Adherence to disease modifying anti-rheumatic drugs and the effects of exposure misclassification on the risk of hospital admission

Abstract

Objective

Describe the effect of different exposure classification strategies for disease modifying anti-rheumatic drugs (DMARDs) on drug-outcome associations.

Methods

We studied the association between DMARD initiation and all-cause hospitalizations in rheumatoid arthritis (RA) patients (1995-2005). Initiators of DMARDs and oral glucocorticoids were followed a maximum of 180 days. We compared two strategies for exposure classification: 1) A persistent exposure required (PER) approach, in which follow-up stopped when the regimen changed; and, 2) A persistent exposure ignored (PEI) approach, in which follow-up continued despite regimen changes. For PEI, adherence was assessed using the medication possession ratio. All-cause hospitalization risk was compared among RA regimen initiators using Cox models and methotrexate as the reference.

Results

We identified 28906 episodes of medication initiation. In PER analyses, TNFα-antagonists did not increase hospitalization risk compared with methotrexate, whereas leflunomide did (hazard ratio [HR]: 1.36, 95% CI: 1.1-1.67). Glucocorticoids increased hospitalization risk (HR: 1.29, 1.54 and 2.03 for low, medium and high doses, respectively). PEI results were similar to PER except that infliximab initiation increased the risk of hospitalization compared with methotrexate (HR: 1.46, 95% CI: 1.19-1.8), and most other effects were closer to the null. In PEI, adherence ranged from 73% for etanercept to 6% for glucocorticoids. Adherence to methotrexate was 59%.

Conclusions

Compared with methotrexate initiation, initiation of leflunomide or glucocorticoids consistently increased all-cause hospitalizations in the first 180 days of use. Most PER and PEI estimates were similar; observed differences in risk between these methods were likely due to differences in adherence.

Introduction

Although tumor necrosis factor alpha (TNFα) antagonists revolutionized the treatment of rheumatic diseases, their comparative safety relative to traditional disease modifying anti-rheumatic drugs(DMARDs) has been questioned.(1) Placebo-controlled trials indicated that TNFα-antagonists increased infections among rheumatoid arthritis(RA) patients.(1) Observational studies indicated that TNFα-antagonists increased the risk of infections and heart failure compared with methotrexate in RA patients.(2-4)

Most medication effects are biologically plausible during the actual exposure to the medication and sometimes during a period after medication exposure has ended. Thus, one approach to measuring medication exposure using administrative databases is to restrict the exposure person-time to the period with medication supply available (i.e. persistent exposure required approach, PER).

In an alternate approach, initiation of medication can by itself be considered an exposure of interest. This initiation identifies the exposure category in which the patient remains throughout follow-up. This persistent exposure ignored(PEI) approach is similar to the intention-to-treat analysis of randomized trials, except that exposures are not assigned through randomization.

In assessing the relative safety of TNFα-antagonists in an observational setting, we recognized that DMARDs were often discontinued.(5) We hypothesized that different exposure classification approaches could lead to conflicting results. We thus compared these approaches during our assessment of effects of RA medications on the risk of hospital admissions.

Methods

Study Cohort

We assembled a retrospective cohort of RA patients enrolled in TennCare, Tennessee's Medicaid managed-care program. Enrollees met our RA definition with ≥1 RA coded diagnosis and initiation of a DMARD or oral glucocorticoid regimen, defined as a filled prescription for a study medication after 180 days without exposure to that specific medication.

Initiation of medication determined the beginning of follow-up for each episode (t₀). On t₀, cohort members were aged ≥18 years, had continuous TennCare enrollment during the previous 180 days(baseline) and had baseline use of pharmacy benefits. Since serious medical conditions could affect follow-up and outcomes unrelated to medication exposure, we excluded patients with solid organ transplantation, HIV/AIDS, cancer, and serious kidney, liver or respiratory diseases, identified at baseline. We also excluded patients who had ≥2 baseline healthcare encounters for juvenile rheumatoid arthritis, systemic lupus erythematosus, Crohn's disease or ulcerative colitis.

For each episode, follow-up continued from t₀ through the first of date of death, end of study(December 31^st, 2005), date of hospitalization or the 180^th day of follow-up. Episodes were truncated because previous studies indicated that infection risk following initiation of TNFα-antagonists was time-dependent,(2, 6) and that study medications use was intermittent, making the long term classification of exposure challenging.(5)

Study medications

New episodes of DMARD or glucocorticoid use were identified applying an algorithm that maximized identification of newer and less frequently used study DMARDs. Patients could contribute ≥1non-overlapping DMARD episodes, using the following hierarchy: TNFα-antagonists(etanercept, infliximab and adalimumab), leflunomide, hydroxychloroquine, sulfasalazine, methotrexate and glucocorticoids. Concurrent initiation of a DMARD and glucocorticoid was considered DMARD initiation. Glucocorticoids were further classified based on initiating dose in milligrams: <7.5(low), 7.5-30(medium) or >30(high) of prednisone equivalents.(7)

Exposure measurements

We compared two approaches for exposure classification. For the PER approach, we classified each person-day of follow-up according to the medication days-supply dispensed.(8) Exposure started with the first filled prescription and continued through medication discontinuation(defined as 14 days without drug supply) or the change of the initial regimen (either by medication addition or switch to another regimen). For TNFα-antagonist users, co-therapy with methotrexate was allowed; however, addition of a different TNFα-antagonist or DMARD other than methotrexate ended the episode. For other DMARDs, initiation of any new DMARD including TNFα-antagonists ended the episode. Initiation of oral glucocorticoids would not end a DMARD episode but initiation of a DMARD ended a glucocorticoid episode. Thus, exposure person-time included all person-days of continuous exposure after initiation of treatment including gaps ≤14 person-days.

For the PEI approach, within our defined episodes of use, we included all person-days, regardless of addition of or switching to other regimens or discontinuation of the initial regimen. To assess exposure misclassification in the PEI approach we measured adherence using medication possession ratios(MPR) expressed as percent of person-time exposed to the initial regimen during the episodes. Since the occurrence of an outcome or censoring event interferes with this assessment, we only calculated MPR for those episodes with 180 person-days of available follow-up.

Outcomes

The study outcome was the first all-cause hospitalization, identified using TennCare inpatient claim files. We considered the first listed discharge diagnosis as the primary reason for hospitalization.

Potential confounders

Covariates were measured during the 180-day baseline preceding each episode, including demographics; markers of comorbidity: number of hospitalizations, outpatient and emergency room visits, enrollment on disability, number of different medication classes filled; surrogate markers of disease severity: extraarticular manifestations of disease, number of intra-articular and orthopedic procedures, number of tests ordered for inflammatory markers; and days of drug supply for other DMARDs, oral glucocorticoids, NSAIDs and narcotics.(2, 6, 9)

Other risk factors assessed included previous hospitalization due to infection, chronic obstructive pulmonary disease(COPD), diabetes and cardiovascular diseases and baseline prescriptions for antibiotics, anticonvulsants, antipsychotics, antidepressants, lipid-lowering agents, gastroprotective therapies, anti-arrhythmics, anticoagulants, replacement estrogens and oral contraceptives.

Among DMARD initiators, the average daily dose of oral glucocorticoids used on t₀ was categorized using ranges previously described.(7) One patient could contribute ≥1 episode provided the new-user definition was fulfilled. For these episodes, a new baseline period was defined and a new set of covariates measured.

Statistical analysis

The units of analysis were the new episodes of medication use and Cox proportional hazard regressions assessed the effect of medication exposure on the risk of hospital admission. Since patients could contribute one or more episodes of new use (with an updated set of covariates), we accounted for this clustering using patient's study numbers to define clusters and accounted for this additional intra-group correlation using the Huber–White ‘sandwich’ variance estimator and calculated robust standard errors for all estimates.(10) The proportional hazards assumption was verified using Schoenfeld residuals and log-log plots of survival functions. Models that used exposure propensity scores to summarize covariate information yielded almost identical results to those reported in this article and did not affect our conclusions. This study was approved by the Vanderbilt University IRB and by the Bureau of TennCare. Statistical analyses were performed using Stata 10.1.

Results

Cohort description

Overall, 14586 RA patients contributed 28906 new episodes of medication use. Their median age was 55 years (interquartile range (IQR): 45–64) and 76% were female. Most patients were white (83%) and lived in major metropolitan areas (56%). Approximately 20% had orthopedic procedures performed during baseline. Manifestations of extra-articular disease were recorded in 1% of patients, 11% had diabetes, 14% had COPD and 6% had an infection during baseline.

At baseline, 63% of patients were enrolled in a disability TennCare category and 2% were nursing home residents. During baseline, 39% had an emergency department visit and 22% had been hospitalized. The median number of different medications per patient was 13(IQR: 8-18). The median days supply exposed to glucocorticoids and narcotics was 30(IQR: 0-128) and 30(IQR: 0-137), respectively.

There were 6,665 hospitalizations identified. Cardiovascular, respiratory, musculoskeletal and gastrointestinal diseases(recorded as principal discharge diagnoses) accounted for 67% of all-cause hospitalizations.

PER approach

The PER approach yielded 2,383 hospitalizations and 3,615 person-years of follow-up or 659 hospitalizations per 1000 person-years. Compared with initiation of methotrexate, initiation of TNFα-antagonists, sulfasalazine or hydroxychloroquine did not increase the risk of hospitalizations. Initiation of leflunomide increased this risk by 36% compared to methotrexate. There was a dose-response increase in hospitalization risk among glucocorticoid initiators (HR range: 1.29 through 2.03), compared with methotrexate(Table). Baseline use of glucocorticoids was also associated with an increased risk of hospitalizations (HR: 1.18 (p=0.009) and 1.16 (p=0.254) for medium and high doses respectively), compared with no use or use of low doses at baseline.

Table. DMARD/Glucocorticoids initiation and the risk of hospitalization, Rheumatoid arthritis cohort, TennCare 1995-2005.

PERSISTENT EXPOSURE REQUIRED (PER)			Crude Analysis		Adjusted Analysis*
	Events	Rate / 1000 person-years	HR	95% CI	HR	95% CI
Methotrexate (MTX)	460	494.9	1.00	Reference	1.00	Reference
Etanercept +/- MTX	126	464.1	0.95	(0.78 - 1.16)	1.12	(0.90 - 1.4)
Infliximab +/- MTX	45	489.2	1.00	(0.74 - 1.35)	1.21	(0.88 - 1.67)
Adalimumab +/- MTX	55	489.6	1.00	(0.76 - 1.32)	1.09	(0.81 - 1.48)
Leflunomide	126	654.1	1.32	(1.08 - 1.6)	1.36	(1.10 - 1.67)
Sulfasalazine	98	545.1	1.05	(0.85 - 1.31)	1.16	(0.93 - 1.44)
Hydroxychloroquine	343	492.8	0.99	(0.86 - 1.14)	1.03	(0.90 - 1.19)
Oral Glucocorticoids (low)	171	756.2	1.41	(1.18 - 1.69)	1.29	(1.07 - 1.56)
Oral Glucocorticoids (medium)	687	968.4	1.73	(1.52 - 1.97)	1.54	(1.34 - 1.77)
Oral Glucocorticoids (high)	272	1322.2	2.32	(1.98 - 2.73)	2.03	(1.72 - 2.41)
PERSISTENT EXPOSURE IGNORED (PEI)			Crude Analysis		Adjusted Analysis
	Events	Rate / 1000 person-years	HR	95% CI	HR	95% CI
Methotrexate (MTX)	899	490.4	1.00	Reference	1.00	Reference
Etanercept +/- MTX	228	435.2	0.89	(0.77 - 1.03)	1.03	(0.88 - 1.21)
Infliximab +/- MTX	101	600.6	1.22	(0.99 - 1.49)	1.46	(1.19 - 1.8)
Adalimumab +/- MTX	110	515.5	1.04	(0.86 - 1.27)	1.17	(0.95 - 1.43)
Leflunomide	252	549.7	1.12	(0.98 - 1.29)	1.22	(1.05 - 1.41)
Sulfasalazine	249	454.3	0.93	(0.81 - 1.07)	1.06	(0.92 - 1.22)
Hydroxychloroquine	657	457.8	0.93	(0.84 - 1.03)	0.99	(0.89 - 1.09)
Oral Glucocorticoids (low)	487	573.2	1.17	(1.05 - 1.30)	1.12	(1.00 - 1.26)
Oral Glucocorticoids (medium)	2,766	620.0	1.26	(1.17 - 1.36)	1.21	(1.11 - 1.31)
Oral Glucocorticoids (high)	916	698.3	1.41	(1.29 - 1.55)	1.33	(1.20 - 1.47)

Footnote: HR, hazard ratio; PER, persistent exposure required; PEI, persistent exposure ignored; low dose (<7.5 mg), medium dose (7.5-30 mg) and high dose (>30 mg of prednisone equivalents)

^*Covariates for adjustment were measured during the 180-days baseline and included: age, gender, race, place of residence, nursing home residency, calendar year of the episode, number of hospitalizations, outpatient and emergency room visits, enrollment on disability, number of different medication classes used, extraarticular manifestations of disease, number of intra-articular and orthopedic procedures, number of tests ordered for inflammatory markers; and days of drug supply for other DMARDs, oral glucocorticoids, NSAIDs and narcotics, previous hospitalization due to infection, chronic obstructive pulmonary disease (COPD), diabetes and cardiovascular diseases and baseline prescriptions for antibiotics, anticonvulsants, antipsychotics, antidepressants, lipid-lowering agents, gastroprotective therapies, anti-arrhythmics, anticoagulants, replacement estrogens and oral contraceptives. Models included 48 variables (or 72 parameters) and 2383 and 6665 events for our PER and PEI models, respectively.

PEI approach

Using the PEI approach, 6,665 hospitalizations were identified during 11,802 person-years of follow-up, or 565 per 1000 person-years. Compared with initiation of methotrexate, initiation of etanercept, adalimumab, sulfasalazine or hydroxychloroquine did not increase the risk of hospitalizations. Initiation of leflunomide increased this risk by 22% compared with methotrexate. There was a dose-response increase in hospitalization risk among glucocorticoid initiators (HR range: 1.12 through 1.33). In the PEI approach, initiation of infliximab also increased the risk of hospitalizations by 46%, compared with methotrexate(Table and Figure). In this approach, baseline use of glucocorticoids was also associated with an increased risk of hospitalizations (HR: 1.12 (p=0.015) and 1.18 (p=0.074) for medium and high doses, respectively), compared with no use or use of low doses at baseline.

Figure. Adherence to RA regimens and the risk of hospitalization following initiation of DMARDs or glucocorticoids, Rheumatoid arthritis cohort, TennCare 1995-2005.

Footnote: MPR, medication possesion ratio (median and interquartile range), HYD, hydroxychloroquine; GC, glucocorticoids; PER, persistent exposure required; PEI, persistent exposure ignored; low dose (<7.5 mg), medium dose (7.5-30 mg) and high dose (>30 mg of prednisone equivalents). Models included 48 variables (or 72 parameters) and 2383 and 6665 events for our PER and PEI models, respectively.

Adherence to initial regimens (PEI)

Median MPR for TNFα-antagonists ranged from 73% for etanercept to 68% for infliximab. Median MPR was 69% for leflunomide and 59% for methotrexate, the reference exposure. Median MPR was 49% and 33% for hydroxychloroquine and sulfasalazine, respectively. Glucocorticoid initiators had the lowest median MPR (6% overall, range 5%-16% for medium and low doses)(Figure).

Discussion

In this large cohort of RA patients, leflunomide and glucocorticoids initiation increased the risk of all-cause hospitalizations compared with methotrexate initiation. During short periods of follow-up, PER and PEI led to similar conclusions for most exposure effects. However, differences in some results were notable, likely related to differential adherence to the initial regimens. These differences might be even greater in studies with longer follow-up.

The two approaches for exposure measurement differed in their complexity. The PER approach applied a strict day-by-day classification of exposure and allowed short gaps that were also considered exposed person-time. The simpler PEI approach classified exposure according to the initial regimen, disregarding subsequent changes. Because of more follow-up and events, confidence intervals were narrower in PEI than in PER. However, PEI hazard ratios were consistently lower than PER estimates for glucocorticoids and leflunomide. The attenuation of an adverse effect with the PEI approach would be expected when exposure misclassification increases, provided the association between exposure and outcome is real. Thus, the decrease in risk associated with lower adherence to leflunomide and glucocorticoids suggests that these drugs do increase hospitalization risk.

Although the PER approach minimized exposure misclassification,(8) this approach over-represents patients with good adherence to therapies. Factors such as intention to comply with clinicians' directions, favorable experience in terms of effectiveness and safety, and patients' own decision to continue on therapy are determinants of adherence. The effects of these selection factors might be magnified with longer durations of follow-up. However, these factors are difficult to ascertain and account for in administrative databases. The interpretation of results from PER approaches must consider these selection issues.

The PEI represents an alternative to the PER approach. Changes to the initial regimen are disregarded, reducing the possibility for selection bias. In placebo-controlled randomized trials, suboptimal adherence to the study medication attenuates the medication effects and biases the results towards the null. Outside of those experimental settings, adherence to medications may be lower and use of other co-therapies may be greater.

In our cohort, adherence to the initial regimen varied widely. Thus, adherence to both the exposures of interest and the non-placebo reference group must be considered when interpreting PEI results. The low adherence to glucocorticoid regimens suggests that these drugs were actually prescribed for intermittent use. Contrary to the PER analysis, infliximab initiation increased the risk of hospitalizations, compared with initiation of methotrexate in the PEI analysis. Discontinuation of infliximab was sometimes followed by initiation of other DMARDs, glucocorticoids or continuation of glucocorticoid therapies(data not shown), which could increase the risk of hospitalization without regard to infliximab past exposure. Although adherence was lower in the reference methotrexate group than among infliximab initiators, rates of hospitalization were similar for methotrexate users in both analyses, while they increased for infliximab users in the PEI compared to the PER analysis(Table).

Our findings have implications for studies of RA patients with severe disease who discontinued their TNFα-antagonists due to lack of effectiveness or intolerance. This group of patients might require more aggressive or alternate treatments that could increase their risk of study outcomes. This would be one explanation for the increased risk for hospitalizations observed for infliximab in the PEI approach compared with the PER strategy. However, although the PER approach considered short gaps after medication supply exhaustion as exposed person-time, hospitalizations following discontinuation of treatment may be residual effects of the medication that were missed by the PER approach (4).

Longer periods of follow-up will increase the possibility of suboptimal adherence and exposure misclassification. These changes will be missed by the PEI approach. PER will also be affected by selection issues increasing as follow-up expands and patients with good adherence enrich the cohort with a subpopulation that is likely to differ from those who discontinued treatment early.

Discontinuation and lack of adherence to medication regimens are challenges for assessments of comparative effectiveness of medications. A medication proven to be highly efficacious in clinical trials might have little real-life effectiveness due to low adherence. In this scenario, assuming no effect of concurrent medications, the PEI approach would provide more useful effectiveness information than the PER approach, and could complement efficacy information. Nevertheless, use of concurrent medications is frequent and should be considered.

Our strategies attempted to determine the effects of individual treatments and to maximize sample size for examination of TNF-α antagonists and other less frequently used DMARDs. Although different DMARDs and glucocorticoids were often used in combination, many patients in our population used glucocorticoids alone, and we had sufficient sample size to examine this group separately. Since co-therapy with glucocorticoids was common in patients using DMARDs, we examined DMARD use controlling for baseline use of glucocorticoids. However, initiation of glucocorticoids during a DMARD episode would not have been captured by our methods, and would result in misclassification of the episode as DMARD alone, rather than DMARD plus glucocorticoids. This would have made it more difficult to demonstrate differences between the DMARD and glucocorticoid groups. Our results suggest that baseline use of glucocorticoids was likely a predictor of subsequent use, and that baseline glucocorticoids use among DMARD initiators was associated with hospitalization, as were glucocorticoids, when initiated alone.

For assessment of medication safety, for which clinical trial information is usually limited, misclassification introduced by the PEI approach in a low adherence setting, could bias the results towards the null, potentially obscuring danger signals of interest. This is the likely explanation for the different results for glucocorticoids in our study, where the PEI estimates were closer to the null than PER estimates. Regimen changes are problematic in this setting because their effects will be ignored and attributed to the initial regimen. Similar to confounding by disease severity, the probability of switching or adding medications to an ongoing regimen is likely differential with regard to the initial regimen.

Frequent discontinuation of medications complicates the assessment of comparative safety of RA medications. Similar challenges will be faced in the forthcoming assessments of comparative effectiveness. Although most PER and PEI estimates were similar, some findings conflicted. The optimal approach to deal with these complexities should be based on a thorough assessment of medication utilization patterns.