Alliance and Outcome: A Reciprocal Loop, Not a One-Way Cause

The therapeutic alliance is the most cited common factor in our field, and the number attached to it is remarkably stable: in meta-analysis across hundreds of studies, the alliance-outcome correlation sits near r = 0.28 (Flückiger et al.). That figure is real and it replicates. What it cannot tell you is the direction of the arrow: does a good alliance produce improvement, or does early improvement produce a good alliance? A naturalistic inpatient study from LMU Munich pulls the two apart, week by week.

What the weekly numbers actually show

Diehm and colleagues followed 164 adults through a 10-week inpatient CBASP programme, a treatment built specifically for persistent depressive disorder (85.3% of the sample met criteria). Every week patients rated their depression (BDI-II) and the working alliance (WAI-S-P), producing a dense time series rather than the usual two-point snapshot. Over treatment, depression fell from severe to moderate (mean 30.2 to 23.0; slope b = -0.58, p < 0.001) and the alliance climbed steadily (b = 0.84, p < 0.001).

Then the two series were lagged against each other. A stronger alliance one week predicted lower symptoms the next (beta = -0.08, 95% CI [-0.12, -0.05]); lower symptoms one week predicted a stronger alliance the next (beta = -0.14, 95% CI [-0.21, -0.08]). Both paths were significant, both were small, and the symptom-to-alliance path was, if anything, the larger. Attachment insecurity moderated neither.

The direction problem, and what the design can settle

This is where the sober reading lives. A single alliance-outcome correlation blends two very different things: differences between patients (some people form warmer relationships and are also less ill) and fluctuations within a patient (a given person's alliance rising above their own baseline). The authors modelled them separately. The within-person coupling was significant in both directions (beta = -0.12 for alliance-to-symptoms; beta = -0.09 for symptoms-to-alliance). The between-person component was non-significant in both directions (p = 0.819 and p = 0.953). Patients who happened to report generally warmer alliances did not go on to have systematically lower symptoms.

What the design establishes is temporal coupling: alliance and symptom change move together, and each nudges the other the following week. What it cannot establish is causation. There is no control group, no manipulation of the alliance, and only self-report; an unmeasured third factor – expectancy, early response, a general sense of getting better – could drive both series at once. The authors say as much, noting that without a control group causal inferences remain tentative. The honest verdict is the one worth carrying into supervision: robust association, contested causation.

What this changes at the table

Practically, the finding argues against treating the alliance as a lever you pull to produce outcome, and for treating it as a signal that co-moves with recovery. A dip in the alliance may be an early read-out of a stalling patient as much as a cause of the stall; a warming alliance may partly be the patient reporting that the work is landing. That reframing has teeth for measurement-based care: reading alliance and symptom scores side by side, week to week, tells you more than either alone, and neither should be over-interpreted as the engine. It also cautions against selling the alliance to trainees as the active ingredient. It is a genuine, reciprocal part of the process – not a proven cause, and not a substitute for a working method.