Yair Goldberg

Initial Severity of Schizophrenia and Efficacy of Antipsychotics: Participant-Level Meta-analysis of 6 Placebo-Controlled Studies

IMPORTANCE Antipsychotic drugs constitute the mainstay in the treatment of schizophrenia, and their efficacy is well established in hundreds of randomized clinical trials. However, it is not known whether they are effective or how effective they are across the wide range of baseline symptom severity. OBJECTIVE To examine the influence of baseline severity of schizophrenia on the efficacy of antipsychotic drugs. DESIGN, SETTING, AND PARTICIPANTS Meta-analysis of participant-level data from 3 pivotal randomized trials of acute schizophrenia (n = 611) and 3 pivotal trials in patients with predominantly negative symptoms of schizophrenia (n = 475). INTERVENTIONS Olanzapine or risperidone vs placebo, and amisulpride vs placebo. MAIN OUTCOMES AND MEASURES Change scores on the Positive and Negative Syndrome Scale (PANSS; score range, 30-210) and the Scale for the Assessment of Negative Symptoms (SANS; score range, 0-125) up to 6 weeks after baseline. The relationship between baseline and change scores for the drug and placebo groups was examined with 8 competing mixed-effects models for repeated measures. RESULTS The best-fitting models showed that, for both types of patients, the interactions between baseline symptom severity and treatment were statistically significant (P < .01). The greater the baseline severity was, the greater the magnitude of the differences was between active treatment and placebo. In acute treatment, the mean differences in PANSS change scores were 9.5 points for patients who were mildly ill at baseline (baseline PANSS score of 58), 13.7 for moderately ill patients (baseline PANSS score of 75), 18.8 for markedly ill patients (baseline PANSS score of 95), and 24.0 for severely ill patients (baseline PANSS score of 116). In treatment of predominantly negative symptoms, the mean differences in SANS change scores were 1.7 for those who were moderately ill (baseline SANS score of 55), 5.7 for markedly ill patients (baseline SANS score of 70), and 9.7 for severely ill patients (baseline SANS score of 85). CONCLUSIONS AND RELEVANCE We can expect benefits of antipsychotic drugs for the full spectrum of patients likely to be treated for acute schizophrenia and for highly symptomatic patients with predominantly negative symptoms. Toward the mildest end of the spectrum, clinicians need to be aware that patients benefit less in terms of symptom improvement but may experience full adverse effects of antipsychotics. Clinicians also need to be aware that in addition to the treatment of active symptoms, which was the focus of this study, antipsychotics have another important action, namely to prevent relapses among patients in remission.

Q-LEARNING WITH CENSORED DATA

We develop methodology for a multistage-decision problem with flexible number of stages in which the rewards are survival times that are subject to censoring. We present a novel Q-learning algorithm that is adjusted for censored data and allows a flexible number of stages. We provide finite sample bounds on the generalization error of the policy learned by the algorithm, and show that when the optimal Q-function belongs to the approximation space, the expected survival time for policies obtained by the algorithm converges to that of the optimal policy. We simulate a multistage clinical trial with flexible number of stages and apply the proposed censored-Q-learning algorithm to find individualized treatment regimens. The methodology presented in this paper has implications in the design of personalized medicine trials in cancer and in other life-threatening diseases.

Initial symptom severity of bipolar I disorder and the efficacy of olanzapine: a meta-analysis of individual participant data from five placebo-controlled studies

BACKGROUND The efficacy of antipsychotics across the initial severity range in patients with acute mania remains unclear. Therefore, we examined the influence of baseline severity on the efficacy of olanzapine. METHODS We did an individual participant data meta-analysis of double-blind, randomised controlled trials that compared olanzapine with placebo, identified through searches of the ClinicalStudyRequest.com database on Feb 2, 2016. We included patients with acute mania associated with bipolar I disorder. We examined the association between baseline and change scores on the Young Mania Rating Scale (YMRS; range 0-60) up to 3 weeks for olanzapine versus placebo groups using eight increasingly complex competing mixed-effects models for repeated measures. FINDINGS We identified 33 reports, five (15%) of which were eligible and contained data for 939 patients (552 received olanzapine; 387 received placebo). The interaction between baseline severity and treatment was significant (β=0·22, 95% CI 0·05-0·39; p=0·013). The greater the baseline severity, the greater the magnitude of the differences between olanzapine and placebo was expected. The mean estimated YMRS scores were reduced at 3 weeks in both groups, but were greater with olazapine than placebo by 2·56 points for patients with a baseline score of 20-25 (9·26 for olanzapine vs 6·70 for placebo; effect size 0·35, 95% CI 0·11-0·60), by 4·74 points for a baseline score of 25-35 (14·25 vs 9·51; 0·58, 0·34-0·86), and by 8·01 points for a baseline score of 35-60 (21·72 vs 13·71; 0·70, 0·31-1·23). INTERPRETATION Benefits of olanzapine can be expected for patients across the full spectrum of symptom severity who are likely to be treated for acute mania. Less severely ill patients seem to benefit less in terms of olanzapine efficacy, but still experience the same side-effects as more severely ill patients. Thus, clinicians and patients should carefully consider the benefit-to-risk ratio of olanzapine and its additional, prophylactic effect against relapse in the long term. The generalisability of these results to other antipsychotics, trial designs, and medical conditions remains to be established. FUNDING None.

LDR-LLE: LLE with Low-Dimensional Neighborhood Representation

The local linear embedding algorithm (LLE) is a non-linear dimension-reducing technique that is widely used for its computational simplicity and intuitive approach. LLE first linearly reconstructs each input point from its nearest neighbors and then preserves these neighborhood relations in a low-dimensional embedding. We show that the reconstruction weights computed by LLE capture the high -dimensional structure of the neighborhoods, and not the low -dimensional manifold structure. Consequently, the weight vectors are highly sensitive to noise. Moreover, this causes LLE to converge to a linear projection of the input, as opposed to its non-linear embedding goal. To resolve both of these problems, we propose to compute the weight vectors using a low-dimensional neighborhood representation. We call this technique LDR-LLE. We present numerical examples of the perturbation and linear projection problems, and of the improved outputs resulting from the low-dimensional neighborhood representation.

Exposure to genocide and the risk of schizophrenia: a population-based study.

BACKGROUND No evidence exists on the association between genocide and the incidence of schizophrenia. This study aims to identify critical periods of exposure to genocide on the risk of schizophrenia. METHOD This population-based study comprised of all subjects born in European nations where the Holocaust occurred from 1928 to 1945, who immigrated to Israel by 1965 and were indexed in the Population Register (N = 113 932). Subjects were followed for schizophrenia disorder in the National Psychiatric Case Registry from 1950 to 2014. The population was disaggregated to compare groups that immigrated before (indirect exposure: n = 8886, 7.8%) or after (direct exposure: n = 105 046, 92.2%) the Nazi or fascist era of persecutions began. The latter group was further disaggregated to examine likely initial prenatal or postnatal genocide exposures. Cox regression modelling was computed to compare the risk of schizophrenia between the groups, adjusting for confounders. RESULTS The likely direct group was at a statistically (p < 0.05) greater risk of schizophrenia (hazard ratio = 1.27, 95% confidence interval 1.06-1.51) than the indirect group. Also, the likely combined in utero and postnatal, and late postnatal (over age 2 years) exposure subgroups were statistically at greater risk of schizophrenia than the indirect group (p < 0.05). The likely in utero only and early postnatal (up to age 2 years) exposure subgroups compared with the indirect exposure group did not significantly differ. These results were replicated across three sensitivity analyses. CONCLUSIONS This study showed that genocide exposure elevated the risk of schizophrenia, and identified in utero and postnatal (combined) and late postnatal (age over 2 years) exposures as critical periods of risk.

Comment on “Dynamic treatment regimes: technical challenges and applications”

Inference for parameters associated with optimal dynamic treatment regimes is challenging as these estimators are nonregular when there are non-responders to treatments. In this discussion, we comment on three aspects of alleviating this nonregularity. We first discuss an alternative approach for smoothing the quality functions. We then discuss some further details on our existing work to identify non-responders through penalization. Third, we propose a clinically meaningful value assessment whose estimator does not suffer from nonregularity.

Initial severity and efficacy of risperidone in autism: Results from the RUPP trial

BACKGROUND Risperidone is a common psychopharmacological treatment for irritability in autism spectrum disorder (ASD). It is not well-established how effective risperidone is across the initial symptom severity range. This study aims to examine the influence of baseline severity on the efficacy of risperidone in the treatment of ASD. METHODS Participants were from the NIMH funded RUPP multisite, randomized, double-blind trial that compared risperidone to placebo to treat autistic disorder with severe tantrums, aggression, or self-injury. Participants were aged 5 to 17, and randomly assigned to risperidone (n=49) or placebo (n=52). Baseline and change scores were computed with the Aberrant Behavior Checklist (ABC) parent assessed scales with irritability as the primary outcome, as well as the clinician assessed ABC Irritability subscale, and Clinical Global Impression Scale. RESULTS The relationship between baseline severity and change scores for the risperdone and placebo groups was examined with eight competing three-level mixed-effects models for repeated measure models. Significant (P<0.01) interactions between treatment and baseline severity were observed for parent ABC ratings of irritability and lethargy only. Greater magnitudes of the differences between risperidone and placebo were observed from moderate to very severe baseline severity on irritability and lethargy. Initial severity values over approximately 30 had a strong effect on symptom change [irritability: effect size (ES)=1.9, number needed to treat (NNT)=2, lethargy ES=0.9, NNT=5]. CONCLUSIONS Parents may expect benefits of risperidone on irritability and lethargy with moderate to severe symptoms of ASD. TRIAL REGISTRATION Registry name: ClinicalTrials.gov, trial identifier: NCT00005014, URL: http://www.clinicaltrials.gov/ct2/show/NCT00005014?term=NCT00005014&rank=1, registered on March 31, 2000.

Support vector regression for right censored data

We develop a unified approach for classification and regression support vector machines for data subject to right censoring. We provide finite sample bounds on the generalization error of the algorithm, prove risk consistency for a wide class of probability measures, and study the associated learning rates. We apply the general methodology to estimation of the (truncated) mean, median, quantiles, and for classification problems. We present a simulation study that demonstrates the performance of the proposed approach.

Joint modeling of dropout and outcome in three pivotal clinical trials of schizophrenia

BACKGROUND Dropout is a serious challenge to clinical trials in psychiatry, yet standard outcome analyses with mixed models do not account for dropout, while joint modeling uses dropout from a survival model to adjust the outcome from a mixed model, but is untested in clinical trials of schizophrenia. AIMS To compare mixed and joint modeling in three acute phase pivotal placebo controlled trials of schizophrenia. METHOD Data were reanalyzed on 611 in-patients with acute schizophrenia who participated in three pivotal randomized controlled trials that compared placebo with olanzapine or risperidone (dropout rates placebo: 62.6% and medication: 37.4%). The outcome measures were BPRS or PANSS total change scores. Mixed-effects models for repeated measures and joint models were computed and compared to examine the time-treatment interaction. Effect size comparisons were made. RESULTS Antipsychotic treatment was superior to placebo across analyses. Time treatment interactions were significant (p<.05) for the mixed (beta=2.33) and joint models (beta=2.62). Compared with mixed modeling, joint modeling reduced the estimated change score for treatment (21.24 vs 19.74) and placebo (1.64 vs -1.11). The effect size differences between placebo and treatment groups were greater for joint (ES=.89) than mixed modeling (ES=0.83). Sensitivity analysis replicated this trend of results in each of the three trials. CONCLUSION Compared to mixed modeling, joint modeling results in a greater separation between treatment and placebo groups. This offers preliminary evidence that joint modeling may be useful in the analysis of antipsychotic placebo controlled RCTs.

Continuous statistical models: With or without truncation parameters?

Lifetime data are usually assumed to stem from a continuous distribution supported on [0, b) for some b ≤ ∞. The continuity assumption implies that the support of the distribution does not have atom points, particularly not at 0. Accordingly, it seems reasonable that with an accurate measurement tool all data observations will be positive. This suggests that the true support may be truncated from the left. In this work we investigate the effects of adding a left truncation parameter to a continuous lifetime data statistical model. We consider two main settings: right truncation parametric models with possible left truncation, and exponential family models with possible left truncation. We analyze the performance of some optimal estimators constructed under the assumption of no left truncation when left truncation is present, and vice versa. We investigate both asymptotic and finite-sample behavior of the estimators. We show that when left truncation is not assumed but is, in fact present, the estimators have a constant bias term, and therefore will result in inaccurate and inefficient estimation. We also show that assuming left truncation where actually there is none, typically does not result in substantial inefficiency, and some estimators in this case are asymptotically unbiased and efficient.