Restrictive drug coverage policies can induce substantial drug exposure misclassification in pharmacoepidemiologic studies

Bias in pharmacoepidemiologic studies using secondary health care databases: a scoping review

BACKGROUND

The availability of clinical and therapeutic data drawn from medical records and administrative databases has entailed new opportunities for clinical and epidemiologic research. However, these databases present inherent limitations which may render them prone to new biases. We aimed to conduct a structured review of biases specific to observational clinical studies based on secondary databases, and to propose strategies for the mitigation of those biases.

METHODS

Scoping review of the scientific literature published during the period 2000-2018 through an automated search of MEDLINE, EMBASE and Web of Science, supplemented with manually cross-checking of reference lists. We included opinion essays, methodological reviews, analyses or simulation studies, as well as letters to the editor or retractions, the principal objective of which was to highlight the existence of some type of bias in pharmacoepidemiologic studies using secondary databases.

RESULTS

A total of 117 articles were included. An increasing trend in the number of publications concerning the potential limitations of secondary databases was observed over time and across medical research disciplines. Confounding was the most reported category of bias (63.2% of articles), followed by selection and measurement biases (47.0% and 46.2% respectively). Confounding by indication (32.5%), unmeasured/residual confounding (28.2%), outcome misclassification (28.2%) and "immortal time" bias (25.6%) were the subcategories most frequently mentioned.

CONCLUSIONS

Suboptimal use of secondary databases in pharmacoepidemiologic studies has introduced biases in the studies, which may have led to erroneous conclusions. Methods to mitigate biases are available and must be considered in the design, analysis and interpretation phases of studies using these data sources.

Denosumab and breast cancer risk in postmenopausal women: a population-based cohort study

BACKGROUND

Denosumab inhibits the receptor activator of nuclear factor κB (RANK) pathway and is used to treat osteoporosis. Emerging evidence suggests RANK-blockade may play a role in mammary tumourigenesis. Thus, we undertook a population-based study of denosumab use and breast cancer risk in a large cohort of postmenopausal women.

METHODS

We included women 67+ years with prior bisphosphonate use who filled a first prescription for denosumab. They were matched on age, date, cumulative prior use of and time since last use of a bisphosphonate to women with no history of denosumab. Cox proportional hazards was used to estimate the hazard ratio (HR) of breast cancer with denosumab use.

RESULTS

A total of 100,368 women were included in the analysis with 1271 incident breast cancer events. Denosumab use was associated with a 13% decreased breast cancer risk (HR = 0.87; 95% CI 0.76-1.00). There was no relationship between increasing number of denosumab doses and breast cancer risk (P-trend = 0.15).

CONCLUSION

These findings suggest a potential protective effect of ever denosumab use on breast cancer risk in a cohort of older women previously treated with bisphosphonates.

Low-Cost Generic Program Use by Medicare Beneficiaries: Implications for Medication Exposure Misclassification in Administrative Claims Data

BACKGROUND

Administrative claims data are used for a wide variety of research and quality assurance purposes; however, they are prone to medication exposure misclassification if medications are purchased without using an insurance benefit. Low-cost generic drug programs (LCGPs) offered at major chain pharmacies are a relatively new and sparsely investigated source of exposure misclassification. LCGP medications are often purchased out of pocket; thus, a pharmacy claim may never be submitted, and the exposure may go unobserved in claims data. As heavy users of medications, Medicare beneficiaries have much to gain from the affordable medications offered through LCGPs. This use may put them at increased risk of exposure misclassification in claims data. Many high-risk medications (HRMs) and medications tracked for adherence and utilization quality metrics are available through LCGPs, and exposure misclassification of these medications may impact the quality assurance efforts reliant on administrative claims data. Presently, there is little information regarding the use of these programs among a geriatric population.

OBJECTIVES

To (a) quantify the prevalence of LCGP users in a nationally representative population of Medicare beneficiaries; (b) compare clinical and demographic characteristics of LCGP users and nonusers; (c) assess determinants of LCGP use and medications acquired through these programs; and (d) analyze patterns of LCGP use during the years 2007-2012.

METHODS

This study relied on data from the Medical Expenditure Panel Survey (MEPS) from 2007 to 2012. The first 3 objectives were completed with a cohort of individuals in the most recent MEPS panel, while the fourth objective was completed with a separate cohort composed of individuals who participated in MEPS from 2007 to 2012. Inclusion in either study cohort required that individuals were Medicare beneficiaries aged 65 years or greater, used at least 1 prescription drug during their 2-year panel period, and participated in all 5 rounds of data collection during their panel period. MEPS captures medication utilization by surveying individuals on current and previous medication use and verifies this information at the pharmacy level, so prescription fills can be observed irrespective of payment by an insurer or a filed claim. Pharmaceutical utilization was assessed at the individual level for each year of the study period, and LCGP use was recorded as a binary variable for each individual. An LCGP medication fill was identified if the total cost of the drug was paid out of pocket and matched the cost of medications listed on LCGP formularies available from major pharmacy retailers during these years. Cohort demographics and characteristics of interest included age, gender, race, employment status, marital status, family income level, education level, residence in a metropolitan statistical area, geographic region, prescription drug coverage, Medicare type, comorbidities, number of unique medications used, and number of medication fills. Comparisons were made between users and nonusers using chi-square and t-tests. Multivariable logistic regression was used to identify factors associated with LCGP use.

RESULTS

From the most recent MEPS panel, 1,861 individuals were included in the study cohort, of which 53.5% were observed to be LCGP users. The 995 LCGP users in this cohort represented over 20 million Medicare beneficiaries who used LCGPs from 2011 to 2012. Significant differences between LCGP users and nonusers existed in terms of race, educational attainment, comorbidity burden, type of Medicare insurance, number of unique medications used, and number of medication fills. Each additional unique medication filled increased the odds of LCGP use by 12% (95% CI = 1.09-1.14). Individuals with insurance in addition to Medicare (i.e., Tricare/Veteran's Affairs or Medicaid) had less than half the odds of using LCGPs compared with those with Medicare or Medicare managed care insurance coverage only. The proportion of LCGP users and the proportion of LCGP fills out of all medications available through LCGPs increased from 2007 to 2012.

CONCLUSIONS

There is a high rate of LCGP use among Medicare beneficiaries aged 65 years or greater. Claims-based research and quality assurance programs focusing on the benefits and harms of medications available through these programs are at risk of underestimating the true medication exposure in this population and should account for this possibility in sensitivity analyses. Managed care organizations should incentivize the reporting of LCGP medication use or make adjustments to generic medication benefit structures to more effectively capture true medication exposure.

DISCLOSURES

No direct sources of funding were used to conduct this study. Data acquisition was supported by the University of Kentucky Center for Clinical and Translational Science through funding from NIH NCATS grant #UL1TR000117. Brown is the Humana-Pfizer Research Fellow at the Institute for Pharmaceutical Outcomes & Policy at the University of Kentucky College of Pharmacy and is provided salary from these corporations. However, neither company provided any direct funding for the current study nor provided any input or guidance for the design, methods, or drafting of the manuscript. Pauly has no financial disclosures or conflicts of interest. Portions of these results were presented at the 20th International Society for Pharmacoeconomics and Outcomes Research International Meeting; May 16-20, 2015; Philadelphia, Pennsylvania. Study concept and design were primarily contributed by Brown, along with the other authors. Brown took the lead in data collection and interpretation, along with Pauly and Talbert. All authors participated in the writing and revision of the manuscript.

Prevalence of Low-Cost Generic Program Use in a Nationally Representative Cohort of Privately Insured Adults

BACKGROUND

Administrative claims data are used for a wide variety of research and quality assurance purposes. Despite their utility, they are prone to medication exposure misclassification if medications are purchased without utilizing an insurance benefit. Low-cost generic programs (LCGPs) offered at major chain pharmacies are a relatively new and sparsely investigated source of exposure misclassification. Since they were implemented in 2006, LCGPs are now available at 8 of the 10 largest pharmacy chains and include a wide variety of medication classes. LCGP medications are often purchased out of pocket; thus, a pharmacy claim may never be submitted and exposure may go unobserved in claims data. There are little data regarding the utilization of these programs, and estimates of their use can provide important insights into the potential impact LCGPs may have on exposure misclassification in claims data.

OBJECTIVES

To (a) quantify the prevalence of LCGP users in a privately insured adult population, (b) assess patterns of LCGP use, and (c) compare clinical and demographic characteristics associated with LCGP users and nonusers.

METHODS

The study cohort consisted of 19,037 privately insured adults aged 18-64 who participated in the Medical Expenditure Panel Survey (MEPS) from 2007-2011. MEPS captures medication utilization at the pharmacy level, so prescription fills can be observed irrespective of a claim being filed. Pharmaceutical utilization was assessed at the individual level for each year of the study period, and LCGP use was recorded as a binary variable for each individual. An LCGP medication fill was identified if the total cost of the drug was paid out of pocket and matched the cost of medications listed on LCGP formularies available from Target, Walmart, CVS, or other major pharmacy retailers during these years. Cohort demographics and characteristics of interest included age, gender, race, employment status, marital status, family income, education level, residence in a metropolitan statistical area (MSA), prescription drug coverage, geographic region, comorbidities, and number of unique medications and medication fills. Comparisons were made between users and nonusers using chi-square and t-tests. Multivariable logistic regression was used to identify factors associated with LCGP use.

RESULTS

Out of the entire study cohort (N = 19,037), 6,921 (36.4%) individuals were identified as LCGP users, representing 34 million LCGP users annually. Users tended to be older, had higher Charlson Comorbidity Index scores, filled more prescriptions per person, and used more unique medications. Proportions of LCGP users and uses nearly doubled from 2007-2011, while total prescription utilization per person remained relatively stable. Over 10% of all prescription fills were filled through LCGPs. Of all LCGP fills, approximately 42% were for cardiovascular medications, 12% for antidiabetics, and 14% for levothyroxine. Greater than 30% of fills for antigout, metronidazole, angiotensin-converting enzyme inhibitors, levothyroxine, metformin, and diuretics were obtained through LCGPs, as were 18.9% of all warfarin fills. Compared with the reference category aged 18-34, adults aged 35-54 had an adjusted odds ratio (AOR) of being an LCGP user of 1.39 (95% CI = 1.29-1.50) and adults aged 55-64 had an AOR of 1.86 (95% CI = 1.70-2.04). Additionally, those with prescription drug coverage were nearly twice as likely to be LCGP users (AOR = 1.96; 95% CI = 1.64-2.35) compared with those without. Gender, income, comorbidity burden, region, year of panel entry, and number of unique medications also significantly predicted LCGP use.

CONCLUSIONS

There is a high rate of LCGP use in the privately insured adult population. Users of LCGPs tend to be older, have more chronic comorbidities, and use more medications than nonusers. Claims-based research and quality assurance programs focusing on the benefits and harms of medications available through these programs are at risk of greatly underestimating the true medication exposure in this population and should account for this in sensitivity analyses. Managed care organizations should incentivize the reporting of LCGP medication use or make adjustments to generic medication benefit structures to more effectively capture true medication exposure.

Confounding by drug formulary restriction in pharmacoepidemiologic research

PURPOSE

The potential consequences of confounding due to drug formulary restrictions in pharmacoepidemiologic research remain incompletely understood. Our objective was to illustrate this potential bias using the example of fluticasone/salmeterol combination therapy, an oral inhaler used to treat asthma and chronic obstructive pulmonary disease, whose use is restricted in the province of Quebec, Canada.

METHODS

We identified all new users of fluticasone/salmeterol in Quebec's administrative databases and classified those who received their initial dispensing of fluticasone/salmeterol between 1 September 1999 and 30 September 2003 as users from the liberal period and those who received it between 1 January 2004 and 31 October 2006 as users from the restricted period. The primary outcome was time to first hospitalization for respiratory causes within 12 months of cohort entry.

RESULTS

Our cohort included 72 154 new users from the liberal period and 5058 from the restricted period. Compared with use during the liberal period, use during the restricted period was associated with an increased rate of hospitalization for respiratory causes (crude hazard ratio [HR] = 1.41, 95% confidence interval [CI] = 1.32, 1.51). Subsequent adjustment for age, sex, and hospitalization for respiratory causes in the previous year attenuated the association (HR = 1.05, 95%CI = 0.98, 1.12). Further adjustment for other potential confounders resulted in a lower rate during the restricted period (HR = 0.78, 95%CI = 0.73, 0.83).

CONCLUSIONS

Formulary restrictions can result in substantial and unexpected confounding and should be considered during the design and analysis of pharmacoepidemiologic studies.

Restrictive reimbursement policies: bias implications for claims-based drug safety studies

Restrictive reimbursement policies-including those based on non-formulary drug status and prior authorizations-can create situations in which patients' use of prescription medications is not fully captured in administrative claims data. This can create bias in drug safety studies that depend solely on these data. An analysis in two Canadian provinces found that primary administrative databases captured only 61 % of dispensations of drugs for which restrictive reimbursement policies were in place. A subsequent simulation study found that, in certain circumstances bias due to exposure misclassification resulting from restrictive reimbursement policies can be quite large in analyses comparing outcomes between drug exposure groups. Investigators need to be knowledgeable about the data they analyze and know whether restrictive reimbursement policies are in place that might affect the capture of drugs of interest. It is also critical to understand the mechanisms by which restrictive reimbursement might cause bias in claims-based drug safety studies, the direction and magnitude of the potential bias, and strategies that could be used to mitigate such bias.

Uptake and characteristics of zoledronic acid and denosumab patients and physicians in Ontario, Canada: impact of drug formulary access

The addition of Limited Use criteria (less restrictive access) for zoledronic acid resulted in an immediate and significant increase in uptake and resulted in differences in patient/physician characteristics. In comparison, the uptake of denosumab (only listed with Limited Use) was rapid. Thus, formulary access restrictions have significant implications for prescribing.

INTRODUCTION

We sought to describe the use of zoledronic acid and denosumab by physicians and patients over time and examine the impact of a 2012 provincial formulary modification that removed the administrative burden on physicians when prescribing zoledronic acid.

METHODS

We identified users of zoledronic acid and denosumab using Ontario pharmacy claims data. The number of new patients and physicians was plotted and examined over time. Interrupted time series analysis examined the impact of a formulary modification to zoledronic acid use and prescribing. Descriptive characteristics of patients and prescribers were summarized pre- and post-formulary modification for zoledronic acid and overall for denosumab.

RESULTS

We identified 1463 zoledronic acid patients treated by 627 physicians and 16,736 denosumab patients treated by 2904 physicians. In the first 2 months on the market, we identified a rapid uptake of denosumab (>450 physicians and >1200 patients) in contrast to zoledronic acid (<10 physicians and <10 patients). Zoledronic acid use increased significantly in the 2-month post-formulary change, yet no change in denosumab was observed. Prior to the formulary modification, more zoledronic acid patients had a history of osteoporosis therapy (41 vs. 26%) or bone density testing (30 vs. 10%). Compared to zoledronic patients (post-formulary modification), more denosumab patients had prior osteoporosis therapy (55 vs. 26%), yet fewer had a gastrointestinal diagnosis (6 vs. 11%).

CONCLUSION

We identified a rapid uptake of denosumab in only 15 months of observation. A provincial formulary modification to zoledronic acid resulted in an increase in utilization and impacted patient characteristics.

Limited impact of drug exposure misclassification from non-benefit thiazolidinedione drug use on mortality and hospitalizations from Saskatchewan, Canada: a cohort study

PURPOSE

Our purpose was to measure the effect of non-benefit drug use on observed associations between exposure and outcome, thereby documenting an empirical example of the potential magnitude of biases introduced when exposure status is misclassified from a restrictive drug coverage policy.

METHODS

New users of antidiabetic agents were identified with a 1-year washout period between January 1, 1995, and December 31, 2005, in Saskatchewan, Canada, and were followed until December 31, 2008. Within this population-based cohort, persons were classified as users of benefit or non-benefit thiazolidinediones (TZDs) according to their first prescription record between January 1, 2006, and December 31, 2006 (non-benefit prescription records were not captured before 2006). An intention-to-treat approach was used to categorize TZD exposure over time. We evaluated the potential bias introduced by drug exposure misclassification by evaluating bootstrapped differences in hazard ratio (HR) estimates of all-cause hospitalization or death between users and nonusers of TZDs obtained from analyses that contained complete drug use (non-benefit and benefit drug use) versus benefit drug use only (non-benefit drug use was misclassified as unexposed). All analyses were replicated within the same cohort of new users of antidiabetic agents for clopidogrel and β-blocker (bisoprolol or carvedilol) users versus nonusers because these agents were also subject to exposure misclassification from non-benefit drug use during the period of the study.

FINDINGS

Among 27,333 new users of antidiabetic agents, we identified 5759 TZD users (28% non-benefit) and 21,574 nonusers of TZDs. The crude HR for hospitalization or death among TZD users versus nonusers was higher in a database that contained benefit-only prescriptions than in a database that contained all prescriptions (HR = 1.11 [95% CI, 1.05-1.18] vs HR = 0.99 [95% CI, 0.94-1.04]). However, the differences in HRs after adjustment for demographic characteristics, health care utilization, comorbidities, and medications suggested minimal bias was introduced when TZD exposure was misclassified in the benefit-only database (adjusted HR [aHR] = 1.04 [95% CI. 0.98-1.10] vs aHR = 0.99 [95% CI, 0.94-1.04]; bootstrapped aHR difference = +0.05 [95% CI, 0.02-0.08]). Minimal differences in aHRs were also observed within analyses of clopidogrel (1551 users [24% non-benefit]; bootstrapped aHR difference = +0.01 [95% CI, -0.04 to 0.06]) and β-blocker users (351 users [42% non-benefit]; bootstrapped aHR difference = +0.06 [95% CI, -0.09 to 0.20]) versus nonusers.

IMPLICATIONS

Although patient characteristics and outcomes differed between users of non-benefit and benefit drugs, misclassification of drug exposure did not meaningfully bias estimates of all-cause mortality and hospitalization after covariate adjustment in our study.

Evaluating the introduction of a computerized prior-authorization system on the completeness of drug exposure data

PURPOSE

Administrative databases that only capture records for benefit-approved prescriptions may underestimate exposure because they do not capture non-benefit prescriptions. Using a natural experiment, we illustrate the impact of automating a prior-authorization policy on the completeness of drug exposure.

METHODS

Using Saskatchewan (Canada) databases, weekly counts of benefit-approved and total prescription records in 2006 for new users of antidiabetic agents were examined across four categories: thiazolidinediones (TZDs), metformin, glyburide, and insulin. On July 1, 2006, Saskatchewan's public drug plan implemented an automated, online-adjudicated, prior-authorization process for TZDs; previously, prior approval was paper based. No such policy changes occurred for other drugs. We estimated the effect of this policy change on drug exposure using interrupted time-series analyses.

RESULTS

We examined 223 552 prescription records: 19% were for TZDs, 48% for metformin, 20% for glyburide, and 13% for insulin. Prior to automation, there were, on average, 571 benefit-approved TZD records per week; however, the number of benefit-approved TZD records increased immediately after the automated process was introduced by 240 prescriptions per week (95% CI 200-280, p < 0.001). The average proportion of TZD benefit-approved records was 73% before and increased to 93% immediately following policy change (20% absolute change, 95% CI 18.7-20.4%). No changes were observed for metformin, glyburide, or insulin (p > 0.1 for all).

CONCLUSIONS

Automating prior authorization for TZDs immediately increased the proportion of captured TZD records, suggesting in our study that one-fifth of TZD exposure was previously misclassified. If replicable, this indicates that even subtle changes in reimbursement policy may affect the validity of drug exposure data.