Rogers Memorial Hospital Presented on the symposium entitled “Envisioning the Clinical Integration of Network Analysis and CBT: New Developments.” McNally, R. J. (Chair). Association of Behavioral and Cognitive Therapies. October, 2016. New York, NY. Comorbid Obsessive-Compulsive Disorder and Depression: A Network Analytic Approach Richard J. McNally Patrick Mair Harvard University Beth L. Mugno Bradley C. Riemann Rogers Memorial Hospital
Comorbidity As many as two-thirds of people with OCD develop major depression OCD usually precedes onset of depression Successful behavior therapy for OCD often reduces depression when overall severity is moderate
Traditional View of Comorbidity (Cramer et al Traditional View of Comorbidity (Cramer et al., 2010, Behavioral & Brain Sciences)
Network View of Psychiatric Comorbidity (Cramer et al Network View of Psychiatric Comorbidity (Cramer et al., 2010, Behavioral & Brain Sciences)
What is the Network Structure of OCD and Depression?
Method Subjects Assessed at intake 408 patients with primary OCD Rogers Memorial Hospital Assessed at intake Yale-Brown Obsessive Compulsive Scale (Y-BOCS) 10 items (0 – 4) Quick Inventory of Depression Symptomatology –Self-Report (QIDS-SR 16) 16 items (0 -3)
Bayesian Network A Bayesian network analysis returns a directed acyclic graph (DAG) The “hill climbing” algorithm tests for conditional independence relations among nodes
Averaged Bayesian Network 1000 iterations of the network (“bootstrapping samples”) An edge remains only if its appears in at least 85% of the iterations The direction of an edge? X > Y percentage Y > X percentage
Averaged Bayesian Network The “importance” of an edge BIC value (Bayesian Information Criterion) The larger the absolute magnitude, the more “damaging” to model fit if one were to remove the edge from the network Some edge examples: Sadness > guilt Strength = 0.97 Direction = 0.903 Comptime > compdistress Strength = 0.99 Direction = 0.598
Results
Averaged Bayesian Network
Averaged Bayesian Network (“Scutari Method”)
Unresolved Issues Candidate causal system? No important unmeasured variables? Are DAGs clinically plausible?
Thank you!
Additional slides that time limits prohibited me from presenting on the NYC ABCT symposium.
Edge thickness signifies the probability of an edge’s direction occurring in the direction predicted. 1000 bootstrap samples and 85% threshold for edge retention (Sachs method)
Edge thickness reflects the percentage of time an edge was pointing in the direction depicted in the graph. Thin edges may signify possible cycles.
Graphical LASSO Least Absolute Shrinkage and Selection Operator Edges are partial correlations Computed via “regularization” Small edges driven to zero Likely “false alarms” vanish from the graph Undirected network
1000 bootstrap samples Estimate the robustness of the edges Estimate the robustness of the centrality metrics
Results
Partial Correlation Network (graphical LASSO)
Confidence intervals for edge strength from the glasso with OCD and depression symptoms October 26 2016