IBM® SPSS® Amos™ 28

Note: This video shows how to perform a d-separation analysis (Kline, 2016, ch 8, Pearl, 2009, Pearl, Glymour and Jewell, 2016).

There are three ways to perform a d-separation analysis. You can either:

1.Click View > Analysis Properties > Output and put a check mark next to D-separation. If you choose this option, the d-separation analysis will be performed when you click Analyze > Calculate Estimates. The output of the analysis will appear in the text output.
 
or

2.Click Analyze > D-Separation Preview. If you choose this option the d-separation analysis will be performed immediately. The output of the analysis will be displayed in a new window.
 
or

3.Click Tools > Export to > DAGitty. This copies your model to the clipboard and attempts to open a webpage at dagitty.net in your web browser. The video (linked to above) shows how to paste your model from the clipboard into the dagitty.net web page.

Using option 1 or 2, a d-separation analysis produces the following table for the model and data in Example 39 of the user's guide.

 

 

Looking at the first row of the table, the model implies that q3 and q1 are independent when q2 is "held constant". That is, q3 and q1 are independent in any subpopulation of people who share the same q2 score. This implies that the partial correlation between q3 and q1 with q2 "held constant" is zero in the population. The corresponding sample partial correlation is .333. The final two columns provide a test of the null hypothesis that the population partial correlation is zero, taking into account the facts that the sample partial correlation is .333 and that the sample size (N) is 22. 1.541 is a t statistic that has a t distribution with N - 3 = 19 degrees of freedom if the partial correlation is zero in the population (Weatherburn, 1968, page 256). The two-tailed "p value" is .140. That is, with a correct model the probability is .140 that a sample partial correlation would be as far from zero as it was in this sample. Each row of the table is interpreted similarly. None of the p values in the table is very close to zero, so that from this point of view the model in Example 39 is compatible with the data.

You can use option 2 even if you have no data, only a model. In that case you will get only the first column of the above table.

If you use option 3, dagitty.net gives the following list of partial correlations.

q1 ⊥ q3 | q2

q1 ⊥ q4 | q3

q1 ⊥ q4 | q2

q2 ⊥ q4 | q3

DAGitty's list of partial correlations is a subset of Amos's. If the model is correct and the partial correlations on DAGitty's list are zero in the sample, then the other partial correlations on Amos's list will also be zero.