A comparison of diabetes medication adherence and healthcare costs in patients using mail order pharmacy and retail pharmacy

Abstract

Objective: To compare long-term diabetes medication adherence and healthcare costs in patients using mail order pharmacy versus retail pharmacy.

Methods: The MarketScan database was used to identify patients who filled prescriptions for oral anti-diabetes medications in a retail pharmacy for at least 6 months before switching to mail order pharmacy for at least 12 months. These patients were matched to others who used retail pharmacy continuously for at least 18 months. A propensity score was used to create matched groups of patients comparable on probability of switching to mail order, weighted Poisson regression was used to analyze differences in medication adherence, and Tobit regression was used to compare costs.

Results: A total of 14,600 patients who switched to mail order were matched to 43,800 patients who used retail pharmacy continuously. The average adjusted adherence in retail pharmacy was 63.4% compared to 84.8% after switching to mail order. Per-member-per-month total healthcare and total medical costs were on average $34.32 and $37.54 lower in the mail order group, respectively. Diabetes-related medical costs were on average $19.14 lower in the mail order group, while pharmacy costs were $14.13 higher.

Limitations: Limitations include a patient population under the age of 65, no information on pharmacy benefit design, and limited follow-up time relative to that necessary to identify long-term diabetes complications.

Conclusions: After adjusting for measured confounders of medication adherence and disease severity, individuals who switched to mail order pharmacy had higher medication possession ratios and trended toward lower total and diabetes-related medical costs over time.

• adherence
• healthcare costs
• mail order pharmacy

Introduction

Although there is substantial evidence to support the theory that poor medication adherence rates are associated with increased healthcare service utilization and higher overall medical costs[1–8], there are few reports which indicate that improvements in a patient's medication adherence will result in a decrease in long-term medical service utilization and costs. The reasons for this lack of evidence are largely unknown, but may be the result of the challenge of adjusting for fundamental differences between patient populations who choose to participate in adherence improvement efforts and those who do not accounting for the impact of disease progression, or lack of data on long-term intervention effectiveness.

Most analyses of medication adherence interventions measure adherence for a short period of time and stop measurement after the intervention has been completed[9]. Studies that have measured long-term adherence to therapy suggest that the effects of adherence programs diminish over time[10],[11]. Although published reports of increased adherence rates among specific therapies dispensed through mail order pharmacies provide support that this strategy may be effective[12],[13], there is little if any evidence that this strategy is beneficial in maintaining increased adherence and reducing medical utilization and costs for an extended period of time. In addition, previous research has not provided evidence that patients who switch from retail pharmacy to mail order pharmacy have increased medication adherence.

The objective of this study was to determine if there is a difference in long-term medication adherence in patients on diabetes medications who switched to mail order pharmacy and to determine if patients who made this switch have lower long-term healthcare costs. It was hypothesized that once patients switch to mail order pharmacy, they have better diabetes medication adherence than patients who continue to receive their medication via a retail pharmacy, that better medication adherence persists over time, and that better medication adherence will result in lower long-term total and diabetes-related healthcare costs.

Patients and methods

Patients

This retrospective claims analysis utilized data from the MarketScan Commercial Claims and Encounters (MCCE) database from Thomson Reuters for the period of January 1, 2005 to December 31, 2007. These data included health insurance claims across the continuum of care (e.g., inpatient, outpatient, outpatient pharmacy) as well as enrollment data from large employers and health plans across the United States who provide private healthcare coverage for more than 45 million employees, their spouses and dependents. This administrative claims database includes a variety of fee-for-service, preferred provider organizations, and capitated health plans. Medicaid and Medicare recipients are not included in this commercial claims database.

Figure 1 describes the formation of the two patient cohorts. The eligible study population included patients 18 through 64 years of age with a primary or secondary diagnosis of type 2 diabetes mellitus (International Classification of Diseases, Ninth Revision [ICD–9] codes 250.xx (excluding 250.x1), 357.2, 362.0x, 366.41, and 648.0). Patients were required to be continuously enrolled in a pharmacy benefit plan for no less than 18 months. To limit the number of patients with difficult to control diabetes, individuals were excluded from the study if they had a diagnosis of type 1 diabetes mellitus (ICD-9 codes 250.x1), if they filled multiple oral antidiabetic agents during the first 6 months of retail follow-up or if they filled a prescription for insulin during the follow-up period[14].

Figure 1. Cohort formation process for population of patients on oral diabetes medications in the MarketScan database between January 1, 2005 and December 31, 2007.

The earliest date on which to establish retail pharmacy use was January 1, 2005 through at least June 30, 2005, and the earliest possible follow-up period in which to establish a change to mail order pharmacy was between July 1, 2005 and at least June 30, 2006. The index date, t₁, was the date that an eligible patient received a prescription for a single oral diabetes medication (Medi-Span Generic Product Indicator [GPI] code beginning 27–15 through 27–99) from a retail pharmacy. Patients whose first oral diabetes medication was filled in mail order were excluded. Next, patients were required to have retail pharmacy claims spanning no less than 6 months after the index date, t₁. This time frame was chosen to provide a minimum baseline interval to calculate medication adherence and to eliminate patients who had pharmacy benefit programs that required the use of mail order pharmacy. These mandatory mail order programs generally require a conversion to mail after two retail prescriptions.

At this point, the mail order- and retail-only cohorts diverge. In the mail order cohort, individuals who switched their index oral diabetes medication from a retail pharmacy to a mail order pharmacy were identified. The date of the first mail order pharmacy fill was considered the index switch date, t_2m. Patients were then required to have mail order pharmacy claims that spanned no less than 12 months, t_2m to t_3m. Patients in the mail order cohort could also have up to two consecutive retail fills (i.e., courtesy refills) at a retail pharmacy during t_2m to t_3m and still be considered part of the mail order cohort. A minimum 12-month period, t_2m through t_3m, was selected as it would provide an adequate number of refill opportunities to judge a patient's medication adherence in mail order. After time point t_3m, a patient remained in the mail order cohort (regardless of the pharmacy channel they received the index diabetes medication) through the end of therapy, t_4m, of their index diabetes medication. There were three scenarios that could constitute the end of therapy for the index diabetes medication: (1) a gap of 45 days or more in the index diabetes medication, (2) the end of insurance eligibility, or (3) the final medication order for the index drug. Patients could have dose changes or medication changes within a drug category and still be considered to be taking the index diabetes medication.

In the retail-only cohort, there was not a similar date to t_2m, the index switch date, because no mail order pharmacy use occurred for the index diabetes medication. To create a comparable date, the length of time between t_1m, the date of the first claim of the index diabetes medication and t_3m, the minimum amount of time in mail order after the switch was determined. Individuals from the retail-only cohort were randomly matched with a greedy matching algorithm[15] at a ratio of 3:1 when their total length of follow-up, t_1r to t_4r, was greater than or equal to t_1m to t_3m. After individuals were matched, a pseudo-index switch date, t_2r, was created for each matched set using the following formula:This created a date that had the same number of calendar days from start to switch as the mail order group. Thus, t_2r, was used as the comparable date to the index switch date, t_2m. Time point t_3r was determined using the following formula:After time point t_3r, a patient remained in the retail cohort through the end of therapy with their index diabetes medication, t_4r. End of therapy was determined as described in the preceding paragraph.

Methods

To identify patients with a similar likelihood of making a switch to mail order pharmacy, a propensity score for the predicted probability to switch to mail order pharmacy was created. The primary goal of the propensity score was to create balance on the potential confounders between treatment groups for the exposure/outcome relationship. The general procedures outlined by Perkins et al. were followed in constructing the model[16]. First, the propensity score was estimated using logistic regression analysis for each subject's probability to use mail order pharmacy (the dependent variable) based on the values of the observed covariates. Measured variables related to the exposure and outcome of interest were included based upon considerations of propensity score model variable selection methodologies[16–18].

The independent variables used in developing the propensity score included patient age at the time of the index diabetes claim measured on a continuous scale, and geography categorized into US census regions according to the state of residence, total number of prescription medication orders, member out of pocket expenses[3],[19], a record of depression (ICD–9 codes 296.2x, and 296.3x)[20], general disease co-morbidity as measured by the Charlson Co-morbidity Index (CCI) score[21], the use of other chronic medications to treat hyperlipidemia (GPI 39-XX), hypertension (GPI 36-XX) or depression (GPI 58-XX), index oral diabetes medication type (GPI codes beginning 27–15 through 27–99), and the Diabetes Complications and Severity Index (DCSI)[22]. The MCCE does not contain results from laboratory tests, thus a modified DCSI score was computed without the input of laboratory values.

The propensity score model was built using step-wise regression techniques using a p-value of less than 0.2 for a variable to enter the model and a p-value of less than 0.05 to be retained in the model. Second- and third-order interaction terms were evaluated through the step-wise selection process. After the initial model was specified, individuals were stratified by deciles of the distribution of the estimated propensity scores. The degree to which balance was achieved was determined by comparing the covariate values across each exposure group for each subclass and covariate. The balance between variables was evaluated by determining pooled Cochran–Mantel–Hanzel statistics for categorical variables and F-statistics for comparison of covariate means before and after stratification. When balance was not achieved (i.e., p-value for difference between two means of <0.5), the model was re-specified to improve the balance.

Next, the distribution of propensity scores was plotted for each exposure group and the degree to which they overlapped was determined. The greater the overlap, the more comparable the populations are with respect to the underlying covariate values. If the degree to which the two populations’ propensity scores overlapped was large, increased balance between the populations was achieved. If overlap was poor, the model was re-specified to improve performance.

The primary outcomes of interest were medication adherence and healthcare costs. The index switch date, t_2m and pseudo index switch date, t_2r, was the point in time that designated the pre-switch period of retail claims and the post-switch period of claims for the calculations of medication adherence and cost values. A medication possession ratio (MPR) was calculated for the time periods t₁ to t₂ and t₂ to t₄ for both the mail order and retail cohorts. The MPR was calculated by dividing the number of days a patient was in possession of the medication after the beginning of the period, by the total number of days in the period plus the days supply between two pharmacy claims for a medication. For example, if a patient had sufficient quantity of medication for 330 days between t₂ and t₄ and there was a total of 360 days between the first pharmacy claim at t₂ and the depletion date of the last pharmacy claim t₄, their MPR was 91.6%.

Unadjusted MPR was calculated using ordinary least squares regression for the mail order versus the retail cohorts. To account for the variable length of follow-up for each patient, a modification of the basic MPR calculation was used. To create an adjusted MPR, a weighted Poisson regression was conducted on our matched cohorts. The dependent variable was the number of days a patient was in possession of the index diabetes medication in the pre-index period and the post-index period. The independent variables included pharmacy channel (mail vs. retail) and deciles of the propensity score. In the post-index period model, the pre-index MPR was included. The weighting variable (i.e., offset) was the natural log of the total number of days between the start of therapy after the match date and the depletion date. The use of the natural log of the total number of days as the model offset variable produced a weighted mean count of the total days covered divided by the total potential number of days of therapy. This created a measure approximating the MPR and a 95% confidence interval for each patient.

The relationship between exposure status and healthcare costs was modeled using Tobit regression to account for the skewed nature of cost data and to avoid the need for log-transformations and post-transformation adjustments[23]. Costs were measured on a monthly basis to account for the differential degree of follow-up for the initial retail claim period and the post-switch claim period. Total healthcare costs included all charges incurred for outpatient pharmacy, and inpatient and outpatient medical services, even if not related to the primary diagnosis of diabetes. The diabetes-specific healthcare costs for a patient included all charges incurred for outpatient pharmacy claims for medications used to treat diabetes (GPI codes beginning 27–15 through 27–99) and inpatient and outpatient medical services with a primary or secondary diagnosis of diabetes (ICD–9 = 250.xx, 357.2, 362.0x, 366.41, and 648.0). Cost values were calculated for two time periods: (1) during the initial retail claim period (2) from the index mail order date or the match date for the retail pharmacy-only group to the end of therapy date.

All data was de-identified in accordance with the Health Insurance Portability and Accountability Act requirements. This research was exempt from approval from an Institutional Review Board based on the Code of Federal Regulation, §46.101b, from the United States Department of Health and Human Resources[24]. Data structures were created and statistical analyses were conducted using SAS 9.2 (SAS Inc., Cary, NC) and Stata MP 10 (StataCorp, College Station, TX).

Results

The patient selection procedure is outlined in Figure 2. The final cohort included 14,732 individuals who switched to mail order pharmacy after at least 6 months in retail pharmacy, and 144,507 patients who remained in retail pharmacy for at least 18 months. 132 mail order patients were excluded due to incomplete data, and 58 retail patients were excluded (i.e., trimmed) because their propensity score was either above or below the maximum or minimum propensity score value found in the mail order pharmacy cohort. 14,600 individuals met all of the study inclusion criteria, and these individuals were matched to 43,800 retail pharmacy-only users. The distribution of key covariates and demographics for the patient population is presented in Tables 1 and 2.

Figure 2. Selection process for population of patients on oral diabetes medications in the MarketScan database between January 1, 2005 and December 31, 2007 (f/u = follow-up).

Table 1.  Characteristics of cohort of patients on oral diabetes medications in the MarketScan database between January 1, 2005 and December 31, 2007.

	Total	Retail only	(%)	Mail order	(%)	Chi-sq
Index diabetes medication order (GPI 4 digit)
Amylin analogs (2715)	75	57	0.1	18	0.1
Incretin mimetic agents (2717)	700	456	1.0	244	1.7
Sulfonylureas (2720)	10,502	7613	17.4	2889	19.8
Amino acid derivatives (2723)	271	170	0.4	101	0.7
Biguanides (2725)	30,095	23,081	52.7	7014	48.0	<0.0001
Meglitinide analogues (2728)	260	162	0.4	98	0.7
Alpha-glucosidase inhibitors (2750)	88	47	0.1	41	0.3
Thiazolidinediones (2760)	8014	5595	12.8	2419	16.6
Combination products (2799)	8395	6619	15.1	1776	12.2
US census region
New England	1811	1301	3.0	510	3.5
Middle Atlantic	2950	1898	4.3	1052	7.2
South Atlantic	9448	6102	13.9	3346	22.9
East South Central	3870	2962	6.8	908	6.2
West South Central	13,348	10,682	24.4	2666	18.3	<0.0001
East North Central	8128	6835	15.6	1293	8.9
West North Central	9787	7761	17.7	2026	13.9
Mountain	2281	1724	3.9	557	3.8
Pacific	6384	4275	9.8	2109	14.4
Unknown	393	260	0.6	133	0.9
Diabetes complication severity index
0	43,642	33,111	75.6	10,531	72.1
1	8125	5889	13.4	2236	15.3
2	4260	3088	7.1	1172	8.0	<0.0001
3	1304	936	2.1	368	2.5
4 or more	1069	776	1.8	293	2.0
Additional characteristics related to adherence
Antidepressant agent use	13,408	9414		3994		<0.0001
Lipid lowering agent use	30,583	20,618		9965		<0.0001
Antihypertensive agent use	34,009	23,884		10,125		<0.0001
Diagnoses for depression	2633	1859		774		<0.0001

Table 2.  Characteristics of cohort of patients on oral diabetes medications in the MarketScan database between January 1, 2005 and December 31, 2007.

	Mean	95% CI lower	95% CI upper
Number of pre-index retail days
Retail	354.09	352.70	355.49
Mail	354.09	351.68	356.51
Number of follow-up days
Retail	453.64	452.50	454.79
Mail	546.98	545.00	548.96
Number of any pharmacy claims
Retail	17.86	17.78	17.94
Mail	14.55	14.41	14.68
Copayment for index diabetes medication
Retail	12.67	12.55	12.78
Mail	11.02	10.81	11.22
Charlson comorbidity index score
Retail	1.84	1.82	1.85
Mail	2.01	1.99	2.04
Age at index diabetes medication
Retail	50.16	50.08	50.25
Mail	52.46	52.31	52.60

Propensity scores were estimated using the index medication type, total number of pharmacy claims during the study period, depression diagnosis, CCI score, US census region, age at the time of index claim, member out-of-pocket cost for the index claim, anti-hypertensive use, anti-depressant use, and the modified DCSI score. Five potential propensity score models were evaluated. Balance, as measured by comparing the covariate values and overlap of propensity score values by exposure group, was best achieved by the fully specified model with third-order interaction terms. The propensity score was not able to create balance in lipid-lowering agent use, thus the variable was maintained as an independent variable in the model. Mean values for these measures are presented in Table 2. Additionally, the step-wise selection process selected 21 interaction terms for the model. The interaction terms are available from the authors upon request.

The mean initial retail follow-up time of 354 days was similar between the retail pharmacy-only and mail order groups, and the MPR within this period was 68.8 (95% CI: 68.5–69.1) in the retail pharmacy-only group and 70.0 (95% CI: 69.5–70.5) in the mail order group. The mean number of days followed after the index mail order date was 454 days in the retail pharmacy-only cohort and 547 days in mail order cohort (difference in means = −93.3, 95% CI −95.6 to −91.0). The unadjusted total MPR for the retail-only group was 67.1% (95% CI: 67.01–67.27) and 80.1 (95% CI: 79.7–80.5) for those who switched to mail order. After adjusting for deciles of the propensity score and pre-period MPR, the adjusted total MPR was 63.4 (95% CI: 63.1–63.7) in the retail pharmacy-only group and 84.8 (95% CI: 84.4–85.1) in the mail order group. These data are shown in Table 3.

Table 3.  Medication adherence data for cohort of patients on oral diabetes medications in the MarketScan database between January 1, 2005 and December 31, 2007.

	Total	Lower limit	Upper limit
Continuous MPR before change*
Retail	68.8%	68.5%	69.1%
Mail order	70.0%	69.5%	70.5%
Continuous MPR after change^†
Retail	63.4%	63.1%	63.7%
Mail order	84.8%	84.4%	85.1%

*Adjusting for: diabetes drug, total medication use, psychiatric diagnoses, medication copay, Charlston co-morbidity index, geography, age, anti-depressant use, lipid lowering use, anti-hypertensive use, and Diabetes complication and severity index.

^†Adjusting for: above and initial retail claim period MPR.

Table 4 presents the change in per-member-per-month total and diabetes-related healthcare costs for the mail order pharmacy group versus the retail pharmacy-only group. Total per-member-per-month overall healthcare and medical costs were on average $34.32 (95% CI: 57.89–10.76) and $37.54 (95% CI: 55.53–19.55) lower, respectively in the mail order group than the retail pharmacy-only group at the end of follow-up for each patient. Total diabetes-related costs were higher in the mail order group, primarily driven by the increase in diabetes medication costs in the mail order pharmacy group. Diabetes-related medical costs were on average $19.14 (95% CI: 25.16–13.11) lower in the mail order pharmacy group at the end of the follow-up period. The average per-member-per-month total and diabetes-related healthcare, medical and medication costs are presented in Tables 5 and 6. Costs incurred during the entire follow-up period for total healthcare, medical and medication costs were generally lower in the mail order population.

Table 4.  Differences in per member per month total healthcare expenditures for cohort of patients using mail order pharmacy for oral diabetes medications vs. retail in the MarketScan database between January 1, 2005 and December 31, 2007.

	PMPM Δ ($)	Lower limit ($)	Upper limit ($)
Total healthcare costs*
Period prior to change	(71.18)	(106.91)	(35.46)
Total follow-up after change	(34.32)	(57.89)	(10.76)
Total medical costs*
Period prior to change	(121.80)	(163.68)	(79.91)
Total follow-up after change	(37.54)	(55.53)	(19.55)
Total medication costs*
Period prior to change	26.82	19.07	34.57
Total follow-up after change	(7.45)	(20.13)	5.22
Diabetes-related healthcare costs*
Period prior to change	(20.10)	(30.72)	(9.48)
Total follow-up after change	6.22	0.04	12.40
Diabetes medical costs*
Period prior to change	(47.03)	(61.65)	(32.40)
Total follow-up after change	(19.14)	(25.16)	(13.11)
Diabetes medication costs*
Period prior to change	12.23	11.04	13.43
Total follow-up after change	14.13	12.35	15.91

*Adjusting for: Total MPR, initial retail claim period MPR, diabetes drug, total medication use, psychiatric diagnoses, medication copay, Charlston co-morbidity index, geography, age, anti-depressant use, lipid lowering use, anti-hypertensive use, and Diabetes complication and severity index.

Table 5.  Per member per month total healthcare expenditures for cohort of patients on oral diabetes medications in the MarketScan database between January 1, 2005 and December 31, 2007.

	PMPM ($)	Lower limit ($)	Upper limit ($)
Total healthcare costs*
Period prior to change
Retail	431.98	403.71	460.26
Mail order	360.80	313.67	407.93
Total follow-up after change
Retail	340.46	311.95	368.96
Mail order	306.13	265.61	346.66
Total medical costs*
Period prior to change
Retail	332.73	309.86	355.60
Mail order	210.94	158.83	263.04
Total follow-up after change
Retail	246.63	218.28	274.97
Mail order	209.09	169.42	248.75
Total medication costs*
Period prior to change
Retail	121.37	116.62	126.12
Mail order	148.20	138.26	158.13
Total follow-up after change
Retail	90.44	78.99	101.90
Mail order	82.99	61.39	104.60

*Adjusting for: total MPR, initial retail claim period MPR, diabetes drug, total medication use, psychiatric diagnoses, medication copay, Charlston co-morbidity index, geography, age, anti-depressant use, lipid lowering use, anti-hypertensive use, and Diabetes complication and severity index.

Table 6.  Per member per month diabetes-related healthcare expenditures for cohort of patients on oral diabetes medications in the MarketScan database between January 1, 2005 and December 31, 2007.

	PMPM ($)	Lower limit ($)	Upper limit ($)
Diabetes-related healthcare costs*
Period prior to change
Retail	94.41	86.78	102.04
Mail order	74.31	61.70	86.92
Total follow-up after change
Retail	97.92	90.59	105.26
Mail order	104.15	92.19	116.10
Diabetes-related medical costs*
Period prior to change
Retail	89.61	79.02	100.20
Mail order	42.58	23.75	61.42
Total follow-up after change
Retail	80.06	71.51	88.60
Mail order	60.92	48.60	73.24
Diabetes-related medication costs*
Period prior to change
Retail	26.85	26.13	27.58
Mail order	39.09	37.62	40.55
Total follow-up after change
Retail	22.78	20.21	25.36
Mail order	36.92	33.68	40.16

*Adjusting for: total MPR, initial retail claim period MPR, diabetes drug, total medication use, psychiatric diagnoses, medication copay, Charlston co-morbidity index, geography, age, anti-depressant use, lipid lowering use, anti-hypertensive use, and Diabetes complication and severity index.

Discussion

The principle objectives of this study were to determine the differences in diabetes medication adherence and healthcare costs in patients who receive their medications through mail order pharmacy versus retail pharmacy. A large sample of individuals who switched from retail pharmacy to mail order pharmacy were identified and matched to individuals in retail pharmacy based upon initial retail claim period and the minimum amount of time of follow-up after the index mail order date. The results indicate that after adjusting for measured confounders of medication adherence and initial retail claim period MPR, patients who switched to mail order pharmacy services had greater adherence to their diabetes medication at the end of their follow-up, and had per-member-per-month total and diabetes-related healthcare and medical costs that trended lower. Diabetes medication-related costs were higher in the mail order group, potentially the product of the increased number of prescriptions associated with greater adherence, or the addition of new medication orders to treat diabetes.

There are several limitations to this study. First, because of the nature of retrospective data analyses, a marker for actual medication adherence was used in this study. While the medication possession ratio has been used extensively in the research literature, and has been endorsed by the National Quality Forum as the preferred method for measuring adherence to chronic mediations[25], it is at best a surrogate marker for actual medication taking activity. The study population was limited to individuals under 65 years of age. Although diabetes is prevalent in the 65 and older population, the absence of those older than 65 is unlikely to bias the results. Age itself has been shown to be a significant predictor of adherence in some chronic medication classes including diabetes[10],[26], and was included in the current analysis. Limiting the analysis to those under 65 restricts the impact of age and eliminates any potential confounding present in that age group but it limits the generalizability of the findings to patients under age 65.

An additional limitation is the availability of just 3 years of follow-up data. Adherence to medications should have its greatest effect over the course of time. While the total follow-up time in the mail order group approached 18 months, this amount of time was likely insufficient to identify the long-term complications of diabetes, such as renal failure, coronary heart disease, retinopathy and peripheral vascular disease[27]. This also limits the interpretability of the study cost data, as any diabetes-related costs are likely to be related to disease processes that have begun occurring prior to the study period and progressing in severity and cost after the study period.

An additional limitation was the potential inability of propensity scores to stratify individuals on their probability to use mail order pharmacy. The regression diagnostics of the propensity score model indicated that the model created good balance between all of the measured confounders (other than lipid lowering agents) of the mail order use and medication adherence relationship. Important unmeasured patient variables such as the proximity to retail pharmacy, inducements to use mail or retail pharmacy, patient satisfaction with the drug's efficacy, medication side effects and behavioral aspects that influence medication adherence may decrease the accuracy of the propensity score model. As the propensity score attempts to adjust for unmeasured differences in the study population, any failure of the propensity score would result in residual confounding.

A further limitation that could confound the relationship between the mail order and retail pharmacy groups with regards to costs is the inability to fully adjust for disease severity. The database does not include information on potentially important confounders to estimating the healthcare costs of diabetes: body mass index and measures of diabetes control (e.g., blood glucose, HgbA_1C). Variation of these variables has been associated with healthcare outcomes and utilization[26],[28]. Diabetes severity was addressed using the DSCI, but it is unknown how this severity index truly performed without more direct measures of diabetes severity.

Finally, there is no information on pharmacy benefit. The choice to enroll in a mail order pharmacy may be dependent upon a patient's pharmacy benefit design. Patients may be restricted to retail pharmacy-only, mail order-only or a plan that encourages a combination of the two. No information was available regarding the patient's pharmacy benefit design. There are two ways that this could influence the use of mail order pharmacy. Patients may have a benefit design that requires the use of a mail order pharmacy for chronic medications. Patients with such a design are typically allowed only two medication fills for a chronic medication in a retail pharmacy before requiring the additional refills be processed through mail service. The impact of this potential influence on mail order use was limited by requiring that a patient have at least 6 months of retail-only pharmacy claims prior to making the switch to mail order pharmacy. These patients may have been less likely to have been forced to choose mail order pharmacy, unless the pharmacy benefit plan changed after the first medication claim. The second way that the pharmacy benefit design could influence the use of mail order pharmacy is if patients were not offered a mail order pharmacy option. There was not a good method for ruling out individuals that did not have a mail order option. The pharmacy benefits management industry, which manages the prescription medication benefits for most health plans, generally encourages mail order utilization in some form as a low cost option and often creates inducements for mail service, which likely limits the number of patients that were subject to a retail-only pharmacy benefit design.

Conclusion

With the above-mentioned limitations in mind, the study results indicate that individuals who switch to mail order pharmacy services are more likely to maintain possession of their medication and continue to fill their medication, and generally utilize fewer total and diabetes-related healthcare services over time, which results in lower costs. These finding have important implications to understanding the relationship between adherence and healthcare costs. While there are many reasons a patient may fail to be adherent with their medication, mail order pharmacy lessens the potential role of medication supply. Through larger quantities and few travel requirements, patients may find mail order pharmacy a less burdensome way to maintain adequate supplies of medication.

This study also supports the theory that chronic use of diabetes medications has a positive impact on overall healthcare costs. While it cannot be summarily concluded that the use of mail order pharmacy is responsible for these adherence improvements and cost differences because of the lack of randomization and long-term follow-up, it does appear that patients who choose to switch to and continue using a mail order pharmacy to fill prescriptions benefit from the experience through better medication adherence over time. Additional research on the long term effects of mail order pharmacy on medication adherence and its relationship to long-term diabetes complications is recommended.

Transparency

Declaration of funding: This study was funded by Express Scripts.

Declaration of financial/other relationships: SD, AV and HS have disclosed that they are employees of Express Script, a pharmacy benefits management company.

Acknowledgements: The authors would like to acknowledge Mike Kattan of Cleveland Clinic, Cleveland, OH USA, and Emily Cox, Yakov Svirnovskiy and Jim Indelicato of Express Scripts for their assistance with data gathering, study design and statistical techniques used in this manuscript.