Factors in choosing atypical antipsychotics: Toward understanding the bases of physicians’ prescribing decisions

https://doi.org/10.1016/j.jpsychires.2005.06.004Get rights and content

Abstract

Objective

Off-label prescribing of medications, polypharmacy, and other questionable prescribing practices have led investigators to examine a large VA pharmacy database to determine if physician prescribing decisions appear reasonable.

Method

The current study addresses the question of physician prescribing of atypical antipsychotics in 34,925 veterans with schizophrenia, using a series of signal detection analyses.

Results

These results suggest that only three factors (hospital size, age, and secondary diagnosis) allow classification of patients prescribed atypicals into three groups with frequencies of use of atypicals ranging from 43% to 79%, and that these results are consistent with reasonable clinical practice.

Conclusions

Results of two-stage signal detection analyses are readily interpretable by clinicians and administrators who are faced with the task of evaluating how physicians prescribe medications in clinical practice. Physicians’ decisions to prescribe atypical antipsychotics are based on both patient and fiscal considerations. This likely reflects a combination of clinical judgment and institutional guidelines.

Introduction

Examination of prescribing decisions by physicians has become the focus of attention as drug costs continue to grow and this contributes to the increasing cost of healthcare in the US (Hoffman et al., 2004, Reinhardt et al., 2004). “Off label” use of medications is common in psychiatry because of the increased availability of psychotropic medications for clinicians to choose from and the lack of clinical trials that adequately address important factors relating to prescription choice (such as specific medication differences, dosing, side-effects, and compliance problems). Indeed, “off label” use of drugs may lead to unnecessary costs and potentially poor or delayed patient outcomes (Petersen, 2002).

A variety of treatment guidelines have been developed to treat schizophrenia. These guidelines provide expert advice on various treatment options that are presumed improve outcomes. Despite the existence of guidelines, the factors that actually predict physician’s decisions have not been explored. Is there a rationale for their prescribing patterns? How can this be determined using available data? To answer this, we review the methodological challenges in using available data, the limitations of using traditional approaches, and finally, we propose a two-stage approach to identifying clinically relevant factors in physician prescribing decisions.

The Veterans Health Administration of the Department of Veterans Affairs (VA) has information from the tens of thousands of patients treated with antipsychotics each year, which provides a valuable analytical resource to address issues regarding prescribing patterns. Using nationwide VA data for a three-month period in 1999, Leslie and Rosenheck (2001a) found that over half the VA outpatients diagnosed with schizophrenia received an atypical antipsychotic and that most prescribing was consistent with approved guidelines. In a second article, the authors examined these data for evidence that constriction of the VA budget had resulted in a reduction of the prescription of relatively expensive atypical neuroleptics (Leslie and Rosenheck, 2001b). In the later work they found, that hospitals that appeared to be under the most fiscal stress prescribed more of the expensive new medications.

As relevant to the VA as these studies are, they do not provide a comprehensive picture of prescribing patterns, in part because of two methodological challenges: (1) the difficulty integrating data that involve both individual-level and facility-level predictors of the physicians’ choice of medication, and (2) the limited availability of analytical techniques that provide easily interpretable and clinically relevant information. Consequently, there continues to be a need to examine such data in a way that can yield results relevant to decision-makers who must determine if patterns of use are clinically and economically reasonable or if they need to be altered.

A methodological challenge with the analysis of huge datasets containing both individual-level and facility-level data stems from the mixture of facility level factors, such as those related to costs of care with individual-level factors such as a patients’ diagnoses, ages, and durations of hospitalization. When assessing correlations for samples drawn from possibly different populations (or different VA facilities), computing the correlation by combining all the subjects together leads to statistical anomalies, termed Simpson’s Paradox (Bickel et al., 1975, Hand, 1979, Kraemer, 1978, Simpson, 1951, Wagner, 1982). Briefly, even if a given factor were completely uncorrelated with the outcome (prescription decisions) at every facility, it might appear to be correlated in the combined analysis, because those facilities with high rates of the outcome were also high or low on those factors (a across-facility correlation). Alternatively, if a given factor were perfectly positively correlated within each facility, it might appear to have no correlation because the cross-facility correlation was negative and cancelled it out. Thus, it is important to deal with across-facility correlation and within-facility correlation separately.

A second methodological challenge stems from the emphasis on statistical rather than clinical significance. Any non-random association will be found to be statistically significant provided the sample size is large enough. However, statistically significant results may not carry any clinical or practical significance (Borenstein, 1997, Borenstein, 1998, Jacobson and Truax, 1991, Kraemer, 1993, Wilkinson, 1999). Furthermore, the most commonly applied method to identify such associations, logistic regression, has a number of additional limitations (Breiman et al., 1984, Kiernan et al., 2001, Kraemer, 1992, Loh and Vanichsetakul, 1988). For example, assumptions of linearity must be made and such analyses are weak in detecting interactions, able only to detect those interactions the user already suspects and incorporates into the model. Finally, the output of logistic regression is at best a risk score that clinicians or administrators cannot subsequently use easily clinically.

In this paper we shall use an alternative signal detection method called Receiver Operating Characteristic (ROC) analyses (Kiernan et al., 2001, Kraemer, 1992). ROC is distinguished from the other partitioning methods by the fact that: (1) it uses a measure of clinical, not statistical, significance to determine splits and (2) it focuses on developing the most compact results.

This work makes use of the same VA pharmacy database analyzed by Leslie and Rosenheck, 2001a, Leslie and Rosenheck, 2001b to compare and contrast results found in the original logistic regression analyses. We shall be able to address questions such as which patient subgroups are prescribed atypicals and whether these patterns are clinically appropriate. The results from this study not only identify the most important variables for prediction, but also develop classification rules so that each patient could be assigned to one group with a specified probability of having the outcome.

We provide a two-stage approach to identify clinically relevant factors in physician prescribing decisions. In Stage 1 of the analyses we shall analyze facility-level data to identify which facilities are more or less likely to recommend an atypical antipsychotic medication to schizophrenics. These results are then used to stratify the facilities into more homogeneous subgroups. In Stage 2, we will use ROC analyses within each such subgroup, to provide an easily interpretable set of results in terms of patient subgroups associated with higher or lower prescription of atypical antipsychotics.

Section snippets

Database

The dataset used in these analyses was the same dataset of 34,925 patient records used by Leslie and Rosenheck in their 2001 study (Leslie and Rosenheck, 2001b) which originated in national VA administrative databases. These data include all VA outpatients who were diagnosed with schizophrenia during the 1999 fiscal year (from October 1, 1998 to September 30, 1999). Patients were identified as having a diagnosis of schizophrenia if they had two or more outpatient encounters in a mental-health

Stage 1: Facility-level analysis

As mentioned, only hospital size was a statistically significant predictor of atypical use for the facility-level analysis. Small hospitals (<200 patients) are more likely to prescribe atypical neuroleptics (63%) than Medium hospitals (58%) and Large hospitals (57%). These results are presented in Table 3.

Stage 2: ROC Individual-based analyses

In Stage 2, ROC analyses were completed for each of the three size categories of facility resulting in three different “decision trees”. Each decision tree (see “Big” hospitals in Fig. 1) has

Discussion

The result that patients who are younger (under 50 years) and more likely to require hospitalization as well as bipolar diagnosis are more likely to be prescribed atypical neuroleptics suggests that factors involving choice of these medications appears reasonable clinically. The fact that smaller hospitals are more likely to use atypicals independently of these three other factors is consistent with a lack of inpatient facilities at smaller centers and the need to treat aggressively on an

Results

In conclusion, people who are faced with the daunting task of interpreting vast amounts of medical data, such as the VA pharmacy and clinical data examined in this study, now have a series of tools that when applied to these data may result in more easily interpretable results than one might initially assume. The current analyses suggest that only three factors allow classification of patients into three groups with frequencies of use of atypicals ranging from 43% to 79%. Such results may be

Acknowledgements

This research was supported by the Sierra-Pacific Mental Illness Research, Education, and Clinical Center (MIRECC), National Institute of Aging Grant AG17824, and the Medical Research Service of the Department of Veterans Affairs.

References (23)

  • M. Borenstein

    Hypothesis testing and effect size estimation in clinical trials

    Annals of Allergy Asthma & Immunology

    (1997)
  • D.H. Solomon et al.

    Determinants of selective cyclooxygenase-2 inhibitor prescribing: are patient or physician characteristics more important?

    American Journal of Medicine

    (2003)
  • P.J. Bickel et al.

    Sex bias in graduate admissions: data from Berkeley

    Science

    (1975)
  • Borenstein M. The shift from significance testing to effect size estimation. In: Bellak AS, Hersen M, editors. Research...
  • L. Breiman et al.

    Classification and Regression Trees

    (1984)
  • D.J. Hand

    Psychiatric examples of Simpson’s paradox

    British Journal of Psychiatry

    (1979)
  • J.M. Hoffman et al.

    Projecting future drug expenditures–2004

    American Journal of Health-System Pharmacy

    (2004)
  • N.S. Jacobson et al.

    Clinical significance: a statistical approach to defining meaningful change in psychotherapy research

    Journal of Consulting and Clinical Psychology

    (1991)
  • M. Kiernan et al.

    Do logistic regression and signal detection identify different subgroups at risk? Implications for the design of tailored interventions

    Psychological Methods

    (2001)
  • J.D. Killen et al.

    Interactive effects of depression symptoms, nicotine dependence, and weight change on late smoking relapse

    Journal of Consulting and Clinical Psychology

    (1996)
  • H.C. Kraemer

    Individual and ecological correlation in a general context: investigation of testosterone and orgasmic frequency in the human male

    Behavioral Science

    (1978)
  • Cited by (0)

    The software used in this analysis is in the public domain (http://mirecc.stanford.edu).

    View full text