Avoiding Bias and Standardizing Costs in Formulary Decisions

Fadia T. Shaya, PhD, MPH, Assistant Professor, University of Maryland School of Pharmacy, Baltimore, MD, USA

Now more than ever, managed care plans need to standardize their approach to appropriate benefit design, within a competitive landscape. Moreover, they need to streamline the tools they use in assessing the comparative cost-effectiveness of different therapy options.

Indeed, coverage decisions made by Pharmacy and Therapeutics committees affect the health of large populations. These decisions are evidence-based and require a standardized framework to present population-based specific data, timely and relevant information that is transparent both to formulary committees and to manufacturers. Managed care plans increasingly realize the value of their own data in guiding them through these decisions. These data have been historically generated and used for billing purposes and often used internally. Even so, their sensitivity is only at 85% [1]. They have also been used in auditing, and in the quality assurance process, in safety assessments and drug utilization reviews.

Managed care data have also served for a number of external uses, including epidemiologic and economic modeling, cost-effectiveness analysis and rate-setting in policy-making. They can also be used to monitor utilization by day of therapy, or to compare utilization trends for comparators by procedure and by diagnosis. These retrospective analyses of the data can complement information from clinical trials. Although they have the advantage of affording large sample size for analysis, and can be completed within relatively shorter time frames and with fewer resources than clinical trials, they do carry a large burden of bias and potential errors. It is very important for users of these data to understand the source of this bias and the options to correct for it. Indeed, major coverage decisions that can potentially affect millions of lives hinge on the interpretation of these data, hence the awesome responsibility to ensure data integrity and validity in the analysis.

 Three main types of bias have to be accounted for in the analysis of data. These include selection, confounding and information bias. Often researchers overlook bias, or fail to adjust for it, producing results that lack validity. In the following, we will discuss one example of bias under each type.

When analyses examine the prevalence and not the incidence cohort, there is a risk of prevalence bias, resulting from confusion in the association with prevalence related to the duration of the disease rather than to its incidence. In this example, given that only survivors receive the medication, the result is a positive association between drug use and outcome. The tradeoff is a possible underestimation of the rate of adverse outcomes.

Of all the sources of bias, confounding is probably the most common. A common source is confounding by indication, or by reason for prescription, whereby channeling of patients to different therapies occurs at the outset of medical encounters. Different persons with different risk factors that are linked to the clinical outcomes, may be more likely to be prescribed a given medication, thereby rendering the two drug cohorts intrinsically different.

Finally, an example of information bias is nondifferential misclassification bias, where errors occur randomly, and the degree of misclassification is similar for patients and independent of both exposure and health status conditions. This bias may lead to a decrease in the strength of the association between the drug and the outcome. It is especially bound to happen when the outcome variable is dichotomized to a yes/no variable, with no consideration to the length of exposure.

A number of methods adjust for different bias. To address selection bias, one could use random sampling, consider only incident cases, and minimize loss to follow-up. To remedy information bias, it is important to determine objective criteria for exposure. For example, it would be important to indicate whether exposure is continuous, or whether it does not matter. There are a number of methods used for minimizing confounding, including randomization, matching, stratification, and multivariate analysis. For the specific case of confounding by indication, propensity scores have been used.

The sources of bias highlighted above should be critically addressed in any data evaluation of effectiveness. In addition, there is much discussion in parallel, about the optimal way to assess costs. While we are still far from the standardization of an approach to costs, it is becoming increasingly urgent to develop such a template. Indeed, given the amplitude of decisions hinging on cost-effectiveness analyses, and the fact that these decisions often capitalize on the cost factor, it is critical that the components of costs be readily comparable across therapies.

 Such standardization would call for specific types of costs to be included in all analyses, perhaps with some allowance for slight variations only, by type of therapy. Specific categories of costs would be included, including for example disease-specific physician, pharmacy and hospital (inpatient and outpatient costs). Whether all other cost are to be included as well should also be determined by the template. On the other hand, the determination of whether or not indirect costs are relevant to the analysis, should not be driven by the availability -or lack thereof- of data, but rather by their relevance to the ultimate coverage decisions.

With the influx of tens of millions of lives into managed care, pursuant to Medicare Part D taking effect, it is more critical than ever to refine the tools used in comparative cost-effectiveness. The demographics of these new populations are very different from the typical profile of a managed care enrollee to which most of managed care cost-effectiveness assessments have applied, to date. In addition, while in the past the random managed care patient would stay in a given plan an average of 2.5 years, the new random elderly patient may or not be as transient. If we expect Medicare enrollees to show lower migration rates, than the relatively short time horizons used in most analyses will need to be re-considered. It is more urgent than ever for researchers as well as users of data to understand and learn how to control for bias, and standardize the approach to costs, given the huge implications of coverage decisions hinging on the results of analysis.

REFERENCES

1 Mouchawar J, et al. The sensitivity of Medicare billing claims data for monitoring mammography use by elderly women. Med Care Res Rev 61:116-27.

  Issues Index | 2006 Issues Index