Identifying patient subgroups who benefit most from a treatment: using administrative claims data to uncover treatment heterogeneity

Abstract

Objectives:

To illustrate how claims data can be used to (1) develop outcome scores that predict response to a traditional treatment and (2) estimate the economic impact of individualized assignment to a newer treatment based on the outcome score. An example application is based on two treatments for attention deficit hyperactivity disorder (ADHD): osmotic-release oral system methylphenidate (OROS-MPH) and lisdexamfetamine dimesylate (LDX).

Methods:

Adolescents with ADHD initiating OROS-MPH (n = 6320) or LDX (n = 6394) were selected from the MarketScan claims database. A model was developed for predicting risk of switching/augmentation with OROS-MPH using multiple baseline characteristics. The model was applied to an independent sample to stratify patients by their predicted risk and, within each stratum, risk of switching/augmentation and ADHD-related total costs were compared between OROS-MPH and LDX patients using inverse probability of treatment weighting.

Results:

The prediction model resulted in substantial stratification, showing risk of switching/augmentation with OROS-MPH ranging from 11.3–42.1%. In the two strata where OROS-MPH had highest risk of switching/augmentation, LDX had significantly lower risk of switching/augmentation than OROS-MPH (by 7.0–8.2%) and lower ADHD-related annual total costs (by $264–$625 per patient).

Limitations:

The current study has used the risk of switching/augmentation as a proxy measure for treatment efficacy to establish the prediction model. Future research using a clinical measure for ADHD symptoms is warranted to verify the findings.

Conclusions:

Combining multiple patient characteristics into a predicted score for treatment outcomes with a traditional treatment can help identify subgroups of patients who benefit most from a new treatment. In this analysis, ADHD patients with a high predicted score for switching/augmentation with OROS-MPH had a lower rate of switching/augmentation with LDX. Assigning OROS-MPH and LDX treatments based on the predicted scores that are heterogeneous in a patient population may help improve clinical outcomes and the cost-effectiveness of care.

Keywords::

• Comparative effectiveness
• Stratified medicine
• Patient subgroups
• Attention deficit hyperactivity disorder
• Methylphenidate
• Amphetamine

Introduction

Comparative effectiveness research has traditionally focused on the average effects of treatments in relatively broad patient populations. However, patients often have heterogeneous responses to therapies. As there is generally no one-size-fits-all treatment choice, identifying and assigning treatments to subgroups of patients who are most likely to benefit from them may improve effectiveness and reduce costs1–3. The generation and application of such evidence is often referred to as individualized, personalized, or, more recently, stratified medicine, and holds the promise of improving health outcomes while limiting healthcare costs1,2,4–7.

The role of individualized medicine in clinical practice has become increasingly important. Many of the most advanced examples come from oncology, which has benefitted from a wealth of molecular markers. For example, cetuximab and panitumumab are only recommended for metastatic colorectal cancer patients with the wild-type KRAS gene8,9. Trastuzumab should be used only in HER2-positive breast cancer patients10,11. Identification of heterogeneous patient subgroups can help to tailor treatment selection in clinical practice9. However, the conventional methods for identifying patient subgroups, i.e., pre-specified or post-hoc subgroup analyses, face the dilemma of a priori vs post-hoc specification12–15. Pre-specified subgroup analysis provides statistical rigor but is restricted by the number of subgroups to be tested and offers limited learning from the data because of the limitations on the number of characteristics that could be considered. Although post-hoc exploratory subgroup analysis provides an opportunity to learn from the data, the multiple testing involved makes it prone to low statistical power12,14,16.

This dilemma has been addressed by a personalized medicine approach, which analyzes post-hoc subgroups defined by a risk score combining multiple patient characteristics using clinical trial data5–7. The method systematically selects and combines multiple baseline covariates to define subgroups through a regression model. Compared to separate post-hoc subgroups defined by single characteristics, this regression-based approach limits the potential for selecting patient characteristics based on the researchers’ subjective perceptions and the loss of power due to multiple testing for many characteristics. Although a model is used to combine characteristics, the evaluation of the model and the assessment of its performance for patient stratification are non-parametric. Once established, such models could be used to guide treatment selection for individual patients. The value of assigning treatments based on these models could be evaluated by plotting a selection impact curve that estimates the population response as a function of risk scores estimated from the models and comparing that to the existing practice4,17.

In this study, we have adopted the methods for personalized medicine developed by Cai and colleagues5–7, and the selection impact curve developed by Song and Pepe17 and applied them to the real-world administrative claims data using attention deficit hyperactivity disorder (ADHD) as an example.

ADHD is a common psychiatric disorder with an estimated prevalence among adolescents of 6–10% in the US, 3–9% in the UK, and ∼3% worldwide3,18,19. Psychostimulants, mainly composed of methylphenidate and amphetamine, are recommended as first-line medications for ADHD20,21. Concerta® (Osmotic-release oral system methylphenidate: OROS-MPH) and Vyvanse® (lisdexamfetamine dimesylate: LDX, an amphetamine) are the two most commonly used long-acting drugs from each class22. In the US, OROS-MPH was launched in 2000, and LDX was launched in 2007, when OROS-MPH was considered the market leader for long-acting methylphenidate formulations. The current study aims to identify subgroups of adolescents with ADHD who are most likely to benefit from LDX or OROS-MPH using a stratified medicine method, and to compare the impact of assigning treatment based on stratified medicine vs the real-world practice (the observed treatments with LDX and OROS-MPH in claims) on two outcomes: risk of switching/augmentation and ADHD-related total costs. To better illustrate our method, we also compared the results from stratified medicine with those from pre-defined subgroups analyses.

Methods

Data source

The study was conducted using the Thomson Reuters MarketScan® Commercial Claims and Encounter Research database (2007–2009), a data source from ∼100 payers in all census regions, with 20 million individuals annually23. The database contains information on enrollment, medical claims, and pharmacy claims.

Study cohorts

Adolescents (aged 13–17) initiating LDX or OROS-MPH on or after 06/29/2007 (the launch of LDX in the US) were selected. The earliest prescription date of LDX or OROS-MPH was defined as the index date. Patients were excluded if they filled other prescriptions for ADHD on the index date. Patients were further required to be continuously enrolled in the healthcare plan for at least 6 months before (baseline period) and 12 months after (study period) the index-date and have ≥1 ADHD diagnosis (ICD-9 code of 314.0 x)24 during this 18-month period. Patients were categorized into one of two study cohorts based on the drug received on the index date: OROS-MPH or LDX. In addition, patients were classified into two pre-defined sub-populations based on whether they received any other ADHD medications during the baseline period.

Baseline variables and study outcomes

Baseline variables were assessed during the 6-month pre-index period and included age, gender, baseline use of ADHD medications, comorbidities, as well as all-cause and ADHD-related (i.e., services associated with ADHD diagnosis code of 314.0 x) medical resource utilization (Appendix 1).

Study outcomes, risk of switching/augmentation (used as a proxy measure for treatment efficacy), and ADHD-related total costs were measured during the 12-month study period. Switching was defined as use of another ADHD medication within 30 days of discontinuing the index drug; augmentation was defined as use of another ADHD medication while on the index drug. ADHD-related total costs, based on reimbursements from the managed care plans to providers, included costs for medical services associated with ADHD diagnoses (ICD-9 code of 314.0 x) and ADHD prescription drug (Appendix 1) costs. All costs were estimated from a third-party payer’s perspective and were inflation-adjusted to 2009 dollars.

Statistical methods

The prediction model described below was developed using R software version 2.11.1 (http://cran.r-project.org/), and the remainder of the analyses were performed using SAS® 9.2 (SAS Institute, Inc., Cary, NC). A result was considered statistically significant if the p-value was less than 0.05 (based on two-tailed tests).

Comparisons of study outcomes in the overall population and in the pre-defined sub-populations

Risk of switching/augmentation and ADHD-related total costs were estimated and compared between the OROS-MPH vs LDX cohorts in the overall population using weighted Chi-square and weighted t-tests, respectively, with inverse probability of treatment weighting (IPTW) to adjust for potential confounding25–28. The probability of receiving OROS-MPH and LDX was estimated using a logistic regression with baseline characteristics as the independent variables. Baseline characteristics were re-evaluated after IPTW to assess the balance of the two study cohorts. Within each patient sub-population, study outcomes were compared between OROS-MPH and LDX using IPTW. p-values for trends across the pre-defined sub-populations were estimated using generalized linear models with binomial distribution and normal distributions for risk of switching/augmentation and ADHD-related costs, respectively, adjusting for baseline characteristics using IPTW and index drugs.

Stratified medicine

A stratified medicine approach was used to identify patient subgroups with different risk of switching/augmentation if treated with OROS-MPH vs LDX (Appendix 2). First, OROS-MPH patients were randomly selected to create the training sample, which was comprised of half of the OROS-MPH patients; the remaining OROS-MPH patients and all of the LDX patients formed the validation sample. Based on the training sample, a prediction model was established to estimate risk of switching/augmentation with OROS-MPH including all baseline variables as potential predictors. The Least Absolute Shrinkage and Selection Operator (LASSO)29 method with a 10-fold cross-validation was used to select the best model and generate coefficient estimates for the retained variables. The model was set up to predict the risk of switching/augmentation associated with OROS-MPH because this treatment was approved earlier and was considered the more established treatment paradigm. Next, the model was applied to the validation sample to generate scores predicting risk of switching/augmentation associated with OROS-MPH; patients were then stratified into quartiles based on those scores. The predicted risk was compared to the observed risk associated with OROS-MPH to validate the model. Within each stratum, the observed risk of switching/augmentation and ADHD-related total costs were compared between LDX and OROS-MPH using weighted Chi-square tests or weighted t-tests with IPTW to adjust for baseline differences. A more conservative assessment, accounting for the multiple comparisons across strata, used the Holm method to control the familywise type I error rate30. Lastly, the p-values for trend across quartiles were estimated based on generalized linear models with binomial distribution and normal distributions for risk of switching/augmentation and ADHD-related costs, respectively, adjusting for baseline characteristics using IPTW and index drugs.

Treatment allocation

The ADHD-related total costs (based on 2009 costs) and risk of switching/augmentation were estimated and compared among different treatment allocation strategies:

1. Real-world practice: the observed outcomes without any adjustment.

2. Pre-defined sub-populations: weighted mean outcomes estimated assuming that LDX was assigned to the pre-defined sub-population where its benefit vs OROS-MPH was more pronounced, while OROS-MPH was assigned to the other sub-population.

3. Stratified medicine: weighted mean outcomes estimated assuming that a certain proportion (0%, 25%, 50%, 75%, or 100%) of patients with the lowest predicted risk of switching/augmentation with OROS-MPH were assigned to that treatment; the remainder were assigned to LDX.

In addition, a hypothetical scenario of reducing the pharmacy costs of OROS-MPH by 20% with the remaining costs unchanged was also considered to simulate the potential price decrease due to the recent availability (05/2011) of generic OROS-MPH.

Results

Description of study cohorts

A total of 12,714 adolescents with ADHD who initiated treatment with OROS-MPH (n = 6320) and LDX (n = 6394) were identified. At baseline, patients in the LDX cohort were significantly more likely to use ADHD medications and have all-cause and ADHD-related outpatient and psychotherapy visits (Table 1).

Table 1.  Baseline characteristics for adolescents.

Baseline characteristics	LDX	OROS-MPH	p-value^a
	(n = 6394)	(n = 6320)
Demographics
Age (years) on index date, mean ± SD	14.9 ± 1.4	14.9 ± 1.4	0.3030
Male, n (%)	4440 (69%)	4138 (65%)	< 0.0001*
ADHD medications, n (%)
Amphetamine short-acting	491 (8%)	156 (2%)	<0.0001*
Amphetamine long-acting	2092 (33%)	679 (11%)	<0.0001*
Methylphenidate short-acting	280 (4%)	313 (5%)	0.1253
Methylphenidate long-acting	1041 (16%)	600 (9%)	<0.0001*
Non-stimulant	681 (11%)	428 (7%)	<0.0001*
Any ADHD medication	3589 (56%)	1798 (28%)	<0.0001*
Comorbid conditions,^b n (%)
Any mental health condition	1546 (24%)	1703 (27%)	0.0003*
Asthma	234 (4%)	228 (4%)	0.8753
Neurological disorders	224 (4%)	157 (2%)	0.0008*
Utilization (number of visits), mean ± SD
Inpatient	0.0 ± 0.3	0.0 ± 0.4	0.9074
Outpatient	5.2 ± 5.8	4.9 ± 5.9	<0.0001*
Emergency room	0.2 ± 0.5	0.2 ± 0.5	0.5427
Other medical services	0.4 ± 1.5	0.3 ± 1.4	0.5726
ADHD-related inpatient^c	0.0 ± 0.1	0.0 ± 0.1	0.4713
ADHD-related outpatient^c	1.6 ± 2.6	1.1 ± 1.9	<0.0001*
ADHD-related emergency room^c	0.0 ± 0.1	0.0 ± 0.1	<0.0001*
ADHD-related other medical services^c	0.1 ± 0.5	0.0 ± 0.5	0.0645
ADHD-related psychotherapy^d	0.8 ± 2.4	0.5 ± 1.7	<0.0001*

LDX, lisdexamfetamine dimesylate; OROS-MPH, osmotic-release oral system methylphenidate; SD, standard deviation.

^a p-values were estimated from Chi-square tests for categorical variables and Wilcoxon rank-sum tests for continuous variables.

^b Comorbid conditions included in the study were selected based on a literature review^28–30.

^c Inpatient, outpatient, emergency room, and other medical services were classified based on the places of services. A medical service was considered as ADHD-related if it was associated with an ADHD diagnosis code (ICD-9 code of 314.0x).

^d Psychotherapy visits were defined as visits to a mental health facility or visits associated with psychotherapy CPT codes. To be qualified as ADHD-related, the events need to associate with an ADHD diagnosis code (ICD-9 code of 314.0x)

* p-value significant at α = 0.05.

Comparisons of study outcomes in the overall population and the pre-defined sub-populations

In the overall population, compared with OROS-MPH, LDX was associated with a 4.5% lower adjusted risk of switching/augmentation (18.4% vs 22.9%; p < 0.0001) and $116 lower adjusted annual ADHD-related total costs per patient ($1456 vs. $1572; p < 0.0001). LDX was also associated with significantly lower adjusted risk of switching/augmentation in the pre-defined sub-populations with and without prior treatment (4.6% and 3.4%, respectively) (Figures 1a and b).

Figure 1.  Comparison of study outcomes for LDX vs OROS-MPH^†. (a) Overall population^‡; (b) Pre-defined sub-populations.^‡,$. †Study outcomes were measured over the 12-month post-index period. ‡Mean and p-values were estimated using weighted t-tests for continuous variables. Tests adjusted for baseline characteristics by using IPTW. ^$p-values for trends across the pre-defined sub-populations for risk of switching/augmentation and ADHD-related costs were estimated based on generalized linear models with binomial distribution and normal distributions, respectively, adjusting for baseline characteristics using IPTW and index drugs. Both p-values were <0.001. *p-value significant at α=0.05.

Stratified medicine

Baseline variables retained in the prediction models by LASSO are presented in Table 2. Use of ADHD medications, presence of comorbidity, and frequent use of medical services at baseline were associated with a higher risk of switching/augmentation with OROS-MPH treatment.

Table 2.  Prediction model for switching/augmentation on OROS-MPH.

Baseline characteristics	Odds ratios*
Baseline ADHD medications (Reference = no ADHD medications)
Amphetamine short-acting	1.41
Amphetamine long-acting	1.91
Methylphenidate short-acting	1.75
Methylphenidate long-acting	1.40
Non-stimulant	2.69
Comorbid conditions (Reference = no comorbid conditions)
Any mental health condition	1.09
Asthma	1.28
Neurological disorders	1.25
Baseline utilization (Reference = no baseline utilization)
Number of inpatient visits	1.23
Number of outpatient visits	1.01
Number of other medical service visits	1.03
Number of ADHD-related inpatient visits	1.21
Number of ADHD-related emergency room visits	2.36
Number of ADHD-related psychotherapy visits	1.01
Index Month (Reference = index month not in April–June)
April–June	0.96
Hosmer-Lemeshow statistic of the model	0.2381

OROS-MPH, osmotic-release oral system methylphenidate.

* Odd ratio > 1 indicates the increased probability of experiencing switching/augmentation for each variable compared to the reference category.

Stratifying patients into quartiles based on the predicted risk of switching/augmentation with OROS-MPH resulted in substantial heterogeneity in the observed risk associated with OROS-MPH, ranging from 11.3–42.1% across quartiles. The observed risks with OROS-MPH were similar to the predicted values in each quartile, and the p-value from a Hosmer-Lemeshow test was greater than 0.05, an indication of good predictive power (Table 3).

Table 3.  Comparison of outcomes for LDX vs OROS-MPH. Observed and predicted risk of switching/augmentation.

Quartiles^a			Observed outcome rates
	Predicted outcome rates on OROS-MPH		LDX weighted rates^b			OROS-MPH weighted rates^b			p-value^b,c	Adjusted p-value^d
	Mean	Min–Max	n	Mean	95% CI	n	Mean	95% CI
Quartile 1	13.2%	(12.5%, 13.6%)	1316	8.1%	(6.6%, 9.7%)	1082	11.3%	(9.3%, 13.3%)	0.0114*	0.0456*
Quartile 2	15.0%	(13.6%, 17.5%)	1339	14.2%	(12.2%, 16.2%)	1030	17.2%	(14.8%, 19.7%)	0.0542	0.0542
Quartile 3	20.8%	(17.5%, 23.9%)	1863	19.6%	(17.8%, 21.4%)	529	27.8%	(19.6%, 36.1%)	0.0263*	0.0526
Quartile 4	33.7%	(23.9%, 98.6%)	1876	35.1%	(32.9%, 37.3%)	519	42.1%	(36.7%, 47.5%)	0.0139*	0.0456*

LDX, lisdexamfetamine dimesylate; OROS-MPH, osmotic-release oral system methylphenidate.

^aPatients in the validation sample were stratified into quartiles based on the predicted risk scores of switching/augmentation on OROS-MPH estimated using the prediction model, which was established using the training sample to predict the risk of switching/augmentation associated with OROS-MPH.

^bp-values and weighted rates with 95% confidence interval for the actual treatment received were evaluated within each quartile, adjusting for baseline characteristics using IPTW.

^cp-value for trend across quartiles was also tested and was < 0.0001, estimated based on generalized linear model with binomial distribution adjusting for baseline characteristics using IPTW and index drug.

^dp-values were adjusted using the Holm method to control the familywise type I error rate.

*p-value significant at α = 0.05.

In the quartiles with higher risk of switching/augmentation with OROS-MPH, the risk associated with LDX was lower than OROS-MPH (by 8.2% in quartile 3, p = 0.0263 and 0.0526 with and without adjustment, and 7.0% in quartile 4, p = 0.0139 and 0.0456 with and without adjustment). ADHD-related costs were also lower comparing LDX vs OROS-MPH in these quartiles (by $265 per patient per year in quartile 3, p = 0.0227 and 0.0908 with and without adjustment, and $625 in quartile 4, p = 0.0394 and 0.1182 with and without adjustment) (Table 4). Based on the Whole Acquisition Costs, a 30-day supply would cost $572 for OROS-MPH (54 mg) and $447 for LDX31. The estimated mean costs of index drug in the claims analyses were $636–$845 for OROS-MPH and $626-$873 for LDX across different quartiles (Table 4).

Table 4.  Comparison of outcomes for LDX vs OROS-MPH. ADHD-related costs.

Quartiles and cost outcomes^a	LDX weighted costs Mean ± SD^b	OROS-MPH weighted costs Mean ± SD^b	p-value^b,c	Adjusted p-value^d
Quartile 1
Pharmacy costs	737.6 ± 702.3	754.4 ± 811.7
Index drug costs	626.3 ± 579.9	635.8 ± 761.0
Medical costs	333.3 ± 1225.8	307.9 ± 871.6
Total costs	1071.0 ± 1480.4	1062.3 ± 1278.5	0.8280	1.0000
Quartile 2
Pharmacy costs	843.9 ± 795.9	876.0 ± 970.4
Index drug costs	686.4 ± 689.8	703.7 ± 888.1
Medical costs	522.4 ± 2590.4	489.5 ± 1844.7
Total costs	1366.3 ± 2756.5	1365.6 ± 2165.9	0.9919	1.0000
Quartile 3
Pharmacy costs	989.9 ± 725.3	1194.9 ± 2276.0
Index drug costs	745.0 ± 615.0	844.9 ± 1920.6
Medical costs	460.2 ± 1930.4	520.0 ± 4879.6
Total costs	1450.2 ± 2083.2	1714.9 ± 5409.9	0.0227*	0.0908
Quartile 4
Pharmacy costs	1281.8 ± 980.5	1356.2 ± 2174.2
Index drug costs	873.1 ± 713.3	842.5 ± 1703.8
Medical costs	773.2 ± 5091.8	1323.5 ± 13,705.8
Total costs	2055.0 ± 5223.3	2679.7 ± 13,962.7	0.0394*	0.1182

LDX, lisdexamfetamine dimesylate; OROS-MPH, osmotic-release oral system methylphenidate.

^aStudy outcomes were measured over the 12-month post-index period.

^bMean, standard deviation (SD), and p-values were estimated using weighted t-tests within each quartile, adjusting for baseline characteristics using IPTW.

^cp-value for trend across quartiles was also tested and was <0.001, estimated based on a generalized linear model with normal distribution adjusting for baseline characteristics using IPTW and index drugs.

^dp-values were adjusted using the Holm method to control the familywise type I error rate.

*p-value significant at α = 0.05.

Treatment allocation

Real-world practice was associated with a lower risk of switching/augmentation (19.9% vs 20.2%) and lower ADHD-related total costs ($1481 vs $1495 per patient per year) than treatment allocation based on the pre-defined sub-populations (i.e., assigning patients using baseline ADHD medications to LDX and patients without baseline medications to OROS-MPH). Among all the allocations based on stratified medicine, treating all patients with LDX was the most cost-effective option, with a risk of switching/augmentation of 18.3% and $1449 ADHD-related total costs per patient (Figure 2a). In addition, it was also more cost-effective compared to the real-world practice and treatment allocation based on the pre-defined sub-populations.

Figure 2.  Outcomes achievable with individualized treatment with LDX and OROS-MPH. LDX, lisdexamfetamine dimesylate; OROS-MPH, osmotic-release oral system methylphenidate; ADHD, attention deficit hyperactivity disorder. The squares connected by the dashed black line illustrate different ways of allocating patients based on the stratified medicine approach; black diamond illustrates treatment allocation by pre-defined sub-populations (i.e., sub-populations with and without prior ADHD treatments); white diamond represents the observed real-world practice.

In the hypothetical case in which the pharmacy costs of OROS-MPH were reduced by 20%, the treatment strategy with 50% LDX and 50% OROS-MPH resulted in lower ADHD-related total costs per patient ($1390 vs $1411) and a lower risk of switching/augmentation (19.7% vs 19.9 %) compared to the real-world practice (Figure 2b).

Discussion

Recently developed methods for personalized medicine have been successfully applied to clinical trial data to identify patients who benefit most from a particular treatment5–7,32. We have shown that this method could be adapted to real-world administrative claims data. In the example of ADHD, this approach identified patient subgroups with heterogeneous responses to OROS-MPH and LDX. This is also the first study applying this methodology to psychiatric disorders. When multiple baseline characteristics were combined to stratify patients, a dramatic stratification was produced, with risk of switching/augmentation on OROS-MPH ranging from 11.3–42.1%. Patients who were most likely to switch or augment with OROS-MPH showed significantly lower risk of such events with LDX (by 7–8%), a greater difference than that observed in the overall population and pre-defined sub-populations (∼4%).

Stratified medicine can help inform decision-makers in their attempts to optimize patient care by making tradeoffs between costs and effectiveness. In the current example, real-world treatment assignments were not optimal in terms of either effectiveness or costs compared to treatment allocations based on risk stratification. The value of stratified medicine is exemplified in the hypothetical scenario where the pharmacy costs of OROS-MPH decreased by 20% (a possible scenario, given that generic OROS-MPH is already available). Under this assumption, assigning 50% of the patients who were predicated to have highest risk of switching/augmentation with OROS-MPH to LDX and the rest to OROS-MPH was associated with better treatment outcomes and lower ADHD-related costs than real-world practice.

The stratified medicine method described here could be used to increase the application of real-world observational databases to future comparative effectiveness research intended to optimize patient care. The method could also use measures other than augmentation or switching, e.g., hospitalization or composite benefit-risk measures, as study outcomes. In addition, as alternatives, models could be developed to directly estimate risk scores for differences in responses between OROS-MPH and LDX, or simply for responses to LDX. However, given the established nature of OROS-MPH as a treatment and the potential correlations between responses to the two treatments, it is more informative to stratify patients based on responses to OROS-MPH. Future studies can explore different modeling strategies.

The results of stratified medicine will be beneficial to a variety of healthcare professionals. The importance of identifying and analyzing outcomes heterogeneity in patient subgroups has already been recognized by some healthcare payers and technology assessment authorities such as the National Institute for Health and Clinical Excellence (NICE) and France’s Transparency Commission2,33. For example, a recent briefing paper from NICE emphasizes the importance of including subgroups in cost-effectiveness analyses and decision-making2. Payers could optimize the value of healthcare by allocating treatments based on the stratified medicine results to improve outcomes and reduce costs in different therapeutic areas. Such an approach could help them focus further research efforts on cost-effective interventions based on personalized treatments. The findings from stratified medicine could also be formatted into index cards or computer-based programs to assist physicians and patients in selecting the most beneficial treatments. However, before applying the findings from an observational study of stratified medicine in real-world decision-making, an important next step is to validate these findings using independent data.

The study is subject to certain limitations. First, as in any observational study, there may be unobserved confounding between the two study cohorts. Second, because the claims data have limited measures of clinical symptoms, the risk of switching/augmentation was used as the treatment outcome here. Future research using a real-world database with rich clinical information (e.g., ADHD symptoms relief) is warranted to verify our findings. Since stratified medicine involves the interpretation of treatment effects across multiple strata, adjustment for multiple comparisons can provide a more conservative assessment of statistical significance in exploratory analyses. In the present study, adjustment for multiple testing using the Holm procedure identified fewer statistically significant outcome differences between treatments within strata. Lastly, we used quartiles as an example to illustrate patient stratification. Future studies could use smoothing5,7,34, which can avoid stratification and treatment allocation by a defined number of subgroups. However, because treatments in claims are not randomized, it is important that the adjustment for selection bias be made, or at least assessed, locally, rather than relying on a single propensity score model fit to the full data. Such two-level smoothing, for the localized propensity score and for the localized assessment of IPTW adjusted treatment effects, was beyond the scope of the current study.

Conclusion

The individualization of medical care plays an increasing role in informing treatment selection to improve outcomes and reduce costs. We have illustrated how stratified medicine, through the targeted use of new treatments based on risk stratification of patients using traditional treatments, can be informed by analyses of real-world claims data using ADHD as an example. In the example application, risk stratification identified patients with heterogeneous responses to a traditional agent (OROS-MPH) and indicated how targeted use of a newer agent (LDX) could potentially improve treatment outcomes while reducing the costs compared to current practice.

Transparency

Declaration of funding

This study was funded by Shire Development LLC., Wayne, PA, USA.

Declaration of financial/other interests

J.E.S., H.Y., K.P., K.A.B., J.X., and E.Q.W. have disclosed that they are employees of Analysis Group, Inc., a company that received funding from Shire Development LLC to conduct this study. M.H.E., J.S., and P.H. have disclosed that they are employees of Shire Development LLC. Shire develops and markets psychiatric drugs including treatments for ADHD.

Acknowledgments

No assistance in the preparation of this article is to be declared.

Appendix 1. Baseline medications and comorbidities

Table

Treatments*	Drugs included
Amphetamine short-acting	Amphetamine aspartate/sulfate; dextroamphetamine saccharate/sulfate; methamphetamine hydrochloride; generic amphetamine aspartate/sulfate; generic dextroamphetamine saccharate/sulfate; generic methamphetamine hydrochloride
Amphetamine long-acting	Lisdexamfetamine dimesylate; amphetamine aspartate/sulfate long-acting; dextroamphetamine saccharate/sulfate long-acting; methamphetamine hydrochloride long-acting; generic amphetamine aspartate/sulfate long-acting; generic dextroamphetamine saccharate/sulfate long-acting; generic methamphetamine hydrochloride long-acting
Methylphenidate short-acting	Dexmethylphenidate hydrochloride; generic dexmethylphenidate hydrochloride; methylphenidate hydrochloride; generic methylphenidate hydrochloride
Methylphenidate long-acting	Osmotic-release oral system methylphenidate; methylphenidate hydrochloride long-acting; dexmethylphenidate hydrochloride long-acting; generic methylphenidate hydrochloride long-acting,
Non-stimulants	Guanfacine; clonidine; generic guanfacine short-acting; generic clonidine short-acting; guanfacine long-acting; clonidine long-acting; atomoxetine hydrochloride
Comorbidities	Conditions included
Any mental health condition	Oppositional defiant disorder, obsessive-compulsive disorder, conduct disorder, substance abuse, anxiety disorder, depression, adjustment disorder, bipolar disorder/mania, learning disability, tics
Neurological disorder	Epilepsy, cerebral degenerations, extrapyramidal disease and abnormal movement disorders, multiple sclerosis, cataplexy and narcolepsy, hypoxic-ischemic encephalopathy

*Anti-psychotics are not considered in this study because they can be used for other mental health conditions other than ADHD.

Appendix 2. Diagram of stratified medicine method

LDX, lisdexamfetamine dimesylate; OROS-MPH, osmotic-release oral system methylphenidate