Background. The quality of the therapeutic alliance (TA) has been invoked to explain the equal effectiveness of different psychotherapies, but prior research is correlational, and does not address the possibility that individuals who form good alliances may have good outcomes without therapy. Method. We evaluated the causal effect of TA using instrumental variable (structural equation) modelling on data from a three-arm, randomized controlled trial of 308 people in an acute first or second episode of a non-affective psychosis. The trial compared cognitive behavioural therapy (CBT) over 6 weeks plus routine care (RC) v. supportive counselling (SC) plus RC v. RC alone. We examined the effect of TA, as measured by the client-rated CALPAS, on the primary trial 18-month outcome of symptom severity (PANSS), which was assessed blind to treatment allocation. Results. Both adjunctive CBT and SC improved 18-month outcomes, compared to RC. We showed that, for both psychological treatments, improving TA improves symptomatic outcome. With a good TA, attending more sessions causes a significantly better outcome on PANSS total score [effect size −2.91, 95% confidence interval (CI) −0.90 to −4.91]. With a poor TA, attending more sessions is detrimental (effect size +7.74, 95% CI +1.03 to +14.45). Conclusions. This is the first ever demonstration that TA has a causal effect on symptomatic outcome of a psychological treatment, and that poor TA is actively detrimental. These effects may extend to other therapeutic modalities and disorders.