Association between insulin adherence and hemoglobin A1c: A systematic review and meta-analysis

Abstract

BACKGROUND:

The public and private sectors have improved insulin affordability to better access and outcomes. However, measuring the clinical and economic impact of these policies is difficult because of limited data availability. Quantifying an association between insulin adherence, based on prescription fills, and glycemic control could make use of large national claims databases.

OBJECTIVE:

To evaluate the magnitude of association between insulin adherence measured by prescription fills and glycemic control in adults with type 2 diabetes mellitus.

METHODS:

A search in Cochrane Central Register of Controlled Trials, EMBASE, and PubMed from inception until October 2022 was performed. Studies were included if insulin adherence, measured by prescription fills, and glycemic control were reported for adults with type 2 diabetes mellitus. Studies were limited to cohort, randomized controlled trial (RCT), or case-control. A random-effects model using the Dersimonian and Laird method was performed to meta-analyze the association of adherence on hemoglobin A1c (A1c). Heterogeneity was evaluated using the I² statistical test. Risk of bias was assessed through the Newcastle-Ottawa Scale.

RESULTS:

There was a total of 11 articles included in the qualitative analysis; 10 studies involving 19,742 patients were included in the quantitative analysis. Adherence was measured as both a categorical (adherent vs nonadherent) and a continuous (change in adherence) variable. Three cohort studies involving a total of 6,077 patients compared the A1c difference between insulin-adherent and nonadherent patients as defined by the study, generally with a fill rate of greater than or equal to 80%. Compared with nonadherent patients, adherent patients had an A1c mean difference of −0.49% (95% CI = −0.76% to −0.22%; P < 0.05; I² = 40.7%). Out of these 3 studies, 1 had a low risk of bias, 1 moderate, and 1 high. Six cohort studies and 1 RCT with a total of 13,665 patients evaluated adherence as a continuous variable. An increase in adherence of 1% was associated with an A1c mean difference of −0.04% (95% CI = −0.06% to −0.01%; P < 0.05; I² = 72.4%). Out of these 7 studies, 4 had a low risk of bias and 3 had a high risk of bias.

CONCLUSIONS:

Insulin adherence based on prescription fill–rate measures was significantly associated with an A1c reduction.

Study registration number: PROSPERO (CRD42021279904)

Plain language summary

The connection between glucose control and consistent insulin use was explored. Prescription fill measures were used to measure consistent insulin use. These measures use paid claims. Those who pay for insulin more may be more likely to take their insulin consistently. We found a link between consistent insulin use and better blood sugar levels. Paid claims may be used to evaluate insulin adherence.

Implications for managed care pharmacy

Our study found a statistically and clinically significant relationship between insulin adherence, based on prescription fill records, and hemoglobin A1c (A1c). Different policies, such as the Inflation Reduction Act, have been recently implemented to improve insulin affordability. In the absence of widely available A1c data and more readily available claims data, insulin adherence can be used to estimate the short-term and long-term clinical and economic effects of these policies.

Self-reported insulin adherence has ranged from 43% to 86 %, with cost being cited as a main reason for suboptimal adherence.¹ From 2014 to 2019, the average annual insulin price increased by 55%.² In 2020, 16.9%-30% of diabetic patients reported rationing their insulin because of unaffordability.³ Both the private and the public sectors have taken action to mitigate cost-related nonadherence. Numerous states since 2019 have established state-specific 30-day copay caps, which range from $25 to $100.⁴ In early 2023, the Inflation Reduction Act capped monthly insulin copay at $35 for Medicare beneficiaries.⁵ In March 2023, Eli Lilly, Novo Nordisk, and Sanofi announced plans to reduce insulin prices, taking place in 2023 and January 2024, with prices being reduced by 65%-75% or with monthly costs being capped to $35.⁶

Measuring the magnitude of the long-term clinical and economic impacts of these policy changes can be challenging. Although hemoglobin A1c (A1c) is a good clinical marker that can be used to predict short-term and long-term diabetic outcomes, individual A1c outcomes are not readily available at state and national levels. However, there are large national and state claims databases that can be used to potentially evaluate insulin adherence based on prescription fills. The clinical and economic effects of these policies are mediated through reducing cost-related nonadherence. By establishing an association between insulin adherence and A1c outcomes, claims data can be used to estimate the short-term and long-term effects of these policies. Although smaller studies have measured this relationship, there is not yet a comprehensive review summarizing and pooling these outcomes in adults with type 2 diabetes mellitus (T2DM).^7,8 This systematic review aims to fill that gap by evaluating magnitude of association between insulin adherence measured by prescription fills and glycemic control in this population.

Methods

The study protocol was registered with PROSPERO (CRD42021279904). The Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) reporting guideline was followed when reporting the review.⁹ Any differences between the initial protocol and the final analysis were described in the Supplementary Materials ^{(648.8KB, pdf)} (Supplementary Table 1 ^{(648.8KB, pdf)} , available in online article).

SEARCH STRATEGY AND STUDY SELECTION

The electronic databases, PubMed, Embase, and Cochrane Central, were searched from inception to October 12, 2022. Search terms included various insulin types and adherence measurements (Supplementary Table 2 ^{(648.8KB, pdf)} ). Reference lists were screened from retrieved articles. No language restriction was applied.

Studies meeting the following criteria were included: (1) if they were randomized controlled trials (RCTs), case-control studies, or cohort studies; (2) if a measure of glycemic control, including but not limited to A1c, and a measure of insulin adherence based on objective prescription data were reported; and (3) if the participants were adults with T2DM. Prescription fill–based adherence measures include medication possession ratio (MPR), proportion of days covered (PDC), continuous measure of medication gaps, continuous single-interval medication availability, and more.¹⁰ Self-reported measures of adherence, such as the Morisky Medication Adherence Scale, were excluded as these measures are subjective.

Eligible studies, after removing duplicates, were screened by 2 reviewers (D.N. and X.L.) independently. Any discrepancies were resolved by a third reviewer (S.N.) through consensus.

DATA EXTRACTION

Data extraction was performed independently by 2 reviewers (D.N. and X.L.), and the third reviewer (S.N.) was referred to in the case of disagreements. Data extraction was completed using a standardized form, which included the authors’ names, publication year, baseline characteristics, insulin type, insulin adherence measurements, and glycemic control outcomes. The methodological quality of the included studies was assessed by 2 reviewers (D.N. and S.N.) using the Newcastle-Ottawa Scale.^11,12

DATA SYNTHESIS AND ANALYSIS

A random-effects model using the Dersimonian and Laird method was performed to evaluate the relationship between medication (ie, insulin) adherence and improvement in A1c levels.¹³ Two different analyses were performed: (1) adherence as a categorical variable and (2) adherence as a continuous variable based on how adherence was reported in the primary study. For the first analysis, the threshold for adherence was defined by the primary study and could vary between studies; regardless, the raw results were incorporated in the analyses. Results are presented as the mean difference (MD) in glycemic control with their 95% CI. When adherence was measured as a continuous variable, a linear relationship between the change in adherence and the change in A1c was assumed. Results were presented as the MD in glycemic control correlated with a 1% increase in adherence. Beta coefficients were used if provided. If not, the correlation was estimated by dividing the MD in A1c by the MD in adherence between the 2 groups. Statistical heterogeneity was assessed using I² statistics.1⁴ An I² estimate more than 50% was interpreted as evidence of a substantial level of heterogeneity.¹⁵ Publication bias was assessed through the Egger test.¹⁶

Subgroup analyses were performed based on the following: (1) prescription fill measures of adherence and (2) insulin type. Insulin type was either defined as long-acting, short-acting, mixed, or undefined. Sensitivity analyses were performed by excluding studies fulfilling the following criteria: (1) studies using the same prescription fill database within the same time frame, (2) high risk of bias, (3) small sample size (<25th percentile of the overall sample sizes), (4) unclear adherence definitions (terms such as MPR or PDC not explicitly stated), and (5) prescription fill history not documented electronically.¹⁷ Criterion 1 was completed in case patients were duplicated in the analyses. Criteria 2 and 3 were used to exclude studies at higher risk of biasing the results. Criteria 4 and 5 were used to exclude prescription fill adherence measures of potentially poor quality. All analyses were performed with Stata version 17.0 (StataCorp).

Results

STUDY SELECTION

A total of 3,124 records were identified through the database search: 388 duplicates were removed and then 2,569 were discarded after screening title and abstract. Of the remaining 167 full-text articles, 7 studies were eligible for inclusion (Supplementary Table 3 ^{(648.8KB, pdf)} ). An additional 4 articles, identified from screening the reference lists of the eligible studies, were included. In total, 11 studies were included.^7,8,18-26 The PRISMA flow diagram showing the search and selection process is displayed in Figure 1.

FIGURE 1.

Preferred Reporting Items for Systematic Reviews and Meta-Analyses Flow Diagram

STUDY CHARACTERISTICS

The descriptive characteristics of the included studies are provided in Table 1. Among 11 studies, 10 were retrospective cohort studies^{7,8,18-23,25,26} and 1 was an RCT.²⁴ In the RCT, patients were randomly assigned on the basis of 2 discharge interventions, and the insulin adherence and A1c outcomes were reported for each group.²⁴ The data on adherence and glucose control for the full trial cohort were compiled, extracted, and included in our review akin to data from a prospective cohort study. Of these 11 studies, only 10 were used for quantitative analysis^{7,8,18,19,21-26}; although the remaining 120 reported the magnitude of the association between insulin and adherence, it was unclear how many patients were receiving insulin. From these 10 studies included in quantitative synthesis, 19,742 patients were included, with a median sample size of 1,217 patients (range 102-6,222).^{7,8,18-23,25,26} For all 11 studies, the average age of participants ranged between 52 and 70 years, with the majority being male (range 27%-98%). Most studies were conducted in the United States (n = 9),^18-26 with the remaining being pulled from Sweden (n = 1)⁷ and Scotland (n = 1).⁸ Data were pulled from prescription fill databases (n = 10)^8,18-26 and paper fill records (n = 1).⁷ All studies followed patients for at least 3 months (range 4 months to 6.75 years).

TABLE 1.

Study Characteristics

First author (year)	Country	Study design	Adherence data source	n a	Time	Baseline characteristics				Intervention characteristics
						Mean age (SD) (y)b	Sex (%)	Mean A1C (SD)	Mean adherence, % (SD)	Insulin type	Waste?	Adherence measurec	Glycemic measure
Ayyagari (2015)	United States	RCT	IMPACT (2001-2010) (United States)	4,876	1 y	54	M: 54 F: 46	NA:9.0 (2.4);A:9.3 (2.1)	NR	Basal insulin	Not noted	Interfill periods	Change in A1c
Cramer (2005)	United States	RCT	VA PBM (2001– 2002) (Eastern United States and Puerto Rico)	6,222	2 y	NR	NR	8.0 (1.7)	77 (17)	Any	Not factored (assumed constant)	Similar to MPR	Change in A1c
Donnelly (2007)	Scotland	RCT	DARTS/ MEMO Collaboration (1995-2001) (Tayside, Scotland)	1,099	6.75 y	62 (12)	M: 52 F: 48	8.5 (1.3)	71 (18)	Not noted	Not noted	Similar to PDC	Change in A1c
Eby (2013)	United States	RCT	Innovus InVision (2006-2010) (United States)	760	1 y	P:54 (11)V:56 (12)	P: M: 55 F: 45 V: M: 54 F: 46	NR	NR	Mealtime insulin (RAI, PMX insulin)	Not noted	aMPR	Change in A1c
Egede (2014)	United States	RCT	VA PBM (1997-2006) (Southeastern United States)	NR	5.4 y	66 (12)	M: 97 F: 3	NR	NR	Not noted	Not noted	aMPR	Change in A1c
Grabner (2013)	United States	RCT	HIRD (United States)	1,466	1 y	52 (NR)	M: 57 F: 43	P: 9.5 (2.0);V: 9.4 (2.2)	NR	Insulin glargine	No waste assumed	aMPR	Change in A1c
Hood (2021)	United States	RCT	VA PBM (2014– 2017) (United States)	951	9 mo	63 (NR)	M: 98 F: 2	9.3 (1.7)	73 (19)	Any	Not noted	PDC	Change in A1c
Kindmalm (2007)	Sweden	RCT	6 health centers (Skaraborg, Sweden)	102	1 y	64 (NR)	NR	NR	NR	Not noted	Not noted	Similar to MPR	A1c levels at the end
Magny-Normilus (2021)	United States	RCT	SureScripts (2012 - 2013) (Boston, MA)	122	4 mo	I:65 (11)C:64 (12)	I: M: 42 F: 58 C: M: 27 F: 73	I: 8.4 (2.0)C: 8.5 (2.3)	NR	Any	Partial waste assumed	aMPR	Change in A1c
Slabaugh (2015)	United States	RCT	Humana (MAPD plan) (2009-2012) (United States)	1,335	2 y	P:69.4 (8.6)V: 70.1 (8.6)	P: M: 53 F: 47 V: M: 57 F: 43	NR	NR	Basal insulin	Not noted	aMPR PDC	Change in A1c
Wei (2014)	United States	RCT	IMPACT (2006-2012); Humana (2006-2012) (United States)	2,809	1 y	60 (NR)	M: 50 F: 50	IMPACT: 8.6 (1.7); Humana: 8.2 (1.5)	NR	Not noted	Not noted	aMPR	Change in A1c

^a This represents the patients receiving insulin who are represented in the systematic review/meta-analysis.

^b This represents the whole cohort described in the study, including but not limited to the ones receiving insulin.

^c The MPR was noted as being adjusted if any modifications to a standard MPR calculation was made. A = adherent; aMPR = adjusted medication possession ratio; BBI = basal-bolus insulin; DARTS = Diabetes Audit and Research in Tayside; C = control group; DET-C = continued insulin detemir; F = female; GLA-C = continued insulin glargine; A1c = hemoglobin A1c; HIRD = HealthCore Integrated Research Database; I = intervention group; M = male, MAPD = Medicare Advantage with Prescription Drug; MEMO = Medicine Monitoring Unit; mo = months; MPR = medication possession ratio; NA = nonadherent; NR = not reported; P = pen group; PBM = pharmacy benefit management; PDC = proportion of days covered; PMX = premixed; RAI = rapid-acting insulin; RCT = randomized controlled trial; V = vial group; VA = Veterans Administration; y = years.

A1c was the only glycemic measure reported. Multiple prescription fill measures were used. Measures used included MPR or adjusted MPR (n = 5),^19-21,24,25 PDC (n = 2),^8,23 and other calculated measures (n = 3).^7,18,22 One study reported both MPR and PDC.26 MPR is the “total day supply in a defined period” divided by the “number of days in the defined period.”¹⁰ PDC is the “days supplied between first and last day of a time period” divided by the “days between first and last day of a time period.”¹⁰ From the other calculated measures, 2 reported the equation used, which was noted to be similar to MPR (Table 2).^7,18 For the remaining, 1 used “interfill periods” (Table 2).²² There was no consistent definition for an adjusted MPR (Table 2 and Table 3).^{19,20,21,24,25,26} For 1 study, the time between fills was used rather than the days supplied.¹⁹ For 2 studies, this value was adjusted because the value was capped at 100%.^20,24 For the remaining 3, the value was adjusted for insulin package sizes.^21,25,26

TABLE 2.

Subgroup Analyses

Subgroup analyses	No. studies	No. participants	Effect size (95% CI)	I 2 %
Studies (adherence as a categorical variable)
N/A
Studies (adherence as a continuous variable)
1. By adherence definitions
MPR	5	10,882	−0.03 (−0.05 to −0.01)	33.73
PDC/adjusted MPR	1	951	−0.21 (−0.32 to −0.10)	N/A
Not specified/other
2. By insulin type
Long-acting	2	3,778	−0.02 (−0.03 to −0.01)	0
Short-acting	1	760	−0.05 (−0.08 to −0.01)	N/A
All insulins/not noted	3	7,295	−0.11 (−0.22 to −0.00)	81.66

MPR = medication possession ratio; N/A = not applicable; PDC = proportion of days covered.

Studies either reported adherence as a categorical or as a continuous variable. Three studies evaluated adherence as a categorical variable, comparing A1c between adherent and nonadherent groups.^7,8,22 Two of those studies used an adherence cutoff of greater than or equal to 80%.^7,8 The remaining study required that patients had a sufficient supply 100% of the time.²² Eight studies evaluated adherence as a continuous variable and reported the change in A1c based on the change in adherence^18-21,23-26; however, only 7 of these could be included in the quantitative synthesis.^{18,19,21,23,24,26,27} The remaining study did not report the number of patients receiving insulin included in the analysis.²⁰

QUALITY OF THE INCLUDED STUDIES

A detailed description of the study quality assessments is presented in Supplementary Table 4 ^{(648.8KB, pdf)} . Regardless of the original study intent, each quality assessment was performed with the exposure being insulin adherence and the outcome being glycemic control. The RCT did not directly compare the effect of adherence on A1c, so this was evaluated methodologically as a cohort study using the Newcastle-Ottawa Scale, as the studied intervention was a discharge program, not adherence (Supplementary Table 4 ^{(648.8KB, pdf)} ).²⁴ Five studies were noted to be at a low risk of bias.^{8,19,20,23,26} Five were noted to be at a moderate risk of bias.^{18,21,22,24,25} One study was noted to have a high risk of bias.⁷ Comparability and A1c follow-up were the aspects most affecting bias. Only a few of the studies matched the patient cohorts or adjusted the analyses for confounders. For most studies, A1c was only reported for a proportion of the initial cohort (Supplementary Table 4 ^{(648.8KB, pdf)} ).

ADHERENCE MEASURED AS A CATEGORICAL VARIABLE

Three studies involving 6,077 patients were included in the meta-analysis comparing A1c MD between insulin-adherent and nonadherent patients (Figure 2).^7,8,22 Pooled estimates of the 3 studies demonstrated a statistically significant association between being adherent and A1c MD of −0.49% (95% CI = −0.76% to 0.22%, P < 0.05; I² = 40.7%). The funnel plot showed mild asymmetry; however, the Egger regression test (P = 0.56) showed no evidence of small-study effects (Supplementary Figure 1 ^{(648.8KB, pdf)} ).

FIGURE 2.

Pooled Results (Adherence as a Categorical Variable)

ADHERENCE MEASURED AS A CONTINUOUS VARIABLE

Seven studies involving 13,665 participants were included in the meta-analysis evaluating the effect of a change in adherence on A1c (Figure 3).^{18,19,21,23-26} Pooled estimates of the 7 studies demonstrated a statistically significant association between a 1% increase in adherence and a mean reduction in A1c of −0.04% (95% CI = −0.06% to −0.01%; P < 0.05; I² = 72.4%). Based on the visual inspection of the funnel plot (Supplementary Figure 2A ^{(648.8KB, pdf)} ) and contour-enhanced funnel plot (Supplementary Figure 2B ^{(648.8KB, pdf)} ), as well as on quantitative measurement that used the Egger regression test (Supplementary Figure 2C ^{(648.8KB, pdf)} ) (P = 0.002), there are small-study effects given the regression test and the asymmetrical nature of the funnel plot.28 Even after removing a study because of heterogeneity, the contour-enhanced funnel plot and Egger regression test showed signs of publication bias (P = 0.048) (Supplementary Figure 2D ^{(648.8KB, pdf)} , Supplementary Figure 2E ^{(648.8KB, pdf)} ).²³

FIGURE 3.

Pooled Results (Adherence as a Continuous Variable)

OTHER SUBGROUP ANALYSES

The findings from additional subgroup analyses are presented in Table 2. When only looking at MPR, the association remained statistically significant, with decreased heterogeneity (−0.03%, 95% CI = −0.05% to −0.01%; P < 0.05; I² = 33.7%). For the subgroup analysis for long-acting insulin, the association was statistically significant with no heterogeneity (−0.02, 95% CI = −0.03% to −0.01%; P < 0.05; I² = 0%). When the insulin type was not noted, the association was not statistically significant, with increased heterogeneity compared with the primary analysis (−0.11%, 95% CI = −0.22% to −0.00%; P = 0.05; I² = 81.7%).

SENSITIVITY ANALYSES

Adherence Measured as a Categorical Variable. After sensitivity analyses were performed, the findings remained consistent (Supplementary Table 5 ^{(648.8KB, pdf)} ). Kindmalm et al was the only study meeting any one of the criteria for the sensitivity analyses: high risk of bias, small sample size, unclear adherence definition, or study using paper prescription fill records.⁷ When removed, the resulting MD was −0.43% (95% CI = −0.78% to −0.08%; P < 0.05; I² = 59.1%).

Adherence Measured as a Continuous Variable. After excluding studies with a population overlap (−0.05%, 95% CI = −0.08% to −0.02%; P < 0.05; I² = 71.5%), small-sized studies (−0.03%, 95% CI = −0.05% to −0.01%; P < 0.05; I² = 73.5%), and studies with unclear adherence definitions (−0.04%, 95% CI = −0.06% to −0.01%; P < 0.05; I² = 73.6%) (Supplementary Table 5 ^{(648.8KB, pdf)} ), the finding remained robust.

Discussion

Insulin adherence, when measured by prescription fill measures, was associated with improved glycemic control. When measuring adherence as a categorical variable, adherent patients had a 0.49% reduction in A1c compared with nonadherent patients, which is comparable to the reduction in A1c associated with adherence to most oral antidiabetic medications, reported to be between 0.5% and 1.25%.²⁹ Although publication bias was assessed, there were a small number of trials, so both the funnel test and the Egger test were limited in their capacity to detect bias.¹⁶ Correspondingly, when analyzing adherence as a continuous variable, a 1% increase in adherence was associated with a 0.04% reduction in A1c. Assuming a linear relationship, a 10% increase in adherence, which is 3 more days of insulin coverage in a 30-day period, would be associated with a 0.4% decrease in A1c. However, the heterogeneity found in the analysis was high. The myriad of insulin adherence measurements included contributed to the high level of heterogeneity found. For example, adjusted MPR, interfill periods, and adjusted PDC were used in the studies found. Heterogeneity was reduced when subgroup analyses for prescription fill measures were performed. A standardized insulin adherence measurement could help compare results across studies or draw meaningful conclusions about the effectiveness of different insulin regimens to improve adherence, as this could provide a common language.

Our systematic review identified 2 commonly used measures of insulin adherence: PDC and MPR. PDC distinguishes itself from MPR by accounting for nonpersistence within the measurement period, whereas MPR solely considers medication quantity dispensed.³⁰ The PDC is then constrained to a maximum of 100%, whereas the MPR can exceed this threshold. Consequently, the PDC offers a more conservative and comprehensive estimate of adherence, particularly when considering real-world circumstances, such as medication switching or insulin expiration.³⁰ Once used, each insulin vial or pen typically has a shelf life of 28 days, unlike oral tablets, which have a year-long expiration once dispensed. This characteristic holds true for both long-acting and short-acting insulin formulations, mitigating concerns regarding dosing variability. Although MPR calculations may assume an extended lifespan for insulin vials, PDC mitigates this extrapolation, providing a more accurate reflection of adherence patterns. Although PDC is ideally suited for measuring insulin adherence, MPR may remain prevalent because of its simplicity in calculation.³¹

Another aspect of adherence that should be considered would be the threshold at which patients would be considered to be adherent. Conventionally, a threshold of 80% or higher, regardless of the clinical context, is defined as being adherent.³² Many previous studies show that being adherent (a threshold of ≥80%) is clinically and financially meaningful. Roebuck et al showed that annual medical spending was significantly decreased by $4,413, providing a benefit-cost ratio of 6.7:1 if diabetic individuals were adherent to their oral antidiabetic medications (MPR ≥ 80%).³³ Another article from Roebuck et al found that adherence (PDC ≥ 80%) to chronic disease medications, including oral antidiabetic medications and decreased hospitalization and emergency department visits by 8%-26% and 3%-12%, respectively.³⁴ Additionally, low and moderate levels of adherence were also associated with decreased overall health care costs.³⁴ However, some findings have indicated that the cutoffs may be disease-specific or based on pharmacological classes.³⁵ A retrospective analysis conducted by Lim et al showed that the optimal cutoff points for adherence to oral antidiabetic medications among patients with type 2 diabetes may be slightly higher than the conventional threshold of 80%, especially if PDC was used as an adherence measure if there was a shorter assessment period and if there was a stricter A1c goal.³⁶ Hence, the adherence threshold used for insulin needs to be carefully explored to ensure that it is clinically meaningful. Although 2 of the studies in our analysis used a threshold of 80%, the remaining study used a threshold of 100% and comprised most of the weight for the categorical analysis. If the current cutoff point is not optimal for insulin, it should be investigated further.

LIMITATIONS

All the studies included in our meta-analysis were evaluated as nonrandomized studies. These observational studies have many inherent limitations, such as the ability to infer causality, as these are subject to multiple confounders. For example, Hood et al showed a greater magnitude of A1c change, possibly because of U-500, a more potent insulin, being initiated after the baseline adherence measurement in this pre-intervention and post-intervention analysis.²³ Future studies could consider the use of insulin vials/pens as an instrumental variable for the causal relationship between insulin adherence and glycemic control.

Furthermore, our study did not evaluate a particular measure of insulin adherence; it is unclear what insulin adherence measurement would be best. There is no standardized measure for insulin adherence established, like there is for oral antidiabetic agents.¹⁰ Managed care databases are not able to capture actual physician insulin prescriptions since prescribers may adjust insulin doses, especially on initiation. For example, Hood et al had to use the dispensed total daily dose rather than the prescribed total daily dose to calculate the days’ supply.²³ However, in the analysis, they only included patients with high-dose insulin regimens in both pre-index and post-index periods to better reflect real-world high-dose regimens. Second, the heterogeneity found in the analysis was high, in part because of the different prescription fill measures. One study by Slabaugh et al showed multiple available insulin adherence measures, including MPR, PDC, and adjusted PDC.²⁶ The adjusted PDC was included in our primary findings, but the other 2 were used in the other analyses. It should be acknowledged that all prescription fill measures share a common limitation. A paid claim does not confirm actual use of the medication; however, this risk may be low in patients with diabetes requiring insulin. Prescription fill measures remain widely used; they are adopted by the Centers for Medicare & Medicaid Services as a quality measure.³⁷

Third, some studies included in our analysis were not designed to assess the association between prescription fill and glycemic control. For example, Ayyagari et al aimed to investigate whether the choice of insulin delivery device (pen vs vial/syringe) has differences on effectiveness and adherence,²² whereas Magny-Normilus et al investigated whether intensive discharge interventions among patients with T2DM were associated with better glycemic control.²⁴ Subsequently, for our study aim, these studies were potentially subject to multiple confounders. Although 6 of our 11 studies had a moderate or high risk of bias, most of the reported study characteristics, such as health plan type or geographical region, between the groups were not thought to be significant confounders.^{7,18,21,22,24,26} Still, many factors could affect glycemic control and distort the association with insulin prescription fill measures. For example, since treatment decisions may change over time in response to patient outcomes, time-dependent confounding may influence insulin adherence and outcomes. Moreover, patient characteristics may also change over time in the longitudinal setting. Some studies adjusted for this but not all; for example, Ayyagari et al did use marginal structural models to adjust for time-dependent confounders.²² Although variables that confound the relationship between prescription fill measures and glycemic control may not have always been captured, the quality of data from these studies were still considered to have been satisfactory, hence the inclusion of the studies in our review.

Conclusions

Our study shows a statistically and clinically significant relationship between insulin adherence, based on prescription fill records, and A1c. This suggests that (1) improvements in insulin adherence reflect an overall improvement in T2DM outcomes and (2) changes in insulin adherence may potentially be used to estimate changes in A1c when A1c data are not present. Various policies in the public and private sectors have been introduced to mitigate cost-related insulin adherence; however, the short-term and long-term outcomes of these interventions are unclear.^4,6 This association can make use of claims databases to estimate A1c in the short-term and potentially be used in economic modeling to estimate long-term effects. It should be noted that these claims databases do not capture those who are self-insured or accessing insulin by other means. Still, economic evaluations have already been published using insulin PDC, and our association can provide a more comprehensive estimate for use in these studies.^38,39

Future research should focus on establishing a standardized insulin adherence measure and adherence threshold. Medication nonadherence is fairly common, especially among patients with chronic diseases. Investigators should further explore the reasons for insulin nonadherence and develop effective methods to minimize nonadherence issues. Our study is a starting point to determine which levels of insulin adherence could be appropriate and how much effort into adherence improvement could help patients achieve a specific A1c goal. Providers can encourage patients to improve insulin adherence based on their current adherence level and goals of glycemic control using our pooled association.