Depression and anxiety co-occur so often that comorbidity is closer to the rule than the exception. The question is why — and two frameworks answer it differently. The answer you accept quietly changes how you read every score a patient hands you.

The latent variable answer: a shared cause behind the symptoms

For most of the twentieth century, the default model was reflective: symptoms are indicators of an underlying condition. Low mood, loss of interest, sleep disturbance — these are not the disorder; they are effects of it, the way a fever is an effect of an infection. Comorbidity, on this view, means two latent variables (depression, anxiety) are correlated, or share a common cause upstream. The symptoms co-occur because the causes do.

This is the logic that justifies summing items into a total score. If nine PHQ-9 items all reflect one latent depression factor, adding them up is a reasonable estimate of that factor. The items are interchangeable indicators; the sum is the signal.

The network answer: symptoms cause each other

The network theory of mental disorders rejects the common cause and puts the causation between the symptoms themselves. Insomnia produces fatigue; fatigue produces concentration problems; concentration problems produce worry about performance; worry feeds back into insomnia. There is no hidden depression entity doing the work — the disorder is the pattern of mutually reinforcing symptoms.

On this account, comorbidity has a concrete mechanism. Depression and anxiety co-occur because specific symptoms bridge the two constructs. Trouble concentrating, sleep disturbance, restlessness — these belong to both syndromes and act as the connective tissue. Deactivate a bridge symptom and the two clusters loosen; leave it active and they pull each other along. Comorbidity is not two correlated factors but one connected graph.

Why the difference is not academic

Consider what each model implies for a total score.

Under the latent view, a PHQ-9 of 14 is a point estimate of depression severity, full stop. Under the network view, two patients with the same 14 can have entirely different structures — one anchored on a highly central symptom that props up the rest, another with the same points spread across loosely connected symptoms. The number is identical; the clinical situation is not. The network asks which symptoms, and how they connect, not just how many.

This is exactly where a symptom network becomes a reading tool rather than a replacement for judgment. Estimating the network (via a regularized partial-correlation method such as EBICglasso; see Epskamp, Borsboom & Fried, 2018) and inspecting centrality tells you which symptoms sit at the crossroads and which sit at the edges — a description of structure, not a prescription for what to do about it.

The honest position: two lenses, not a verdict

It would be tidy to declare the network model the winner. It is not that simple, and overclaiming does the field no favors. The latent variable tradition earned its dominance for real reasons — measurement invariance, decades of validated instruments, tractable statistics. The network approach adds a structural view the sum score cannot express. Held side by side, they answer different questions about the same data.

That is the stance worth keeping: not integration into one true model, and not a contest to be settled, but two lenses laid against the same scores — each showing something the other cannot. When depression and anxiety arrive together, the latent lens asks what they share; the network lens asks what connects them. Both readings are worth having in front of you.

---

Method note: network estimation in this post refers to regularized partial-correlation networks (EBICglasso). For the estimation and stability workflow, see Epskamp, Borsboom & Fried (2018), “Estimating psychological networks and their accuracy,” Behavior Research Methods. Bridge symptoms follow the bridge-centrality framework of Jones, Ma & McNally (2019).