Dr. Volavka and Colleagues Reply

To the Editor: We appreciate the opportunity to discuss our results. First we address the concerns about the risperidone doses we used. Dr. Meibach suggests that “It has been well established that risperidone doses above 10 mg/day are less effective than lower doses,” referring us to his publication (Marder and Meibach, 1994). However, that statement is not supported by their data. Their dose of 16 mg/day of risperidone was more effective than 2 mg/day or 10 mg/day; 6 mg/day was “as effective as 16 mg” (p. 825).

Drs. Osser and Akhter suggest that “most data and expert opinions indicate better results with risperidone doses below 8 mg/day” than with higher doses. To support their statement, they cite five articles presenting original data. Except for the report by Marder and Meibach (1994), the articles present no data on doses ≥8 mg/day; thus, higher doses are not compared with the lower ones in these articles (Ho et al., 1999; Lane et al., 2000; Conley et al., 2001; and Love et al., 1999). Therefore, they cannot provide empirical support for Drs. Osser and Akhter’s statement. A records review (Love et al., 1999) found that patients receiving 2–4 mg/day of risperidone were more likely to be discharged than those taking 6 mg/day. However, the patients’ doses were determined by clinical judgment, and thus the more seriously ill patients (who were therefore less likely to be discharged) were more likely to receive the higher dose.

Would the efficacy of risperidone have been better had we used lower doses? We address this issue indirectly. Blood samples for the determination of antipsychotic plasma levels were drawn at several time points. Our validated assay of risperidone and 9-hydroxyrisperidone (in progress) is expected to yield results comparable to those obtained by others (1). Similar to others (1), we summed the levels of risperidone and 9-hydroxyrisperidone, creating a variable we labeled RISSUM. Since plasma levels of RISSUM are known to correlate with risperidone dose, the clinical outcome in patients with low levels of RISSUM may yield insights about what would happen had we used lower risperidone doses in our study. At the end of the 8-week fixed-dose period, RISSUM values were available for 36 patients. The median RISSUM value was 42.5 ng/ml. (Incidentally, in patients treated with 6 mg/day of risperidone, the average RISSUM value reported by others was 47.9 ng/ml [1]). At 8 weeks, the average improvement in total score on the Positive and Negative Syndrome Scale for the subgroup of 18 patients whose RISSUM values exceeded the median was 5.39 points (SD=16.89). The analogous number for the 18 patients scoring below the median was 2.85 points (SD=18.81). The difference was not statistically significant, but the response tended to be better with higher plasma levels. This suggests that lower doses would not have been more effective in this group.

That said, we recognize that many clinicians feel that our target risperidone dose (8 mg/day) was too high. We would have avoided that criticism—and perhaps reduced some side effects—by using 6 mg/day instead. We did not do that for the reasons explained in the article. Although empirical evidence suggests that our efficacy results would not have differed substantially, the lower dose would have been closer to current mainstream prescription patterns, thus making the study appear more relevant for many clinicians. Finally, as we stated in the article, the dose of risperidone in the variable-dose period of the study was probably too high. Dr. Meibach is concerned about the high doses of olanzapine we used; however, doses above 20 mg/day are commonly used in clinical practice (2).

Dr. Adityanjee feels that our average dose of haloperidol in the variable-dosing phase, 27.5 mg/day, “cannot be justified.” However, the average haloperidol dose was 28 mg/day in a large, well-respected study of patients with treatment-resistant schizophrenia (3). To support his feelings about our dosing, Dr. Adityanjee refers us to two articles that included first-episode patients (McEvoy et al., 1991; Farde et al., 1992). It is well known that such patients require (and tolerate) much lower doses than those who are in the later stages of schizophrenia, such as our patients. Another study (Wolkin et al., 1989) showed that the average dose of 55 mg/day of haloperidol did not yield any advantage in comparison with 39 mg/day; these doses were too high to be relevant here. The conclusion drawn from a meta-analysis (Geddes et al., 2000) was that the efficacy of haloperidol is better when doses below 12 mg/day are used, but that conclusion is invalid since studies using doses governed by clinical judgment were included (e.g., reference 3). In such studies, doctors who see a poor response sometimes increase the dose, but this change may not improve efficacy (4). Thus, the high dose may be a consequence—rather than a cause—of poor response.

Another concern pertains to potential unblinding of raters for the Positive and Negative Syndrome Scale due to the raters observing extrapyramidal symptoms; in Dr. Meibach’s view, the presence of extrapyramidal symptoms might give the rater a hint that the patient was randomly assigned to risperidone. However, the extrapyramidal symptom ratings, completed on the same day as the Positive and Negative Syndrome Scale, showed a substantial overlap between risperidone and other treatments (Table 2 in our article). This overlap would have prevented the use of extrapyramidal symptoms to correctly guess the patients’ treatment assignment.

Drs. Osser and Akhter speculate that the cohort of patients enrolled after the olanzapine arm began had a better prognosis and that this was manifested by the improvements with risperidone in that cohort. However, there was only a small difference between the cohorts among clozapine patients. This argues against a cohort effect. Such effect, if present, would have resulted in an increase of the improvement rate with clozapine that would have been similar in size to that observed with risperidone. To support their speculations, Drs. Osser and Akhter quote three articles, all of which compared clozapine and risperidone in patients with treatment-resistant schizophrenia. In one of them (Azorin et al., 2001), clozapine was found superior to risperidone. The other two articles showed no difference, probably because of ineffective doses of clozapine (Bondolfi et al., 1998) or because the study group was too small (Wahlbeck et al., 2000). It is not clear how these articles could support Drs. Osser and Akhter’s speculation. Nevertheless, we admit that we cannot prove that there was no cohort effect.

Drs. Osser and Akhter mention that two patients receiving risperidone developed seizures and suggest that we should have discussed that finding. However, there were no seizures in the patients taking risperidone; a correction of this error in our article was published in the December 2002 issue of the Journal (p. 2132). Finally, Drs. Osser and Akhter note the modest size of the observed improvements. We agree.

Regarding Dr. Gupta’s question on rater blinding for the clozapine patients, the raters of the Positive and Negative Syndrome Scale did not ask the patients about salivation problems, and the patients generally did not spontaneously complain about this. There were no suicides or suicide attempts during the study. There was no significant difference among the treatments in the effects on dyskinesias (as assessed with the Extrapyramidal Symptom Rating Scale). A report on glucose and lipid abnormalities was published in the Journal(5).

Dr. Messori takes us to task for what she sees as our failure to “carry out the statistical analyses originally planned regardless of what happens with random assignment of subjects.” However, this dogma is not universally accepted. Feinstein, an outstanding biostatistician, explained that the random selection process “does not protect against inadvertent fortuitous distortion or ‘the luck of the draw.’” Therefore, after the group is drawn, the investigator “will want to have a method of checking what has happened, and, if substantial distortion occurred despite all the precautions, he [or she] will want to be able to eliminate or adjust for the effects of distortion” (6, pp. 608–609).

The data-analytic cycle consists of two components: model selection and model checking (7). Investigators should implement model-checking procedures to see if their models are internally consistent with the data. “What we should not do is to accept a class of models a priori, put the data through a standard package and accept the output as an appropriate analysis….We believe strongly that editors and referees should not accept articles, and regulators should not accept submissions, unless the authors can show that they have checked their models” (7, p. 2318).

Contrary to Dr. Messori’s claim, our analytical method was not selected in an open-ended manner; it was specified a priori as one of the two methods to be used. Consequently, her alarming suggestion that the double-blind design makes little sense given our analytical approach is not warranted.