Who Gets Better in Therapy? What Research Says About Predicting Treatment Response for Anxiety

Before starting treatment, the question that matters most to patients is whether it will work for them. Not on average, not for a typical patient, but for them specifically. Research has tried hard to answer this question. The honest answer is that it mostly cannot, not yet.

What has been found

Systematic reviews of predictors of treatment outcome in anxiety disorders have identified several variables that are statistically associated with response:

Baseline severity. More severe symptoms are generally associated with more absolute improvement (a higher score has more room to fall), but paradoxically, also with a lower probability of reaching a predefined response threshold (like a 50% reduction) [1]. How you define “response” changes the direction of the finding, which is one reason studies disagree.

Comorbid depression. Several studies find that comorbid depression predicts poorer outcomes for anxiety-specific CBT, but the finding is inconsistent. In a prospective analysis of two CBT trials for anxiety, comorbidity was associated with a poorer course, but the effect was modest [1].

Duration of illness. Longer duration before treatment is sometimes associated with poorer outcomes, possibly reflecting more entrenched patterns. But duration is confounded with severity, comorbidity and age.

Dropout. A systematic review of predictors of dropout from CBT for social phobia found that younger age, lower education and higher severity predicted premature ending, but the predictors varied across studies and many were not replicated [2].

Expectancy. People who expect treatment to help tend to do better. Whether this reflects genuine engagement, placebo-like mechanisms, or simply that less impaired people are more optimistic is unclear.

Why individual predictors fail

The fundamental problem is that individual predictors explain very little variance. A predictor that explains 5% of the variance in outcome is statistically significant in a large study but practically useless for a clinician trying to advise one patient. “Comorbid depression predicts somewhat poorer outcomes on average” does not tell you whether this person with comorbid depression will respond.

Several methodological issues compound the problem:

• Heterogeneity within diagnoses. Two people with generalized anxiety disorder may share a diagnostic label but differ in virtually every other way. The diagnosis is too broad a category for prediction.
• Small trials. Most anxiety treatment trials have 50–100 participants, far too few to detect reliable predictors, especially when multiple predictors are tested.
• Inconsistent outcome measures. Some studies use response rates, others remission rates, others continuous symptom change. The choice of outcome affects which predictors emerge.
• Multiple testing. When studies test many potential predictors, some will reach significance by chance.

Machine learning approaches

Machine learning offers the promise of combining many weak predictors into a stronger composite. Rather than asking whether any single variable predicts outcome, algorithms can integrate demographics, symptom profiles, comorbidity, treatment history and other variables to generate a prediction.

In depression, a machine learning study used data from a large trial to build a model predicting treatment response to a specific antidepressant. The model achieved modest accuracy (around 65%) in an external validation sample [4]. The improvement over chance was real but small.

Comparable work in anxiety is less advanced. The studies that exist tend to be developed on one dataset and lack external validation. A model that works well in the training data may not generalize to a different clinic, population or treatment protocol.

The Personalized Advantage Index

DeRubeis and colleagues proposed the Personalized Advantage Index (PAI), a method for using comparative trial data to estimate which treatment each individual would benefit from most [3]. The logic: in a trial comparing two treatments, you model the interaction between patient characteristics and treatment assignment. For a new patient, you estimate their expected outcome under each treatment and recommend the one with the better predicted outcome.

In proof-of-concept analyses, patients who received their PAI-recommended treatment did better than those who received the alternative. In a subsequent study in a routine clinical setting, a machine learning approach to treatment matching showed that patients who received their algorithmically recommended treatment had better outcomes than those who were mismatched [5].

These results are promising but early. The PAI requires data from a prior head-to-head trial, which limits applicability. External replications are few. And the improvements, while statistically significant, are modest in absolute terms.

What routine monitoring can do

While prediction before treatment remains weak, tracking progress during treatment is more informative. Routine outcome monitoring — measuring symptoms at each session and comparing with expected trajectories — can identify patients who are not responding as expected, triggering a review or change in approach [6].

This is prediction of a different kind: not “will this person respond?” but “is this person responding?” The latter is more tractable because it uses actual data rather than baseline predictions. It also enables a responsive clinical approach: start with the evidence-based first choice, monitor closely, and adapt early if the trajectory suggests problems.

What this means for reading research

When a study reports predictors of treatment response:

1. Check the sample size. Small studies produce unreliable predictor estimates.
2. Ask whether the predictors were pre-specified or identified post hoc by testing many variables.
3. Look for external validation. A predictor found in one dataset must be tested in another.
4. Consider practical significance, not just statistical significance. A predictor that explains 3% of outcome variance is statistically real and clinically useless.

For the related question of whether the therapist matters more than the technique, see therapist effects in outcomes.

When to look for more support. If you have tried a treatment for anxiety and it hasn’t helped, that doesn’t mean nothing will. Response varies by person, and alternatives exist. Talk to your GP or therapist about adjusting the approach. If you are in crisis, call or text 988 (US) or Samaritans on 116 123 (UK and Ireland).