You can try to figure out how stuff works by commenting a line out with C-style comments // and running Graphviz again. Let us go over this code and learn some tricks.
  1. digraph is the keyword that introduces a directed graph.
  2. Lines 2--13 introduce the relations between the latent variables vis (visual factor), text (reading analysis factor) and math (counting factor), as well as their relation to the observed variables x1 through x9.
  3. weight=# option gives some degree of control over the shape of the line. The greater the weight, the more straight and the shorter the line becomes. 1000 is a really big number, so the lines must come as straight as possible.
  4. The option dir=both requests a two-sided arrow.
  5. Lines 15--23 specify that the observed variables be put in boxes rather than default ovals. The group keyword indicates that these nodes belong to the same group, and should be layed out at the same level of the diagram.
  6. rank=same; followed by the list of variables additionally forces placement of all the variables at the same level.
  7. rank=max; followed by the list of variables forces those variables to appear at the very bottom of the diagram.
  8. The desription of the error terms d1--d9 generates the upward arrows. Generally Graphviz thinks about directed graphs as flows that go in a certain direction (from left to right and from top to bottom), so these arrows are going against the flow. Hence we need to force the origins of these arrows to be below their respective x1--x9 variables with rank=max; command just discussed.
  9. shape=plaintext option requests that a node is displayed without any contours around it, so that a label is reproduced as is.
  10. label="" requests that a label for this node is empty. The joint effect is that nothing at all is produced in place of d1--d9 variables.
It is more customary in the SEM literature to denote the covariances by curved two-sided arrows. Graphs like that can be forced through Graphviz, but their aesthetic appeal is probably not that great:

The change that's needed for this graph is to replace the relations between the factors by 

   vis:ne -> math:nw [dir=both];

   vis:e -> text:nw [dir=both];

   text:ne -> math:w [dir=both];
The specification node:position in place of node attaches the arrow at the given compass position. It forces some curvature into the lines, although there is no other control over the shape of the line.

Example 2: Bollen's liberal democracy model

Bollen (1993) builds a multiple traits - multiple methods (MTMM) model of liberal democracy. There are two main factors, or traits (polytical liberty and democractic rule), and three methods factors (sources of data, political science researchers A. Banks, R. D. Gastil, and L. R. Sussman). The path diagram is as follows (the individual error terms are omitted):
Self-check: produce the above diagram using Graphviz. Hints: multiple line lables are produced by inserting the C-style end-of-line character "\n", and the fixed loadings are produced with label option.
© Stas Kolenikov, 2009.