Edinburgh Napier University

  Our trial1 was not designed to change clinical practice. It was a preliminary trial, which needs to be followed up by a larger, pragmatic multicentre study. It is important not to overinterpret our data, and we explicitly advised against discontinuation of medication. We claimed the trial showed that cognitive therapy was safe and acceptable, not safe and effective. No participants were included on the basis of attenuated psychotic symptoms (our entry criteria included positive symptoms above a threshold on the positive and negative syndrome scale [PANSS]). Our combined sample had a mean baseline PANSS score of 72, with the recommended cutoffs for mildly ill being 58 and moderately ill being 75; therefore, our sample consists of people with both levels of difficulty. Duration of psychosis and other sample characteristics are reported elsewhere.2 We can clarify that the proportion of patients starting medication in each group was 27%. An investigation of the components of therapy is being undertaken and will be published.

Standardised effect sizes (and 95% CIs) for the PANSS subscales are as follows: PANSS positive (−0·45, −0·81 to −0·09), PANSS negative (−0·22, −0·51 to 0·07), and PANSS general (−0·47, −0·78 to −0·16). The claim that missing data is over 50% at follow-up points is based on use of the randomised number as denominator, rather than the number who could have potentially provided outcome assessments given the planned variable length of follow-up (adopted to provide value for money to the funder, allowing maximal recruitment, and the most complete data we could gather in the lifetime of the trial). The proportion of non-missing PANSS data at 9 months (the final point at which an outcome assessment could have been provided by all randomised participants) is about 61%, and at 18 months (using those who could have been followed up within that timeframe as the denominator) it is about 67% (table 21)—rates that are not ideal, but acceptable.

Two correspondents question the use of a mixed (random) effects model based on an assumption that outcome data are missing at random. We provided a technical reference to indicate what was meant by missing at random.3 Its meaning is not quite the same as that implied in everyday English, the latter having the technical description more akin to “missing completely at random” (MCAR). Clearly, MCAR does not hold. The missing-at-random assumption will only ever be an approximation of the truth (like the modelling assumptions behind any statistical analyses). Our critics seem to assume that if the correct missing data mechanism had been chosen then the estimated treatment effect would be closer to the null. We suggest they would not have been prepared to accept modifications to the prespecified analysis if it led to a greater treatment effect. As we drift away from the prespecified statistical analysis plans, the risk of selecting methods of analysis that favour the outcomes we are looking for becomes a much more serious cause for concern than minor violations of the assumptions underlying the main analyses. Although missing data lead to lower precision (reduced statistical power) and potential biases, the fact that in the present trial we found statistically significant effects implies that it was not underpowered. Lack of power would only have been a serious concern if we had observed a promising but statistically insignificant result.

In response to the hypothesis concerning loss to follow-up in generic community services versus early intervention services, we had 9-month PANSS outcomes for ten of 20 (50%) and 12 of 17 (71%) patients allocated to cognitive therapy in early intervention services and community services, respectively. The numbers for treatment as usual (TAU) were 15 of 22 (68%) and 8 of 15 (53%). In a comparison of the proportions of missing data in the four groups, χ2=2·52 (df=3, p=0·472). We check the sensitivity of our results to differences in proportions of missing data by adding service type as an extra covariate in the mixed-effects model (again assuming missing at random). The revised estimate of the treatment effect on PANSS total is −6·64 (standard error [SE] 2·17), the original estimate being −6·52 (SE 2·18). With the addition of a treatment by service interaction to the mixed model, the effect of cognitive therapy in early intervention services is estimated to be −10·97 (SE 2·78; p