Multiple Pharmacy Use, Medication Adherence in Older Adults
Medicare claims (Parts A, B, and D) and enrollment data for individuals in the Chronic Conditions Warehouse 10% random sample of Part D beneficiaries who were continuously enrolled in a stand-alone prescription drug plan and alive throughout 2009 were obtained (n = 1,529,825). Beneficiaries were included if they were aged 65 and older on January 1, 2009, and filled one or more prescriptions at a commercial retail or mail order pharmacy during the study period (n = 926,956). Those who were missing any pharmacy identifying information (<1%) or filled any prescriptions at other pharmacy types (e.g., long-term care, home infusion, Indian Health Service, Department of Veterans Affairs, institutional, specialty, nuclear, and compounding; 15.8%) were excluded. The University of Pittsburgh institutional review board approved the study.
Outcome Measures. Medication adherence was measured using the proportion of days covered (PDC), defined as the number of days in a period "covered" by a medication divided by number of days in the period. Unlike another commonly used measure of adherence using pharmacy claims (Medication Possession Ratio), the PDC avoids overestimating adherence when individuals refill a medication before the previous prescription runs out. The PDC is a National Committee for Quality Assurance/National Quality Forum (NQF)–endorsed measure for medication adherence and is used in the Medicare Star Ratings. The measurement period was defined as starting with the beneficiary's date of first fill of a medication in one of the classes of interest and ending on December 31, 2009. The days' supply for each prescription fill was captured from the "days' supply" field in the PDE data set (i.e., no calculations were required to determine days' supply). Days' supply with the last prescription fill in the measurement period was truncated on the last day of observation and not included in the PDC calculation for any supply beyond that to avoid overestimating adherence. Consistent with the NQF quality measures, a dichotomous measure of medication adherence (PDC ≥0.80) was constructed in each of eight therapeutic classes that older adults widely use (beta-blockers, renin angiotensin system antagonists, calcium channel blockers, statins, sulfonylureas, biguanides (metformin), thiazolidinediones, and dipeptidyl peptidase-IV inhibitors).
Second, a dichotomous measure for presence of a potentially clinically significant DDI was constructed. Using a previously developed algorithm and information from medication package inserts, beneficiaries filling two of several interacting medications (available upon request) during the same time period were identified. Presence of a DDI was defined as 1 or more overlapping days in which the beneficiary possessed two interacting medications. Only oral, nontopical dosage forms were included in the DDI analysis.
Independent Variables. Multiple pharmacy use can be defined in several ways (Table 1). One important question is whether multiple pharmacy use is concurrent or sequential, the latter of which may be the case for "snowbirds," who live part of the year in another state, or those who switch pharmacies at some point in the year. As such, three non-overlapping groups were defined: single pharmacy use for the entire year, sequential multiple pharmacy use in the year, and at least one instance of concurrent multiple pharmacy use. Specifically, the number of different pharmacy identification codes from the Part D pharmacy characteristics file were first used to classify participants as using a single pharmacy or multiple pharmacies, and then the fill dates were used to further classify those who used multiple pharmacies as doing so sequentially versus concurrently. Sequential multiple pharmacy use was defined as filling at least one prescription at two or more pharmacies without overlapping fill dates throughout the year. Concurrent multiple pharmacy use was defined as filling at least one prescription at two or more pharmacies with at least some overlap in fill dates throughout the year. In addition, a primary pharmacy was defined for each beneficiary as the pharmacy where the plurality of prescriptions were filled in 2009.
Another important issue in defining multiple pharmacy use is whether it occurs within a pharmacy chain but at different physical locations (affiliated) or at different chains (unaffiliated). Pharmacists operating at different locations within the same chain may not know the individual's medication history in detail but may have access to complete electronic data on prescriptions filled. The "relationship type" variable in the Part D pharmacy characteristics file was used to determine whether the pharmacy had a chain or franchise relationship with another entity. It was hypothesized that the effects of multiple pharmacy use might be different for pharmacies with the same corporate parent than for those with different corporate parents.
Covariates. Covariates were grouped into three main categories: socio-demographic (predisposing), access to care (enabling), and health status (medical need) factors. Sociodemographic variables included age, sex, and race and ethnicity. Access-to-care variables included a composite indicator of low-income subsidy (LIS) and dual eligible status and the presence of a national Part D plan (a plan that offers a benefit package in all 34 U.S. Prescription Drug Plan (PDP) regions). The effect of including fixed effects for the Part D plan as a covariate to adjust for any plan-level use of tools that may affect adherence (cost sharing) or DDIs (medication monitoring) was also assessed. ZIP code was used to calculate beneficiaries' geographic location (rural vs urban), and the total number of unique prescribers for each beneficiary during 2009 was calculated. The beneficiary's total number of unique medications dispensed at all pharmacies in 2009, an indicator for end-stage renal disease, and nine chronic conditions captured in the Chronic Condition Data Warehouse for which medication adherence and DDIs are important (diabetes mellitus, heart failure, hypertension, hyperlipidemia, asthma, chronic obstructive pulmonary disease, osteoporosis, osteoarthritis and rheumatoid arthritis, and depression) were used to indicate health status, and a Charlson Comorbidity Index score was determined. Whether the primary pharmacy was a community or retail pharmacy or a mail order pharmacy, and whether it was an independent, chain, or other were also adjusted for, and a dichotomous variable for any mail order use was created.
Statistical Analysis. Analyses were conducted using Stata version 11.0 (Stata Corp, College Station, TX). Descriptive statistics were examined for all variables, and t-tests and chi-square tests were used to assess differences according to single versus multiple pharmacy use for continuous and categorical variables, respectively. Current best practices for multiple propensity score weighting in multinomial treatment analysis were used to control for potential differences in health status between beneficiaries using single and multiple pharmacies. As such, a multinomial logistic regression model was estimated to calculate the probabilities of an individual using a single pharmacy, multiple pharmacies sequentially, or multiple pharmacies concurrently, adjusted for all sociodemographic, access-to-care, and health status variables. Those probabilities were used to generate inverse probability of treatment (IPT) weights, which are the inverse of the probability an individual receives the "treatment" (multiple concurrent, multiple sequential, or single pharmacy use). To assess the extent to which the propensity score weighting achieved balance of covariates, differences in covariate values of the three groups of pharmacy users before and after applying the IPT weights were examined using linear, logistic, and multinomial logistic regression depending on the nature of the covariate. The IPT weights were then used to weight the observations in two separate logistic regression models examining the effects of multiple pharmacy use on medication adherence and potential DDIs. Finally, adjusted predicted probabilities were estimated for both outcomes using Stata's "margins" command. The delta method was used to calculate 95% confidence intervals (CIs), which allows for correlation between observations. All regression models used robust estimates of variance.
A series of sensitivity analyses was conducted to test for robustness. First, in the primary analysis, multiple pharmacy use was defined based on different physical locations (even within the same chain or parent company). In a secondary analysis, analyses were rerun using the more-restrictive measure of multiple pharmacy use that defined multiple pharmacy use only as filling prescriptions at multiple physical locations with different parent companies. Second, the effect of assessing only community or retail pharmacies (excluding mail order) on both outcomes was tested. Third, the PDC was examined as a continuous variable, and different cut points (e.g., ≥0.60, ≥0.70) were used to assess robustness for the adherence outcome. The effect on the PDC of excluding beneficiaries filling only one prescription, those who were hospitalized throughout the year, and those who were LIS beneficiaries was also examined. Fourth, an analysis was conducted on the DDI outcome using a more-stringent definition of multiple pharmacy use that required that interacting medications be filled at different pharmacies. More-restrictive definitions of the presence of a DDI were also assessed (≥10 and ≥30 overlapping days of interacting drugs). The results of these sensitivity analyses were qualitatively similar in direction and magnitude to the main analyses.
Methods
Study Population
Medicare claims (Parts A, B, and D) and enrollment data for individuals in the Chronic Conditions Warehouse 10% random sample of Part D beneficiaries who were continuously enrolled in a stand-alone prescription drug plan and alive throughout 2009 were obtained (n = 1,529,825). Beneficiaries were included if they were aged 65 and older on January 1, 2009, and filled one or more prescriptions at a commercial retail or mail order pharmacy during the study period (n = 926,956). Those who were missing any pharmacy identifying information (<1%) or filled any prescriptions at other pharmacy types (e.g., long-term care, home infusion, Indian Health Service, Department of Veterans Affairs, institutional, specialty, nuclear, and compounding; 15.8%) were excluded. The University of Pittsburgh institutional review board approved the study.
Study Variables
Outcome Measures. Medication adherence was measured using the proportion of days covered (PDC), defined as the number of days in a period "covered" by a medication divided by number of days in the period. Unlike another commonly used measure of adherence using pharmacy claims (Medication Possession Ratio), the PDC avoids overestimating adherence when individuals refill a medication before the previous prescription runs out. The PDC is a National Committee for Quality Assurance/National Quality Forum (NQF)–endorsed measure for medication adherence and is used in the Medicare Star Ratings. The measurement period was defined as starting with the beneficiary's date of first fill of a medication in one of the classes of interest and ending on December 31, 2009. The days' supply for each prescription fill was captured from the "days' supply" field in the PDE data set (i.e., no calculations were required to determine days' supply). Days' supply with the last prescription fill in the measurement period was truncated on the last day of observation and not included in the PDC calculation for any supply beyond that to avoid overestimating adherence. Consistent with the NQF quality measures, a dichotomous measure of medication adherence (PDC ≥0.80) was constructed in each of eight therapeutic classes that older adults widely use (beta-blockers, renin angiotensin system antagonists, calcium channel blockers, statins, sulfonylureas, biguanides (metformin), thiazolidinediones, and dipeptidyl peptidase-IV inhibitors).
Second, a dichotomous measure for presence of a potentially clinically significant DDI was constructed. Using a previously developed algorithm and information from medication package inserts, beneficiaries filling two of several interacting medications (available upon request) during the same time period were identified. Presence of a DDI was defined as 1 or more overlapping days in which the beneficiary possessed two interacting medications. Only oral, nontopical dosage forms were included in the DDI analysis.
Independent Variables. Multiple pharmacy use can be defined in several ways (Table 1). One important question is whether multiple pharmacy use is concurrent or sequential, the latter of which may be the case for "snowbirds," who live part of the year in another state, or those who switch pharmacies at some point in the year. As such, three non-overlapping groups were defined: single pharmacy use for the entire year, sequential multiple pharmacy use in the year, and at least one instance of concurrent multiple pharmacy use. Specifically, the number of different pharmacy identification codes from the Part D pharmacy characteristics file were first used to classify participants as using a single pharmacy or multiple pharmacies, and then the fill dates were used to further classify those who used multiple pharmacies as doing so sequentially versus concurrently. Sequential multiple pharmacy use was defined as filling at least one prescription at two or more pharmacies without overlapping fill dates throughout the year. Concurrent multiple pharmacy use was defined as filling at least one prescription at two or more pharmacies with at least some overlap in fill dates throughout the year. In addition, a primary pharmacy was defined for each beneficiary as the pharmacy where the plurality of prescriptions were filled in 2009.
Another important issue in defining multiple pharmacy use is whether it occurs within a pharmacy chain but at different physical locations (affiliated) or at different chains (unaffiliated). Pharmacists operating at different locations within the same chain may not know the individual's medication history in detail but may have access to complete electronic data on prescriptions filled. The "relationship type" variable in the Part D pharmacy characteristics file was used to determine whether the pharmacy had a chain or franchise relationship with another entity. It was hypothesized that the effects of multiple pharmacy use might be different for pharmacies with the same corporate parent than for those with different corporate parents.
Covariates. Covariates were grouped into three main categories: socio-demographic (predisposing), access to care (enabling), and health status (medical need) factors. Sociodemographic variables included age, sex, and race and ethnicity. Access-to-care variables included a composite indicator of low-income subsidy (LIS) and dual eligible status and the presence of a national Part D plan (a plan that offers a benefit package in all 34 U.S. Prescription Drug Plan (PDP) regions). The effect of including fixed effects for the Part D plan as a covariate to adjust for any plan-level use of tools that may affect adherence (cost sharing) or DDIs (medication monitoring) was also assessed. ZIP code was used to calculate beneficiaries' geographic location (rural vs urban), and the total number of unique prescribers for each beneficiary during 2009 was calculated. The beneficiary's total number of unique medications dispensed at all pharmacies in 2009, an indicator for end-stage renal disease, and nine chronic conditions captured in the Chronic Condition Data Warehouse for which medication adherence and DDIs are important (diabetes mellitus, heart failure, hypertension, hyperlipidemia, asthma, chronic obstructive pulmonary disease, osteoporosis, osteoarthritis and rheumatoid arthritis, and depression) were used to indicate health status, and a Charlson Comorbidity Index score was determined. Whether the primary pharmacy was a community or retail pharmacy or a mail order pharmacy, and whether it was an independent, chain, or other were also adjusted for, and a dichotomous variable for any mail order use was created.
Statistical Analysis. Analyses were conducted using Stata version 11.0 (Stata Corp, College Station, TX). Descriptive statistics were examined for all variables, and t-tests and chi-square tests were used to assess differences according to single versus multiple pharmacy use for continuous and categorical variables, respectively. Current best practices for multiple propensity score weighting in multinomial treatment analysis were used to control for potential differences in health status between beneficiaries using single and multiple pharmacies. As such, a multinomial logistic regression model was estimated to calculate the probabilities of an individual using a single pharmacy, multiple pharmacies sequentially, or multiple pharmacies concurrently, adjusted for all sociodemographic, access-to-care, and health status variables. Those probabilities were used to generate inverse probability of treatment (IPT) weights, which are the inverse of the probability an individual receives the "treatment" (multiple concurrent, multiple sequential, or single pharmacy use). To assess the extent to which the propensity score weighting achieved balance of covariates, differences in covariate values of the three groups of pharmacy users before and after applying the IPT weights were examined using linear, logistic, and multinomial logistic regression depending on the nature of the covariate. The IPT weights were then used to weight the observations in two separate logistic regression models examining the effects of multiple pharmacy use on medication adherence and potential DDIs. Finally, adjusted predicted probabilities were estimated for both outcomes using Stata's "margins" command. The delta method was used to calculate 95% confidence intervals (CIs), which allows for correlation between observations. All regression models used robust estimates of variance.
A series of sensitivity analyses was conducted to test for robustness. First, in the primary analysis, multiple pharmacy use was defined based on different physical locations (even within the same chain or parent company). In a secondary analysis, analyses were rerun using the more-restrictive measure of multiple pharmacy use that defined multiple pharmacy use only as filling prescriptions at multiple physical locations with different parent companies. Second, the effect of assessing only community or retail pharmacies (excluding mail order) on both outcomes was tested. Third, the PDC was examined as a continuous variable, and different cut points (e.g., ≥0.60, ≥0.70) were used to assess robustness for the adherence outcome. The effect on the PDC of excluding beneficiaries filling only one prescription, those who were hospitalized throughout the year, and those who were LIS beneficiaries was also examined. Fourth, an analysis was conducted on the DDI outcome using a more-stringent definition of multiple pharmacy use that required that interacting medications be filled at different pharmacies. More-restrictive definitions of the presence of a DDI were also assessed (≥10 and ≥30 overlapping days of interacting drugs). The results of these sensitivity analyses were qualitatively similar in direction and magnitude to the main analyses.