Abstract
Objective:
To develop and apply a longitudinal model that adjusts for pre-treatment covariates to examine the trajectory of healthcare costs in duloxetine patients with major depressive disorder (MDD).
Methods:
Retrospective healthcare cost data from Thomson Reuters Marketscan® Database included 10,987 patients with MDD, aged 18–64, receiving duloxetine at low (<60 mg/day), standard (60 mg/day), or high (>60 mg/day) initial doses. A linear mixed-effects model for repeated measures used dose, month, and dose*month as fixed effects and patient (dose) as a random effect, and adjusted for demographics, comorbidities, body system disorders, and prior medication history. Model goodness-of-fit was evaluated with R2. Rates of change (slopes) were estimated from the fitted model and differences in the cost trajectory among dosing cohorts were tested using the F-test. Bootstrapping and propensity score (PS) stratification were conducted to provide sensitivity analyses.
Results:
Main effects and covariates were all significant (p < 0.05). Adjustment by pre-treatment covariates greatly improved the model fit (R2 = 0.43). The model revealed a significant increase in healthcare costs in the 6 months preceding and a significant decrease in the 6 months following duloxetine initiation for each initial dose cohort and the overall cohort (p < 0.05). In both the pre- and post-treatment periods, the high initial-dose cohort had higher healthcare costs than standard or low initial-dose cohorts (p < 0.05). Bootstrapping and PS stratification confirmed these test results.
Limitations:
The analyses performed here were based on non-randomized, observational data, and thus subject to potential biases due to unmeasured confounding.
Conclusions:
Longitudinal models, compared with conventional mean-based methods, provide better opportunities to assess changes in cost trajectory patterns around the time of changes in medical treatment. In insured patients with MDD started on duloxetine, healthcare costs increased before duloxetine initiation, perhaps signaling a clinical deterioration that led to a change in treatment strategy. Healthcare costs then decreased following duloxetine initiation.
Introduction
Major depressive disorder (MDD), affecting ∼7% of adults in the US, is among the most common and costly psychiatric disordersCitation1,Citation2. Antidepressants are the mainstay treatment for adults with MDD. Duloxetine is a second-generation antidepressant and has demonstrated efficacy, safety, and tolerability in patients with MDDCitation3–6. Duloxetine is approved for treatment of MDD at starting doses of 40 or 60 mg/day, acute treatment target doses of 40–60 mg/day, a maintenance treatment target dose of 60 mg/day, and a maximum recommended dose of 120 mg/dayCitation7. No significant differences in efficacy at doses of 60 mg/day or >60 mg/day have been observed in clinical trialsCitation6. However, healthcare payers are interested in whether the use of higher duloxetine doses in practice settings is beneficial from an economic perspective.
When assessing the economic value of pharmaceuticals in real-world clinical settings, healthcare payers have typically focused their attention on evaluating all-cause cost changes surrounding a change in treatment strategy, such as initiating, switching, titrating, augmenting, or discontinuing treatment with a pharmaceutical agent. The traditional static approach to such healthcare cost analysis, based on comparisons of average pre- and post-treatment costs across treatment groups, fails to account for the dynamic pattern associated with those comparisons. We hypothesized that costs increase immediately before a medication switch. If costs decrease after the medication switch at a similar rate, then the cost analyses before and after treatment initiation may simply show no changes in average costs. However, longitudinal analyses would demonstrate the rationale and potential need for the medication change.
In our previous workCitation8, we used a longitudinal model to examine healthcare cost patterns in the 6 months prior to and 12 months following duloxetine initiation in patients with MDD. The analyses focused on the rate of healthcare cost change, or cost trajectory, for patients initiated on or titrated to high-dose therapy. We discovered that, while the pre- and post-treatment initiation average costs had no differences, the all-cause healthcare cost increased prior to and decreased following duloxetine initiation. When compared with patients started at low-dose (<60 mg/day) or standard-dose (60 mg/day) duloxetine, patients who started on high-dose (>60 mg/day) duloxetine had higher total healthcare costs both prior to and following duloxetine initiation, with increases in pharmacy costs largely offset by reduced inpatient care costs. However, this model did not adjust for potential patient-level confounding variables related to demographics and medical history that might be expected to have a significant impact on costs, irrespective of medication choice.
The objective of the current study was to build and apply a robust longitudinal model for healthcare cost analysis with adjustments for age, gender, health plan type, medical comorbidities, body system disorders, and prior medication history. The proposed model would examine patient-level all-cause healthcare costs longitudinally, focusing on (1) calculations of the monthly rates of change (slopes) in cost trajectories before and after duloxetine initiation and (2) comparisons of those trajectories among MDD patient sub-groups. Examining the rate of change (slope) in cost-trajectory patterns rather than the mean is not a common approach in healthcare cost research. However, we believe this more dynamic longitudinal modelling approach provides an opportunity to gain greater insight regarding the impact of changes in pharmaceutical treatment directives on total medical costs than do more traditional comparative static analysis methods. In this manuscript we expand on the previous analyses done using the longitudinal model in our 2011 studyCitation8, by adjusting for age, gender, health plan type, medical comorbidities, body system disorders, and prior medication history. Diseases and conditions are organized by body system (). This study only examined selected psychoactive and pain medications () which are common for MDD patients. We also conduct propensity score stratification and bootstrapping as sensitivity analyses to deal with confounding and data distribution issues.
Table 1. Body system diseases and prior psychoactive and pain medications.
Method
Subjects and design
This was a retrospective analysis of medical, pharmacy, and enrollment data from the Thomson Reuters MarketScan® Commercial Claims and Encounters Database. Records included for consideration were from patients who were initiated on duloxetine during the 2007 calendar year. This analysis included patients who did not have a prescription claim for duloxetine in the 3 months prior to duloxetine initiation, or an earlier duloxetine claim for which the expected days supply of medication carried over into the 3-month window prior to initiation (n = 65,199). Patients included in the analyses were aged 18–64 years, had at least one inpatient or outpatient diagnosis of MDD in the previous 12 months, and had continuous insurance enrollment for 6 months preceding and 12 months following the index duloxetine dose (duloxetine initiation) (n = 10,987) (). Of the patients who had no supply of duloxetine in the 3 months prior to study initiation, approximately a third had a diagnosis of MDD between 3–12 months before entering the study. Of those, approximately half had continuous health plan enrollment in the 6 months prior to and 12 months following the index prescription.
Figure 1. Patient disposition.
Statistical analysis
Initial doses of duloxetine for each patient were calculated using the formula, quantity × strength/days’ supply of medication. Low (<60 mg/day), standard (60 mg/day), and high (>60 mg/day) initial-dose cohorts were created on the basis of the initial dose prescribed to each patientCitation8,Citation9. The average initial doses were 29.8 mg/day for the low-dose group and 106.1 mg/day for the high-dose group. Summary statistics were produced for patients’ demographic and clinical characteristics. Mean and standard deviation (SD) were calculated for continuous variables; frequency and percentages were reported for categorical variables. Pairwise comparisons between initial-dose cohorts were performed, with chi-square tests for categorical variables and t-tests for continuous variables.
A linear mixed-effects model for repeated measures was developed to examine the healthcare cost trajectory patterns and compare the monthly costs between initial-dose cohorts. Monthly patient-level all-cause healthcare costs for the 6 months preceding and 6 months following duloxetine initiation were calculated. In our analysis, the daily drug costs were averaged to each day of the supply period, which started the day the prescription was filled. Inpatient costs were averaged to each day of hospital stay period, which lasted from admission to discharge. Outpatient costs were added to the day of outpatient visit. Then, monthly costs were calculated based on the daily costs. An average healthcare cost for months 4–6 was calculated and used in the model for the month 5 time point. Healthcare cost outliers were trimmed at $50,000 to minimize distorting effects of high-cost outliers. Observations with monthly healthcare costs over $50,000 accounted for 0.11% in the population and the values were outside of mean + 6 × SD. Initial-dose cohorts using low (<60 mg/day), standard (60 mg/day), and high (>60 mg/day) initial doses were analyzed using a model that adjusted for baseline demographics, comorbidities, body system disorders, and prior medication uses. Demographics adjustments included age (18–25, 26–35, 36–45, 46–55, and 56–64 years), sex, health insurance plan type (preferred provider organization [PPO], point-of-service [POS] with capitation, POS, health maintenance organization [HMO], exclusive provider organization [EPO], and comprehensive, consumer-driven health plan [CDHP]), and US region of residence (Northeast, North Central, West, South).
Healthcare costs at each time point during the 6 months preceding and the 6 months following the index prescription were compared between initial dose cohorts using a linear mixed-effects model for repeated measures. The repeated measures approach assessed costs over time using the following model:
where Y represents patient-level healthcare cost (US dollars), μ represents overall mean, α represents dose effect, p(α) represents the random patient-within-dose effect, τ represents month effect, (ατ) represents dose × month interaction, C represents a list of baseline covariates, ε represents the random inter-patient variability, and δ represents the random intra-patient variability.
The initial variables chosen to represent patient-level medical and treatment history were individual comorbidities, prior medication history, and body system diseases. The inclusion of such a large number of explanatory variables in the model prevented the algorithm from converging. The Charlson comorbidity index (CCI)Citation10, and body system index (BSI), a new measure developed for this study, were included to represent patient medical history, and the prior medication index (PMI) was developed to represent medication history in the model. The CCI, defined by Charlson et al.Citation10, is a weighted index that takes into account the number and the seriousness of comorbid diseases. Both the BSI and PMI, not seen in the literature, were developed for this study, and take into account the number of body systems with disorders and medications used, respectively. The BSI was defined as the number of 17 body systems defined in the International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) with disorders in the 6 months prior to treatment initiation (). The PMI was defined as the number of 19 common psychoactive and pain medications () used in the 6 months prior to treatment initiation.
Rates of change (slopes) in US dollars/month for the 6 months preceding, 3 months preceding, 6 months following, or 3 months following treatment initiation were estimated separately from the fitted model. The slopes for each initial-dose cohort and overall cohort were tested for H0: Slope = 0. The predicted costs (US dollars) at each month before or month after treatment initiation were estimated, and differences in the cost trajectory among initial-dose cohorts at each time point were tested using F-test. All statistical tests were at p < 0.05 significance level.
The R2 statistic was used to evaluate how well the baseline covariate adjusted longitudinal model fits the healthcare cost data and how much improvement in goodness-of-fit the adjusted model gained as compared with the previous unadjusted model. The R2 statistic for mixed-effects models was calculated with the following formulaCitation11:
where logLM is the log-likelihood of the model of interest (either the adjusted model or the unadjusted model), logL0 is the log-likelihood of the intercept-only model, and n is the number of observations.
To provide a robustness check of the longitudinal regression model, propensity score stratification was conducted. Propensity scores were calculated for the three initial-dose cohorts in a multinomial logistic regression model based on four demographic variables (), 21 comorbidities (data not shown), as well as 17 body system disorders, and 19 prior medications () as explanatory variables. Quintiles of the propensity scores were added as strata into the regression model. Differences among the three initial-dose cohorts were re-examined after propensity score stratification. Given the skewed nature of the cost data, bootstrapping was also used as sensitivity analysis to provide a non-parametric assessment of the variability of the cost trend estimates.
Table 2. Demographic information for all included patients, and for patients in the three initial-dose cohorts.
Results
Demographics by cohort
The demographic information and clinical characteristics for all included patients and for patients in the three initial-dose cohorts are shown in . The majority of patients were female (74.5%), and the mean age was 45.9 years (SD = 11.0 years) (). Most patients were from the South (42.7%). The majority of patients had insurance coverage through a PPO (58.9%). Total healthcare costs were statistically greater (p < 0.001) in the 6 months preceding duloxetine initiation for patients in the high initial-dose cohort ($11,229) compared with patients in the standard-dose cohort ($8330) or low initial-dose cohorts ($7947).
Longitudinal model
A linear mixed-effect model for repeated measures with initial-dose cohort, month, and initial-dose cohort-by-month interaction included as fixed effects, patient within initial-dose cohort as random effect, adjusting for demographics and prior medical and medication history, was developed. The fixed effects from the baseline-adjusted longitudinal model are shown in . All the main effects and covariates were statistically significant in the model (p < 0.05). There was no significant interaction between dose and month (p = 0.2035), indicating similar healthcare cost trajectory patterns among the three initial-dose cohorts. Model-predicted monthly patient-level healthcare costs, overall and by initial-dose cohort () showed an increase prior to treatment and a decrease following treatment initiation. This trend in healthcare costs appeared to be similar among the three initial-dose cohorts.
Figure 2. Least squares mean of patient-level monthly healthcare costs, overall and by initial-dose cohort (US dollars/month). Treatment month is reported as months prior to treatment initiation (negative values) and months following initiation (positive values). Overall, n = 10,987; High Initial-Dose Cohort, n = 1045; Standard Initial-Dose Cohort, n = 6686; Low Initial-Dose Cohort, n = 3256. US, United States.
Table 3. Type 3 tests of fixed effects from the baseline covariate adjusted longitudinal model.
The pairwise comparisons among initial-dose cohorts, adjusted for demographics, comorbidities, and prior medication uses, are shown in . The three initial-dose cohorts were different (overall, p < 0.05). The high initial-dose cohort had higher healthcare costs than the standard or low initial-dose cohort (overall p < 0.001). The bootstrap-derived 95% confidence intervals (US dollars/month) from the sensitivity analysis for pairwise comparisons among initial-dose cohorts supported the above test results (). Sensitivity analysis using propensity score stratification produced similar results for pairwise comparisons among the initial-dose cohorts with slightly elevated p-values. Specifically, p-values ranging from 0.036–0.048 in the original analysis ranged from 0.056–0.076 in the propensity score-adjusted sensitivity analysis. While there was some loss in significance, the overall results were not very different.
Table 4. Pairwise comparisons and sensitivity analysis among initial-dose cohorts.
The rates of change in healthcare costs (slopes) from the baseline-adjusted longitudinal model (US dollars/month) are shown in . In both the 3- and 6-month periods prior to treatment initiation, all three initial-dose cohorts and the overall cohort showed a significant increase in healthcare costs (p < 0.05). On average, overall healthcare costs increased $108 (standard error [SE] $10.9) per month in the 6-month period prior to treatment initiation, and $240 ($23.8) per month in the 3-month period prior to treatment initiation. During the 6-month period prior to treatment initiation, the low initial-dose cohort had numerically greater monthly increases in healthcare costs than the high initial-dose cohort ($121 vs $99). All three initial-dose cohorts and the overall cohort showed a significant decrease in healthcare costs in the 3 months following treatment initiation (p < 0.001), and the 6 months following treatment initiation (p < 0.05). On average, healthcare costs decreased $165 ($21.1) per month in the 3 months following treatment initiation, and $65 ($10.7) per month in the 6 months following treatment initiation. The high initial-dose cohort had a numerically greater decrease in healthcare cost than the low initial-dose cohort in 3 months ($361 vs $132) and 6 months ($137 vs $44) following treatment initiation. The 3-month change in healthcare costs was observed to have a steeper slope than the 6-month change in healthcare costs prior to and following treatment initiation.
Table 5. Rate of cost change (US dollars/month) and sensitivity analysis prior to and following initiation of duloxetine in patients with major depressive disorder.
Results from the sensitivity analysis for rate of changes in healthcare costs (slope, US dollars/month) are shown in . Bootstrap-derived 95% confidence intervals for the rate of cost change did not cover the value 0, which confirmed the results based on linear regression models.
Model goodness-of-fit was measured with the R2 statistic from likelihood ratio statistics. The R2 statistic was calculated from the fitted model with 87,896 observations. Comparing an R2 statistic between the adjusted and unadjusted models showed differences between including and not including demographics, and prior medical and medication history in the specification of fitted models. Notably, R2 for the baseline covariate adjusted model (R2 = 0.43) was higher than for the unadjusted model (R2 = 0.06).
Discussion
The conventional approach to analysis of healthcare costs has been to compare mean costs among initial-dose cohorts over a fixed period (or periods) of time. The objective of conventional analysis has been to understand changes in mean costs from one period to another, not rates of change within these different periods of time. The longitudinal model not only addresses mean differences among initial-dose cohorts, but also explores and focuses on the trends over time. Longitudinal analyses are useful when the objective is to assess the dynamics of healthcare costs over time, for instance, surrounding changes in medical treatment. A change in medical treatment, such as medication initiation, switching, titration, augmentation, or discontinuation, typically results from a patient’s new health condition or a failed (or ineffective) treatment of a patient’s existing health condition and is associated with changes in costs of care. Longitudinal analyses allow the comparison of changes within distinct periods of time, for example before and after initiation of a drug.
In the current study, healthcare costs were increasing before initiation of duloxetine therapy, perhaps signalling a clinical deterioration that led to a change in treatment strategy, and decreasing after initiation. The initiation of treatment with duloxetine was followed by a decrease in healthcare costs during the 6 months post-initiation. The initial-dose cohorts were different, but the overall trend in rate of change in total healthcare costs was similar amongst the initial-dose cohorts.
Compared with patients in the low or standard initial-dose cohorts, patients in the high initial-dose cohort had higher healthcare expenses both prior to and following duloxetine treatment initiation. In addition, the greatest decrease in healthcare costs was seen in the high initial-dose cohort. The patients in the high initial-dose cohort tended to be older, have more comorbid conditions, and demonstrate greater use of concomitant medicationsCitation11. Therefore, the initial healthcare cost would be expected to be higher in this initial-dose cohort, allowing for a greater decrease in healthcare costs after treatment initiation. More studies would need to be done to determine if the use of higher doses of duloxetine is beneficial from an economic perspective. Studies or analyses such as crossover or randomized trials, observations of changing doses, or self-selection models that permit simulations could be applied to determine whether higher doses are appropriate.
As reported earlierCitation8, the longitudinal model without adjustment for baseline covariates did not address the confounding issue. However, confounding is often an issue in real-world healthcare data analysis. This analysis refined our 2011 resultsCitation8 by accounting for confounding in a more sophisticated version of the longitudinal model. The aim of the current study was to build and apply a more robust longitudinal model adjusting for pre-treatment covariates to examine trajectory patterns of healthcare costs in patients with MDD treated with duloxetine. In an initial attempt, all of the available individual comorbidities, body system disorders, and prior medication uses were included in the model, making the model too complex to converge. The current longitudinal model includes three indexes and accounts for the number of comorbidities (CCI), body system disorders (BSI), and prior medication history (PMI). All three indexes were statistically significant (p < 0.0001), indicating that all the covariates were important additions adjusting for confounding in the model. Including demographics, CCI, BSI, and PMI in the adjusted model improved the goodness-of-fit of the model (unadjusted model R2 = 0.06, adjusted model R2 = 0.43). Future studies are needed to validate and improve the BSI and PMI indexes.
In our 2011 studyCitation8, which used an unadjusted longitudinal model, the rates of change in costs (US dollars/month) prior to and following duloxetine initiation in patients with MDD were not significant for the 6 months prior to duloxetine initiation in the high initial-dose cohort and the 6 months following duloxetine initiation in the low initial-dose cohort. However, in the current study which employed the adjusted longitudinal model, all three initial-dose cohorts showed significant rates of change in total healthcare costs prior to and following initiation of duloxetine (p ≤ 0.0140). The significant differences in rates of change in total healthcare costs before and after duloxetine initiation, compared with the lack of significant differences in the unadjusted model, suggest that those differences were masked by the confounding variables introduced into our revised model. The adjusted model provides a better representation of change in healthcare costs over time. The current longitudinal model was efficient in modelling the cost data and provides a more reliable and stable result compared with the unadjusted model.
Limitations
Several limitations should be considered when evaluating these analyses. Data were collected from a health insurance claims database and do not describe the patient’s clinical information. Because patients in this analysis were privately insured, it is unknown if these findings can be extrapolated to the general US patient population. In addition, the inclusion criteria required patients to undergo a 3-month duloxetine-free period immediately prior to index dose, but healthcare costs were analysed for the 6-month period preceding the index dose. Therefore, duloxetine exposure was possible for 3 months (between month 6 and month 3 prior to index dose); and ∼3.5% of the study population was exposed to duloxetine during that time. The full study limitations were previously published as Cui et al.Citation8 in 2011.
The within-group comparisons over time, such as pre- and post-dosing costs, do not completely address broader economic issues related to treatment strategy. The estimation of daily duloxetine doses was based on insurance claim forms, rather than monitored compliance with physician recommendations regarding medication use. Therefore, some distortion in our results is possible. In addition, the reasons for patient discontinuation, the specific patients who discontinued treatment and the effects of discontinuation on healthcare spending through treatment duration were unknown.
The inclusion of an array of specific comorbidities or comorbidity groupings and prior medications and body system diseases produced a model that was too complicated to converge. Therefore, CCI, BSI, and PMI were used in the model instead. While CCI has attained widespread acceptance, BSI and PMI are newer and, as such, are not yet well validated. Further studies are needed to improve the BSI and PMI by not only considering the number of body system disorders or medications used, but also by taking the severity or importance into account.
The purpose of this analysis was interpreting healthcare patterns associated with the initial duloxetine dose. The analyses across initial-dose cohorts performed here were based on non-randomized, observational data and thus subject to potential biases due to measured and unmeasured confounding. Although demographic variables, CCI, BSI, and PMI were used to address the issue of bias and the model goodness of fit (R2 = 0.43) was higher than the unadjusted model (R2 = 0.06), the linear mixed-effect model could not fully address the selection bias issue. The issue was not fully addressed even after we applied propensity score stratification. Propensity score matching would be a solution; however, this manuscript focuses on estimation of the slope for each dose cohort and not on comparison of the three initial dose cohorts. Further research is needed to assess the causal relationship of dosing and costs.
Conclusions
Longitudinal models enable us to engage in a more sophisticated dynamic analysis of costs than traditional static mean comparison. In this analysis, the adjusted longitudinal model revealed a significant increase in healthcare costs in the 6 months prior to and significant decrease in the 6 months following initiation of duloxetine therapy for each initial-dose cohort and the overall cohort (all p < 0.05). The increase in healthcare costs prior to the initiation of duloxetine therapy, perhaps indicated a declining clinical state that led to a change in treatment strategy. Healthcare costs decreased following initiation of duloxetine.
The addition of the three indices (CCI, BSI, and PMI) allowed the longitudinal model to adjust for baseline covariates. The differences in rates of change in total healthcare costs with the adjusted longitudinal model compared with the unadjusted modelCitation8 illustrate the differences that were masked by confounding variables. The current longitudinal model adjusting for demographics, comorbidities, and prior medication use fits the data well (R2 = 0.43) and provides a more robust model of cost analysis than the unadjusted model.
Transparency
Declaration of funding
Financial support was provided by Eli Lilly and Company, Indianapolis, IN. Employees of Lilly were involved in the study design, analysis of data, and in the decision to submit the manuscript for publication.
Declaration of financial/other relationships
At the time this manuscript was written, ZC, DF, WS, SA, and DN were full-time employees of Eli Lilly and Company, and were minor stockholders of Eli Lilly and Company.
Acknowledgements
Appreciation is expressed to Jody Arsenault, PhD, for writing and editorial contributions. Dr Arsenault is a scientific writer employed full-time by PharmaNet/i3. Eli Lilly and Company contracted the technical writing of this manuscript with PharmaNet/i3, an inVentiv Health Company. Also acknowledged are Baojin Zhu and Xiaomei Peng of Eli Lilly and Company for critical review of this manuscript, Yun Fang of PharmaNet/i3 for data analysis, and Angela Lorio of PharmaNet/i3 for editorial review of this manuscript. Trial Registration: http://www.clinicaltrials.gov Identifier: NA.
References
- Kessler RC, Berglund P, Demler O, et al. Lifetime prevalence and age-of-onset distributions of DSM-IV disorders in the National Comorbidity Survey Replication. Arch Gen Psychiatry 2005;62:593-602
- Greenberg PE, Kessler RC, Birnbaum HG, et al. The economic burden of depression in the United States: how did it change between 1990 and 2000? J Clin Psychiatry 2003;64:1465-75
- Detke MJ, Lu Y, Goldstein DJ, et al. Duloxetine 60 mg once daily dosing versus placebo in the acute treatment of major depression. J Psychiatr Res 2002;36:383-90
- Detke MJ, Lu Y, Goldstein DJ, et al. Duloxetine, 60 mg once daily, for major depressive disorder: a randomized double-blind placebo-controlled trial. J Clin Psychiatry 2002;63:308-15
- Detke MJ, Wiltse CG, Mallinckrodt CH, et al. Duloxetine in the acute and long-term treatment of major depressive disorder: a placebo- and paroxetine-controlled trial. Eur Neuropsychopharmacol 2004;14:457-70
- Goldstein DJ, Mallinckrodt C, Lu Y, et al. Duloxetine in the treatment of major depressive disorder: a double-blind clinical trial. J Clin Psychiatry 2002;63:225-31
- Eli Lilly and Company. Cymbalta [package insert]. Indianapolis, IN: Eli Lilly and Company; 2012
- Cui Z, Faries DE, Zhao Y, et al. Trajectory analysis of healthcare costs for patients with major depressive disorder treated with high doses of duloxetine. J Med Econ 2011;14:662-72
- Liu X, Cui Z, Niu L, et al. Dosing patterns for duloxetine and predictors of high-dose prescriptions in patients with major depressive disorder: analysis from a United States third-party payer perspective. Clin Ther 2011;33:1726-38
- Charlson ME, Pompei P, Ales KL, et al. A new method of classifying prognostic comorbidity in longitudinal studies: development and validation. J Chronic Dis 1987;40:373-83
- Magee L. R2 measures based on Wald and likelihood ratio joint significance tests. Am Stat 1990;44:250-3