Healthcare costs of fast-acting insulin analogues versus short-acting human insulin for Danish patients with type 2 diabetes on a basal–bolus regimen

Abstract

Aims:

Fast-acting insulin analogues (FAIAs) reduce hypoglycaemia and improve administration flexibility compared with short-acting human insulin (SHI). This analysis examines whether these benefits translate into cost offsets when comparing the total treatment costs for FAIA versus SHI used as basal–bolus therapy for treating type 2 diabetes (T2D).

Methods:

Registry data covering the Danish population including demographic variables, prescription, hospital and primary care data formed the basis for analysis. To capture patients on basal–bolus therapy only, inclusion criteria were ≥2 prescriptions of either long-acting insulin analogues (LAIAs) or neutral protamine Hagedorn (NPH) insulin (basal component), and ≥2 prescriptions for either an FAIA or SHI (bolus component) during the inclusion period (1 January–31 December 2005). Patients using LAIAs (n = 521) or NPH (n = 2695) were analysed separately. Within each basal cohort, patients using FAIAs or SHI were matched regarding observable variables using propensity scores. Healthcare costs were analysed for a follow-up period (maximum 2 years post-inclusion).

Results:

Within each cohort, matching produced groups with similar observed covariates. Overall direct healthcare costs in the LAIA cohort were €4183 and €5289 for FAIA and SHI, respectively. In the NPH cohort, costs were €4940 and €4699 for FAIA and SHI, respectively. For both basal cohorts, cost differences between FAIA and SHI were not statistically significant.

Limitations:

As the propensity score model cannot account for unobserved variables, conclusions of causality cannot be made. Moreover, exclusion of indirect costs and application of hospital contact charges accrued in the discharge year only may result in an underestimation of overall healthcare costs.

Conclusion:

Using matched cohorts, treating patients with T2D using basal–bolus regimens containing FAIAs was no more costly to the Danish healthcare system than regimens using SHI. FAIAs provide a flexible administration and optimal glucose control for a similar cost.

Keywords::

• Basal
• bolus
• Costs
• Healthcare
• Registers
• Type 2 diabetes

Introduction

The fast-acting insulin analogues (FAIAs) insulin aspart, insulin lispro and insulin glulisine have enhanced the treatment options for patients with diabetes compared with short-acting human insulin (SHI). In general, FAIAs marginally improve glycaemic control and reduce the occurrence of hypoglycaemic episodes, particularly during the night1–6. Moreover, the rapid absorption of FAIAs means they effectively reduce postprandial glucose even when administered immediately before or after eating7,8. In contrast, SHI requires administration 30 minutes before eating, an interval that may be inconvenient for patients, potentially reducing adherence and thereby resulting in suboptimal glucose control9,10.

While the health benefits of regimens using FAIAs may reduce associated healthcare costs in terms of fewer additional treatment and intervention requirements, FAIAs themselves are more expensive than SHI11. Comparison of all the associated costs for analogue versus human insulin regimens therefore forms an important component of healthcare-sector evaluations and subsequent decisions regarding prescription practice. Only a few direct cost studies analysing the economic impact of FAIAs in a clinical setting have previously been reported12–14. A database analysis encompassing 14 United Healthcare-affiliated health plans across the US found a small but non-significant overall saving of US$216 for insulin lispro compared with SHI12. A subsequent study using data from a large, managed-care organisation in the US also observed a similar non-significant reduction in overall medical costs of US$2327 for insulin lispro compared with SHI, but used a slightly different method of propensity score matching13. However, a US healthcare database study followed a cohort of patients switching to insulin aspart from SHI. Comparison between pre- and post-switch costs found significant post-switch reductions in both total and diabetes-related costs of $2283 (p = 0.0001) and $2028 (p < 0.001) per patient per year, respectively14.

Register-based analyses use real-world patient data, and can illuminate clinical aspects of diabetes treatment that are not captured in traditional clinical studies due to their limited time horizon and rigorous selection of patients. These advantages mean that database studies have often been used as a supplement to clinical studies13,15,16.

The objective of this cost analysis is to use government-maintained registers in Denmark as a source of real-world patient data to compare the direct healthcare costs of patients with type 2 diabetes on a basal–bolus regimen using either FAIAs or SHI in combination with long-acting insulin analogues (LAIAs) or intermediate-acting human insulin (neutral protamine Hagedorn [NPH] insulin). This study is an extension of a previous analysis, comparing LAIAs against NPH insulin in basal insulin-only therapy17, and is the first time that the total treatment costs of FAIAs as part of a basal–bolus regimen have been directly addressed.

Methods

Material

Data were drawn from registers compiled by Statistics Denmark18, covering the total Danish population of 5.4 million inhabitants. Data covering characteristics and healthcare utilisation were extracted for patients on a basal–bolus regimen between 2002 and 2007. Patient characteristics included variables such as age, marital status, employment status, education and gross income.

The Danish Register of Medicinal Product Statistics19 provided information on outpatient prescription purchases but not hospital or over-the-counter (OTC) transactions. The available data include transaction dates, pharmaceutical classification codes, pharmacy selling price and the overall transaction price (comprised of the pharmacy selling price plus the prescription fee). The overall price was used as the cost estimate (this includes copayment by patients).

Hospitalisation cost data from the National Registry of Patients20 included type of hospital contact (inpatient stay, ambulatory and emergency room visits) and diagnosis codes (available from 2002 to 2006). Hospital contact costs were estimated using charges from the Danish Diagnosis-Related Group’s case-mix system, a case-based classification system and its associated funding model. Costs of consultations with the GP or medical specialist were extracted from the National Health Insurance Registry (available from 2002 to 2006)21.

Patients

Patients with type 2 diabetes were defined as those having received at least one prescription for oral antidiabetic drugs (OADs) at any time from 1995 onwards. The inclusion period for analysis lasted 12 months (1 January to 31 December 2005). Patients receiving at least two prescriptions of LAIAs in the inclusion period were allocated to the long-acting insulin analogue cohort (the LAIA cohort). Patients receiving at least two prescriptions of NPH insulin and no LAIAs were allocated to the NPH cohort. The date of the first prescription in the inclusion period was defined as the index event (for current insulin users, the index event was the first prescription refill in 2005; for patients receiving insulin for the first time, the index event was the first basal insulin prescription filled). Exclusion criteria were used to construct comparable cohorts that were not confounded by switches to other regimens. Thus, patients who switched from LAIAs to NPH insulin or vice versa after the index event were excluded. Likewise, patients who used premix insulin or who switched to premix insulin after the index event were excluded from analysis, as premix insulins consist of both a basal insulin component and a fast-acting insulin component in a single formulation and so their use cannot be considered a basal–bolus regimen. Including premix insulin use would have introduced additional variables to the analysis. The remaining patients were divided further into basal-only (previous analysis21) or basal–bolus therapy subgroups. To minimise the potential of confounding effects from the basal insulin component, users of LAIAs and NPH insulin were analysed as separate cohorts.

Patients were defined as being on a basal–bolus regimen if they had at least two prescriptions of prandial insulin in the inclusion period. Each basal cohort was further subdivided according to whether they were receiving FAIAs (the FAIA group) or SHI (the SHI group). Patients were allocated to the FAIA group if they had prescriptions with FAIAs exclusively during the analysis period, and vice versa for the SHI group. The analysis period was the time from the index event to 31 December 2006, death or migration, whichever came first; thus the maximum analysis period was 2 years. Patient selection and inclusion/exclusion criteria for the basal–bolus subgroup are illustrated in Figure 1.

Figure 1.  Patient selection. A total of 42,386 patients with prescriptions for long-acting insulin analogues or NPH insulin in the index period (2005). Various exclusion criteria: no switches in the follow-up period, at least two prescriptions in the index period, no migration between 2004 and follow-up, no patients with missing socio-economic variables. FAIA, fast-acting insulin analogue; LAIA, long-acting insulin analogue; NPH, neutral protamine Hagedorn insulin; OAD, oral antidiabetic drug; SHI, short-acting human insulin; T1D, type 1 diabetes (no OAD prescription between 1995 and 2007); T2D, type 2 diabetes (OAD prescription between 1995 and 2007).

Statistical analyses

Separate statistical analyses were conducted for users of either LAIAs or NPH insulin. However, in both cases, the propensity score method was used to match pairs of similar patients based on observed covariates. The propensity score-matching technique allows the creation of groups with a similar distribution of observed covariates22–25. The propensity scores were estimated by a logistic regression model for predicting the probability (the propensity) of using FAIAs. To control for socio-demographic and disease-specific variables, the models included age and gender, the presence of diabetic complications or acute diabetes-related events evaluated prior to the inclusion period, and GP costs for the previous year. A list defining the included diabetic complications can be found in Supplementary data, Table S1. Socio-economic variables such as income, education, employment and origin were also added to the specification if the added variables could improve the model with significant coefficients. For both the LAIA cohort and NPH cohort, each patient from the FAIA group was randomly matched with replacement to a patient from the SHI group whose propensity score deviated by no more than half a percentage point. For the analyses of the NPH cohort, a match could be found for all the patients in the FAIA group. In the case of the LAIA cohort, a few patients from the FAIA group without a match remained, and so random matching between propensity scores deviating by less than one percentage point was attempted. As the SHI group within the LAIA cohort comprised only 84 patients, they were used repeatedly to generate sufficient matches for the much larger FAIA group. This method enabled pairing of all but seven patients from the FAIA group with a patient from the SHI group for the LAIA cohort, while in the SHI group five patients were not used for matching. Patients who could not be matched were excluded from the study groups.

In the case of the LAIA cohort where the number of patients in the SHI group is very small, expanding the original sample size by matching with replacement does not alter the level of statistical uncertainty. Therefore, it was decided to use non-parametric bootstrapping for analysis of the LAIA cohort, rather than standard parametric methods to simulate confidence intervals. The bootstrapping technique can be used when the distribution characteristics of a parameter are unknown26–28. The parameter's distribution is estimated by repeatedly resampling the original data set of 521 patients with replacements. For the present study, 1000 replications were performed. Each replication included matching and healthcare cost calculations. Confidence intervals were estimated from the bootstrapped standard errors26–28. In contrast, the NPH cohort comprised sufficient patient numbers to allow the use of standard parametric t-statistics.

Direct healthcare costs in the analysis period were estimated as the sum of all prescription costs, GP fees and hospital/ambulatory charges in the study period for each patient and adjusted to annual costs, as total length of participation in the study varied between patients. As the analysis period had a maximum duration of 2 years, the costs were not discounted.

Results

A total of 622 patients were identified as receiving basal treatment with LAIAs and at least two prescriptions of either FAIAs or SHI. Of these, a total of 437 and 84 patients could be allocated to the FAIA group and SHI group, respectively (Figure 1). A total of 2946 patients were identified as receiving basal treatment with NPH insulin and at least two prescriptions of either FAIAs or SHI. Of these, a total of 888 and 1807 patients could be allocated to the FAIA group and SHI group, respectively (Figure 1). These patients from the LAIA cohort or NPH cohort formed the data material for the propensity score models.

Key characteristics for the groups before and after matching are shown in Table 1. For the LAIA cohort, the patients in the SHI group compared with the FAIA group were more likely to be older and female; furthermore, there was a greater tendency towards experiencing acute events prior to follow-up. Propensity score matching resulted in equal mean propensity scores for the two groups for both the LAIA and NPH cohorts. The simulated matches were equal with respect to most characteristics, with the exception of diabetes complications among the LAIA cohort patients, where the matching caused a bigger differential between the two groups than was present originally. Local registry regulations prevented analysis of data for insulin aspart, insulin lispro and insulin glulisine separately; however, according to official statistics, the proportion of patients receiving insulin aspart as their FAIA during 2006 was approximately 90%, while the remaining 10% of patients received insulin lispro29.

Table 1.  Key characteristics before and after simulated matches.

	FAIA		SHI
	All	Matched	All	Matched	Simulated*
Patients on LAIA basal
Number of patients	437	430	84	79	430
Mean propensity score	84%	85%	81%	82%	85%
Proportion of women	48%	47%	62%	59%	45%
Mean age	53.5	53.5	57.1	57.4	54.2
Acute events (prior to follow-up)	10%	10%	18%	15%	7%
Diabetes complications (prior to follow-up)	8%	8%	8%	8%	4%
Patients on NPH basal
Number of patients	888	888	1807	677	888
Mean propensity score	35%	35%	32%	34%	35%
Proportion of women	41%	41%	44%	41%	41%
Mean age	54.4	54.4	57.4	55.6	55.0
Acute events (prior to follow-up)	4%	4%	4%	4%	4%
Diabetes complications (prior to follow-up)	15%	15%	19%	16%	15%

*Simulated number based on 84 LAIA patients and 1807 NPH patients.

FAIA, fast-acting insulin analogue; LAIA, long-acting insulin analogue; NPH, neutral protamine Hagedorn; SHI, short-acting human insulin.

Table 2 shows the results of the logistic regression models used to estimate the propensity of being a patient on FAIAs, based on socio-demographic variables and diabetes-related complications. For the model regarding the LAIA cohort, women had significantly lower odds of being in the FAIA group (odds ratio [OR] = 0.759, p = 0.028). In the NPH cohort, patients above the age of 60 had lower odds of being in the FAIA group (OR = 0.790, p = 0.001). For both cohorts, the only disease-specific characteristic to show a statistically significant effect, albeit a minor one in both cases, was the use of GP services prior to the inclusion period (OR = 1.000, p = 0.042 and OR = 1.000, p = 0.026 for the LAIA and NPH cohorts, respectively). For the NPH cohort model, income and origin variables were also incorporated. High income increased the odds of being in the FAIA group (annual income DKK 300,000+: OR = 1.228, p = 0.013), while being a foreigner (defined as an immigrant or a descendant of immigrants) decreased the odds (OR = 0.776; p = 0.011).

Table 2.  Logistic regression models predicting the probability of being a patient using fast-acting insulin analogues.

	Estimate	SE	Odds ratio	*p-value
Patients on LAIA basal (n = 521)
Intercept	1.799	0.228		<0.001
Female	−0.276	0.125	0.759	0.028
Age 40–59 years	−0.196	0.186	0.822	0.292
Age 60+ years	−0.269	0.198	0.764	0.175
Diabetes complications (prior to follow-up)	0.156	0.456	1.169	0.733
Acute events (prior to follow-up)	−0.309	0.166	0.734	0.064
GP costs (prior to follow-up)	0.000	0.000	1.000	0.042
Patients on NPH basal (n = 2695)
Intercept	−0.679	0.149		<0.001
Female	−0.022	0.044	0.979	0.623
Age 40–59 years	0.041	0.061	1.041	0.503
Age 60+ years	−0.235	0.069	0.790	0.001
Income DKK 100,000–199,000	−0.139	0.067	0.871	0.038
Income DKK 200,000–299,000	0.093	0.075	1.097	0.220
Income DKK 300,000+	0.205	0.082	1.228	0.013
Being a foreigner	−0.253	0.099	0.776	0.011
Diabetes complications (prior to follow-up)	−0.180	0.117	0.836	0.123
Acute events (prior to follow-up)	0.087	0.106	1.091	0.408
GP costs (prior to follow-up)	0.000	0.000	1.000	0.026

*Wald chi-square statistics.

LAIA, long-acting insulin analogue; NPH, neutral protamine Hagedorn; SE, standard error.

Mean direct healthcare costs during the analysis period (normalised to annual costs) are summarised in Tables 3 and 4 for the LAIA and NPH cohorts, respectively. For the LAIA cohort, most cost components were lower for the FAIA group compared with the SHI group. These costs related to other aspects of healthcare, from inpatient, outpatient and emergency room to GP and specialist costs. In addition, when compared with the SHI group, the costs of LAIAs and other medicines were also lower for the FAIA group, although the bolus insulin was more costly in the FAIA group. The overall annual direct healthcare costs, including prescription medicine, amounted to €5289 in the SHI group and €4183 in the FAIA group; however, based on the estimated confidence intervals, the cost differences failed to reach statistical significance.

Table 3.  Costs in the follow-up period, patients on basal LAIA (Euro, normalised to annual mean costs).

				Bootstrapped statistics
	FAIA (n = 430)	SHI (n = 430*)	Difference (FAIA–SHI)	SE	Lower CI	Upper CI
Outpatient, diabetes related	20.3	50.8	−30.4	15.5	−84.5	23.6
Outpatient, other	477.9	865.1	−387.1	142.2	−1305.1	530.8
Emergency room, diabetes related	0.9	2.1	−1.3	0.8	−4.6	2.0
Emergency room, other	24.6	17.6	7.1	3.9	−7.6	21.8
Inpatient, diabetes related	145.0	157.5	−12.4	64.7	−225.8	200.9
Inpatient, other	1070.1	1509.6	−439.5	299.9	−1878.5	999.6
GP costs	175.7	188.3	−12.6	21.0	−69.9	44.7
Specialist costs	121.6	141.8	−20.1	11.7	−67.8	27.5
Bolus insulin	528.6	486.4	42.1	26.3	−36.9	121.2
Basal insulin	816.4	925.8	−109.4	43.0	−265.5	46.6
OADs	33.0	29.7	3.4	8.2	−23.6	30.3
Other medicine	768.9	914.9	−146.0	54.4	−320.3	28.3
All costs	4183.1	5289.4	−1106.4	459.4	−3459.2	1246.5

*Simulated n based on 84 patients.

CI, confidence interval; Difference, difference between groups; FAIA, fast-acting insulin analogue; LAIA, long-acting insulin analogue; OADs, oral antidiabetic drugs; SE, standard error; SHI, short-acting human insulin.

Table 4.  Costs in the follow-up period, patients on basal NPH (Euro, normalised to annual mean costs).

				t-statistics
	FAIA (n = 888)	SHI (n = 888*)	Difference (FAIA–SHI)	SE	Lower CI	Upper CI
Outpatient, diabetes related	40.4	29.5	10.9	16.6	−16.5	38.3
Outpatient, other	670.5	691.6	−21.1	98.5	−183.3	141.1
Emergency room, diabetes related	0.8	1.3	−0.5	0.4	−1.2	0.2
Emergency room, other	24.4	26.3	−1.9	3.2	−7.1	3.3
Inpatient, diabetes related	252.2	188.7	63.5	78.5	−65.7	192.7
Inpatient, other	1565.7	1383.6	182.1	208.8	−161.7	525.9
GP costs	201.3	211.1	−9.8	10.1	−26.4	6.9
Specialist costs	112.8	107.5	5.3	7.1	−6.4	17.0
Bolus insulin	598.4	527.2	71.3	19.8	38.7	103.8
Basal insulin	464.8	429.6	35.2	16.8	7.5	62.8
OADs	28.1	35.1	−7.0	3.1	−12.1	−1.8
Other medicine	980.1	1067.8	−87.7	52.7	−174.4	−0.9
All costs	4939.5	4699.2	240.3	316.6	−281.0	761.7

*Simulated n based on 1807 patients.

CI, confidence interval; Difference, difference between groups; FAIA, fast-acting insulin analogue; NPH, neutral protamine Hagedorn; OADs, oral antidiabetic drugs; SE, standard error; SHI, short-acting human insulin.

Similarly, for the NPH cohort, the costs of the FAIA group and the SHI group were comparable after the propensity score matching. Except for both the basal and bolus insulins, which were more costly in the FAIA group, none of the cost elements were statistically significant. The overall annual direct healthcare costs, including prescription medicine, amounted to €4940 in the FAIA group and €4699 in the SHI group: this difference was not statistically significant.

Discussion

This is the first register-based analysis to compare healthcare costs for patients with type 2 diabetes treated with basal–bolus regimens using either FAIAs or SHI in combination with LAIAs or NPH insulin. The Danish registers provide a detailed compendium of information that could be used to define detailed patient characteristics, allowing socio-economic and other drivers of treatment choice to be identified. For patients with an equal propensity for using FAIAs or SHI, no significant differences exist for most cost items between the two study groups, irrespective of whether the basal insulin is LAIA or NPH insulin. Although the direct cost of treatment is higher for patients using FAIAs compared with SHI, the higher cost of insulin is offset by lower expenses elsewhere in the healthcare system, particularly in relation to hospital-related care. The faster absorption, and the fact that FAIAs can effectively reduce postprandial glucose even when administered immediately before or after eating, could potentially have improved adherence, thereby resulting in better glucose control compared with SHI9,10, which may explain why the overall costs were not found to be higher for the FAIA cohort compared with the SHI cohort, despite the fact that FAIAs are more expensive than SHI. The study is an observational study and is furthermore too short in duration to show any long-term consequences of the two treatments.

Difficulties encountered

To compensate for the lack of randomisation, propensity score matching was used. This is a statistical technique that attempts to reduce selection bias in such situations. Repeated matching allows the best match to be used in each case22, producing groups displaying a similar distribution of the observed covariates24. In this study the simulated matches were equal with respect to most characteristics, with the exception of diabetes complications among the LAIA cohort patients, where the matching caused a bigger differential between the two groups than was originally present. This outcome, which is coincidental, does not invalidate the comparison. On the contrary, it makes the comparison more conservative as the SHI group starts out with a lower base-case diabetes progression. The propensity score method, on the other hand, is dependent on the validity of the regression model and it is possible that other unobserved covariates, such as measurements of baseline glycated haemoglobin level or body mass index, might enhance model robustness. Although the propensity score model eliminates the bias of observed variables, not all relevant confounders have been taken into account and the results should not be interpreted as causal effects.

The major difficulty encountered in registry-based analyses is that the results are dependent on the quality and quantity of data available. Firstly, patient identification is sensitive to potential miscategorisations. In this study, for example, patients with type 2 diabetes who did not use OADs during the previous 10 years would have been missed. Additionally, patients with type 1 diabetes were not explicitly excluded and, although current treatment guidelines in Denmark do not indicate OADs in type 1 diabetes, it is possible that the sample included such individuals. However, these potential concerns should not be overstated. The administrative databases used are managed by institutions within the Danish government and are internationally regarded as high-quality, reliable databases.

Secondly, the analysis only considered costs incurred by treatment at public hospitals, consultations with a GP or specialist including visits for hypoglycaemia management and prescription medicine. Costs of OTC medicine, psychiatric hospitalisations, dentistry, physiotherapy and privately financed healthcare costs were not included due to either a lack of data or a lack of relevance. In addition, hospital contact charges could only be applied to activities occurring during the year the patient was discharged; since hospital ambulatory contacts can extend far beyond a year, these may have been underestimated. Further to this, the prices, charges and fees used in this analysis are only crude approximations of the true opportunity costs of treatment. Indirect costs related to more qualitative measures such as loss of productivity are excluded as they are not captured by the registry databases used as a source of data in this study.

Alternative methodologies

In an ideal world, the best way to compare FAIAs and SHI would be to conduct a randomised controlled trial (RCT) over a long time horizon and record both efficacy and economic costs at the same time. However, this approach is not feasible as it would be far too expensive and unlikely to attract and retain enough patients over the duration to generate sufficient statistical power. Another approach would be to use simulation modelling, incorporating evidence from various sources including RCTs. Modelling can allow the long-term benefits to be predicted, but its usefulness is reliant upon the quality of the source data. Lack of detail or uncertainty around the extrapolation, for example, may adversely affect the robustness of such a model. The register study is therefore a good alternative to either approach because, although it is not randomised, it does represent the situation for the actual patient population and matching can be used to simulate randomisation. In theory, registry-based studies allow analysis over years or even decades, although in this case the length of the analysis period was limited to a maximum of 2 years.

New questions arising from the study

Due to registry restrictions, this analysis was unable to compare the costs of individual FAIAs. According to the figures available for 2006, around 90% of FAIA prescriptions were for insulin aspart, with around 10% for insulin lispro29. It would be interesting to repeat this analysis in populations using defined FAIAs to determine whether any differences between individual analogues exist.

Conclusion

This first registry-based analysis of the costs of FAIAs as part of a basal–bolus regimen found no evidence that using FAIAs was more expensive than patients using SHI, irrespective of the type of basal insulin used.

Transparency

Declaration of funding

This study was funded by a grant from Novo Nordisk A/S, Denmark.

Declaration of financial/other relationships

T.E.C. and T.P. are employees of Novo Nordisk A/S. J.G., COWI A/S, has received funding from Novo Nordisk to work on the registry-based study. J.G. is an employee of Novo Nordisk A/S as of 1 February 2011.

Acknowledgements

The authors acknowledge assistance from Watermeadow Medical UK in literature searching and manuscript preparation, funded by Novo Nordisk A/S.