A systematic method for estimating individual responses to treatment with antipsychotics in CATIE

Abstract

Objective

In addition to comparing drug treatment groups, the wealth of genetic and clinical data collected in the Clinical Antipsychotic Trials of Intervention Effectiveness study offers tremendous opportunities to study individual differences in response to treatment with antipsychotics. A major challenge, however, is how to estimate the individual responses to treatments. For this purpose, we propose a systematic method that condenses all information collected during the trials in an optimal, empirical fashion.

Methods

Our method comprises three steps. First, we test how to best model treatment effects over time. Next, we screen many covariates to select those that will further improve the precision of the individual treatment effect estimates which, for example, improves power to detect predictors of individual treatment response. Third, Best Linear Unbiased Predictors (BLUPs) of the random effects are used to estimate for each individual a treatment effect based on the model empirically indicated to best fit the data. We illustrate our method for the Positive and Negative Syndrome Scale (PANSS).

Results

A model assuming it takes on average about 30 days for a treatment to exert an effect that will then remain about the same for the rest of the trial showed the best fit to the data. Of all screened covariates, only two improved the precision of the individual treatment effect estimates. Finally, correlations between drug effects and PANSS scales suggested that in CATIE it cannot be recommended to simply combine treatment effects across drugs (e.g. to study common drug mechanisms), but it is sensible to study how a given drug affects multiple symptom dimensions.

Conclusions

We demonstrate that treatment effects can be estimated in a way that condenses all information collected in an optimal, empirical fashion. We expect the proposed method to be valuable for other clinical outcomes in CATIE and potentially other clinical trials.

Introduction

Schizophrenia is an often devastating neuropsychiatric illness (Sullivan, 2005) and effective pharmacotherapy is therefore critical from a mental health perspective. The Clinical Antipsychotic Trials of Intervention Effectiveness (CATIE, (Lieberman et al., 2005, Stroup et al., 2003) is an important resource for studying treatment with antipsychotics. As the use of atypical medications has steadily grown, one of the goals of CATIE was to examine whether these expensive medications can reduce morbidity and hospital use and improve community functioning, as compared to the conventional or first generation antipsychotic drugs.

Complications exist in that typically only a proportion of the patients suffering from schizophrenia respond to a specific antipsychotic (Kane, 1999). Misdosing, medication error, drug–drug interactions, and concomitant illnesses can sometimes provide an explanation. However, drug (non)-response may also reflect individual differences in disease etiology and biological variables that influence drug response. The several weeks it may take a clinician to declare a treatment ineffective and recommend an alternative may leave the schizophrenic patient vulnerable to continuing social dysfunction and suicide (Meltzer and Okayli, 1995) or possible side effects of the drug. Any improvement in our ability to match individual patients to the most effective and safe medication would therefore have substantial clinical benefits.

In addition to comparing treatments on a group level, many opportunities exist in CATIE to study individual differences in treatment response. Most notably, CATIE participants with DSM-IV schizophrenia and available DNA have now been genotyped for ~ 492,000 single nucleotide polymorphisms (SNPs, which are the most commonly used genetic markers) using the Affymetrix 500 K genotyping platform plus a custom 164 K chip to improve genome-wide coverage (Sullivan et al., 2008). This study therefore constitutes a tremendous resource for studying genetic modifiers of individual response to antipsychotics. This is particularly so because the genotype and clinical data are available to the scientific community from the controlled-access repository of the National Institutes of Mental Health (www.nimhgenetics.org).

The shift from comparing treatment groups to using the wealth of clinical and genetic data to better understand individual differences in responses presents novel methodological challenges. A major challenge here is how to best estimate individual responses to treatment with antipsychotics. In the case of CATIE this challenge is accentuated by the complex study design. For example, instead of a simple controlled randomized design with a drug and placebo group, CATIE was designed to evaluate the effectiveness of antipsychotic drugs in typical settings and populations so that study results will be maximally useful in routine clinical situations. Other salient features include the extensive data collection at multiple time points for a period up to 18 months. Furthermore, patients can be switched up to two times to a different drug for reasons such as lack of efficacy or drug toxicity.

In this article we propose a systematic method to estimate individual changes in clinically relevant outcomes during drug treatment in a way that condenses all information collected during the trial in an optimal, empirical fashion. Although such changes may not necessary be the (exclusive) result of the pharmacological action of the drug (e.g. there may be placebo effects, artifact due to repeated testing with same instrument etc), for sake of simplicity we will refer to these changes as estimates of the individual treatment effect.

Our model based approach has several potential advantages compared to more standard methods for estimating treatment effects (e.g., calculating post-minus pre-treatment symptom scores). First, using all, versus only a subset, of the assessments will improve the precision of the estimates and consequently the statistical power to detect correlates of the individual treatment effects. The latter is, for example, important in genome-wide association studies (GWAS, where markers spanning the entire genome are tested for association with the outcome of interest) because genetic effects are likely to be small implying the need to maximize statistical power. Second, a wide variety of questions arise in complex trials such as CATIE including how to define treatment effects, how to handle the different treatment lengths, and what to do if only few observations are available. Our modeling approach provides a “natural” framework to address some of these issues. For example, there are multiple ways to define the treatment effect. Thus, instead of using the assessment obtained at the end of the treatment, for each patient we could also use the assessment after a predetermined period of treatment. The latter assessment will be more comparable across patients (e.g. in CATIE there is great variability in treatment duration), less confounded by the effects of being in the trial, and assuming that treatment effects will plateau will provide a good indication of the overall effect. However, rather than assume a priori which model of treatment response is more accurate, the method developed here enables empirical comparisons of various models and allows the selection of the treatment effect measure that is most consistent with the data. Finally, effects of potential confounders of treatment effects (e.g. changes as a result of repeated administration of the same diagnostic instrument) cannot be eliminated unless all data are considered. Such corrections may be important to further increase the precision of the treatment effect estimates.

Our method consists of first studying the best way to model treatment effects, then to screen many possible covariates to select those that will further improve the precision of the treatment effect estimates, and finally generate individual treatment effect estimates based on the best fitting model. We will demonstrate our method using the Positive and Negative Symptom Scale (PANSS, (Kay et al., 1987) that is a primary outcome in the CATIE study. In the final section we will study some of the properties of the estimated treatment effects and provide guidelines for subsequent analyses using these estimates. The method proposed in this article is generic and could potentially be valuable for other outcome variables in CATIE as well as other clinical trials.