A Network Analysis Replication: OCD & Depression Comorbidity

based on A Network Perspective on Comorbid Depression in Adolescents with Obsessive-compulsive Disorder. Preprint version available here.

Why replicate?

There has been some recent debate over the replicability of psychological network modeling. Fortunately, the initial criticisms seem to have been greatly overstated, and resulted in large part because of statistical mistakes.

One positive aspect of this confusion is that it has reminded us the replication is an important issue. It’s important to remember that replication applies to theories and findings from individual studies, not to statistical methods such as network analysis at large (would you test if “linear regression replicates?”). Replication is a good way to ascertain the stable and reliable results from a study. It is also a good way to discover what might be unique across different types of samples.

The original sample

Earlier this year, Psychological Medicine published a network analysis of OCD and depression comorbidity authored by McNally, Mair, Mugno, & Riemann. This paper was somewhat of a pioneer in psychological network analysis: not only was this the first network study to look at the comorbidity of OCD & depression – it was also the first psychological network study to use directed acyclic graphs (DAGs) to generate hypotheses about directionality from cross-sectional data.

McNally et al. examined a sample of 408 adults with primary OCD admitted to treatment at Rogers Memorial Hospital. They reported that sadness, anhedonia, and obsession-related distress were the symptoms with the highest centrality. Their DAG model indicated that OCD symptoms generally predicted depression symptoms, rather than the other way around. They reported that distress associated with obsessions and sadness were the primary “bridge symptoms”.

The new sample

We replicated McNally et al.’s paper by examining OCD and depression in a sample of 87 adolescents. True to the original, we used admissions data from Rogers Memorial Hospital with the same measures: the Yale-Brown Obsessive Compulsive Scale (Y-BOCS) and the Quick Inventory of Depressive Symptomatology (QIDS).

This certainly isn’t a one-to-one replication — adolescents are not the same as adults! This is particularly true when we are dealing with OCD– several important differences have already been noted in the literature.

For instance, comorbidities are different — comorbid disruptive disorders such as oppositional defiant disorder and ADHD are more common in youth than adults. On the other hand, those with earlier onset cases of OCD had lower overall rates of lifetime comorbid depression than those with late onset OCD. There are also some basic lifestyle differences – adults usually have jobs, may be married, and may have children, whereas adolescents are usually single and in school.

With this in mind, we were interested to see both the similarities between the networks, as well as the differences. The similarities (e.g., the parts that “replicate”) are likely to say something universal about how OCD and depression work, and the differences (e.g., the parts that “don’t replicate”) might tell us about what makes adults and adolescents unique (or they might be spurious – we’ll have to be careful).

The results

The GLASSO networks for both the adult sample and the adolescent sample can be seen in the figure below. As you can see, they turn out to be fairly similar with a couple of important differences. We used MDS with Procrustes rotation for plotting, as this is a much more stable way to compare networks compared to the Fruchterman-Reingold. Overall, the correlation between the two networks was large (r = 0.67).

png_blog_1

We can get an even better comparison by looking at the centrality values for each network. We’ll use expected influence centrality because there are some negative edges. Once again, adults are on the left, and adolescents are on the right.

jpg_blog_2

We get very similar patterns. The expected influence centralities correlated at r = 0.75. That’s pretty high, especially considering that we are looking at different populations. It seems that OCD and depression comorbidity are pretty robust across adults and adolescents.

What’s unique? The main difference we observed was that “concentration problems” seems like a more central symptom in the adolescents. This actually makes sense when we start to consider that the adolescent life is often centered on school. Looking through the literature, we found that the number one functional complaint in pediatric OCD was “problems concentrating on schoolwork” and the number two complaint was “problems doing homework”. We obviously weren’t the first to notice that concentration problems can be a major problem for adolescents with OCD.

The takeaway

Network analysis in psychology is still relatively new, and the first replications are just starting to appear. Eiko Fried recently released a great study that examines PTSD across four different datasets (e.g., a self-replication), which shows a high degree of similarity across samples. In contrast, some conclusions about network density predicting treatment response have been recently nonreplicated.

As we move forward, more and more replications will appear. It is inevitable that some results will not replicate—that’s just part of science! However, we can avoid major problems with replication by being careful now – using large samples, running stability analyses, and being careful when interpreting our results.

 

Advertisements

4 thoughts on “A Network Analysis Replication: OCD & Depression Comorbidity”

    1. Eiko,

      That’s a surprisingly tricky question to answer. Unfortunately, DAG estimation depends somewhat on the R version and OS, so I wasn’t able to perfectly reproduce the DAG from the adult paper. In addition, you wouldn’t expect the DAGs to replicate to the same degree as the other networks, because they have a lot of additional constraints (no loops allowed, edges can only be in one direction, a very small number of total possible edges), so it’s somewhat unclear what a “good” replication would look like.

      The correlation between the edge weights between the adolescent DAG and the (imperfectly reproduced) adult DAG was 0.32. The “conceptual replication” seemed more solid — in both cases the direction went from OCD to depression and the bridges went through obsessions, not compulsions. Because the DAGs are so constrained, I think these types of “overall trends” and generated hypotheses are much more useful than trying to interpret the individual edges.

      Like

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

w

Connecting to %s