• adherence;
  • marginal structural models;
  • medication compliance;
  • persistence;
  • type 2 diabetes


  1. Top of page
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. References

Objective:  We applied marginal structural models (MSMs) to estimate the effects of medication adherence with hypoglycemics on reducing the risk of microvascular complications in type 2 diabetic patients.

Methods:  A retrospective longitudinal cohort study for type 2 diabetes patients was conducted using the California Medicaid claims database (1995–2002). Medication adherence and multiple time-varying confounders were measured quarterly over a maximum of 7.5 years follow-up. Cox regression models and MSMs results on the effect of compliance were compared.

Results:  Of 4708 eligible patients, 2644 (56.2%) experienced microvascular complications during the follow-up period. After controlling for baseline covariates, standard Cox models estimated that adherence was associated with increased risk of complication with hazard ratio (HR) of 1.09 (95% confidence interval (CI): 1.00, 1.18). With adjustment of time-varying confounders as exogenous variables, the HR was 0.96 (0.88, 1.04). Using the MSM technique, the HR was 0.76 (95% bootstrap CI: 0.60, 0.92), indicating a significant benefit of medication adherence with hypoglycemics on the reduction of microvascular complications. This result contrasts with the negative results obtained in the hazard model, and is more consistent with prior clinical trial results

Conclusion:  Unlike conventional models, MSMs estimated that higher medication adherence may result in reduced risk of microvascular complications among patients with type 2 diabetes.


  1. Top of page
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. References

The ultimate goal of medication adherence research is to improve health outcomes in target populations. Nevertheless, the effect of medication adherence on health outcomes and the magnitude of such effects have not been rigorously studied. Although the randomized clinical trial is the gold standard for causal inference, medication adherence (synonym of “compliance”[1]) is a patient subjective choice, and cannot be randomly assigned. Therefore, researchers need to use observational studies to understand such relationships.

DiMatteo et al. conducted a meta-analysis on the relationship between adherence and health outcomes, and only identified 44 articles pertaining to medication adherence [2]. They noted that one reason for such underinvestigation of this important relationship in the literature is that the benefit of adherence on outcomes was often taken for granted. In addition, they found that “the studies . . . are all correlational”[2], which reflects the limitations of conventional methodology in addressing the relationship between adherence and outcomes.

In an observational study, the estimation of the relationship between medication compliance and outcomes can be confounded by selection bias, where patients with more severe disease tend to be more adherent to medication, perhaps due to a perception of greater need [3]. In addition, medication treatment for chronic illness is not a static intervention but a prolonged and dynamic process that may vary over time where perceived improvement in treatment outcomes may reduce or increase future adherence. Given the presence of time-varying confounding, reviewers of the compliance literature have cast doubt on whether the effect of compliance could be identified. As Morris and Schulz noted, “researchers have attempted to identify causal relationships between variables, assuming that the variables can be treated as independent. Nevertheless, the phenomenon of medication-taking behavior involves variables that are interrelated with the possibility of feed-back loops”[4]. This time-varying confounding phenomenon is also recognized by DiMatteo et al., as “reverse causality,” which “may play a role . . . , particularly during a long course of treatment”[2].

Time-varying confounding is a ubiquitous phenomenon in health intervention research. Nevertheless, conventional studies often assume adherence as constant, and measure it using a single index over the study period (e.g., Medication Possession Ratio—MPR), whereas the outcomes are also frequently measured during the same study period. The partial correlation between adherence and outcomes is then established by controlling baseline covariates. Because of the ambiguous temporal order and their mutual interaction, the effect of adherence on outcomes and the effect of outcomes on adherence cannot be distinguished. Furthermore, the effects of both time-independent and time-varying confounding could produce biased and puzzling estimates. Therefore, because of the limitations of the traditional study design, lack of methodological rigor [5], and publication bias, the assumption that adherence is beneficial to health outcomes remains unproven [2,6].

Despite lack of convincing evidence, it is commonly taken for granted that higher medication adherence improves outcomes [2], and it is frequently quoted: “drugs don't work in patients who don't take them.” Yet it is important to understand whether higher adherence improves outcomes and the extent of it, because it is the rationale behind adherence-improving interventions and disease management programs, as well as a key element of cost-effectiveness modeling for estimating effectiveness under real-world adherence. This study hypothesizes that higher adherence improves outcomes and an alternative model is proposed to estimate these effects.

In contrast to the conventional model, this study measures both adherence and other outcomes-related covariates (time-varying confounders) repeatedly over time, and accounts for the dynamic interactive effects between adherence and time-varying confounders. This estimation method is executed by applying marginal structural models (MSMs), which is a relatively new model design developed by Robins and colleagues [7,8].

Robins and colleagues developed MSMs for causal inference of time-varying treatment in observational longitudinal studies with the presence of time-varying confounding effect [8–10]. These models operate under the assumption of sequential ignorability, i.e., there is no unmeasured confounder conditional on past treatment and covariate history. MSMs are estimated using inverse probability of treatment weight (IPTW), where the inverse probabilities of receiving the actual treatment given a set of previous covariate history are served as weights (with further stabilization) to estimate the effect of treatment [8,11,12]. As an intuitive interpretation, at any specific time, the purpose of weighting is to create a pseudo-population where time-varying confounders no longer predict adherence level at time; therefore, the confounding effects are removed with weighting and the effect estimate of compliance remains unchanged. Thus, an otherwise complicated model can be converted to a traditional Cox proportional hazards model with the application of patient- and time-specific weights.

The IPTW estimation method used in MSMs is an extension of the propensity score method to a longitudinal model (i.e., a longitudinal version of the propensity score method), where propensity scores were not only estimated and applied at baseline (as in traditional method), but also estimated and applied at each time intervals. In contrast to the conventional approach that assumes that the initial treatment decision is intended over the entire follow-up period, this longitudinal MSM assumes treatment decisions are allowed to change over time conditional on the past treatment response. In this study, the time-varying propensity score (i.e., “probability of treatment”) refers to the probability of choosing higher adherence level conditioning on baseline and time-varying confounders. After applying inverse weighting of the time-varying propensity scores, the treatment of choice would be random, with the assumption of no unmeasured confounders at each point in time (i.e., sequential ignorability).

The principal objective of this study is to estimate the effect of adherence to hypoglycemic medications on preventing microvascular complications (e.g., retinopathy, neuropathy, and nephropathy) among type 2 diabetes patients. A secondary goal of this study is to demonstrate the merit of MSMs when estimating the effect of medication adherence on health outcomes in a longitudinal setting, by comparing it to conventional regression methods.


  1. Top of page
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. References


The data source for this study is a random 20% extract of the fee-for-service portion of the California Medicaid program (Medi-Cal) paid claims and eligibility files from 1995 to 2002. Medi-Cal covers outpatient care, nursing home, inpatient, prescription drugs, as well as other medical services for poor and disabled California residents. Medi-Cal paid claims files include institutional (facility) claims at the claim level, professional services by specific service, and pharmacy claims for each filled prescription. Eligibility files include the monthly-recorded enrollment status and enrollees' demographic information.

Sample Selection

The study cohort includes adult type 2 diabetic patients (identified by International Statistical Classification of Diseases and Related Health Problems (ICD)-9-CM: 250.xx, but with no. 250.x1 or 250.x3 in claims history) who newly started on hypoglycemic medications with a minimum prior period of 6 months with no hypoglycemic prescriptions. The date of the first hypoglycemic prescription filled serves as the index date. In addition, selected patients were required to have at least two prescription fills of hypoglycemics (either oral hypoglycemics or insulin), and continuous enrollment from 6 months before to 12 months after the index date. Figure 1 provides a diagram of the study timeline.


Figure 1. Diagram of study time line. Rx: hypoglycemic prescription.

Download figure to PowerPoint

Patients were excluded if they initiated insulin as the only hypoglycemic during the first quarter after the index date (which is very rare among type 2 diabetes patients). Patients were also excluded if they had a diagnosis of diabetes-related microvascular complications in the preindex period or within 3 months after the index date, because many patients may have experienced complications at the time of starting medication treatment [13]. This exclusion was justified because patients with concurrent complications appear to have a different response to medications [14,15].

The final selected patient cohort was followed until the first occurrence of any of the following events: 1) disenrollment from Medi-Cal; 2) diagnosis of diabetes-related microvascular complications [16]; and 3) the end of data availability (i.e., 12/2002).

Outcomes Measure

Outcome was defined as the time span from the index date to the first diagnosis of diabetic microvascular complications in any medical claim, including neurological disorders, diabetic foot problems, retinopathy, and nephropathy [16].

Adherence and Covariates

Adherence is measured quarterly using the MPR, the ratio of total number of days covered by any prescription of hypoglycemic class within a given quarter over the number of calendar days (90 days). When the days supplied from prescription fills of the same class overlap each other, the overlapped coverage days are not double counted when calculating MPR for this class. MPRs of the different classes of hypoglycemic medications filled within a quarter are averaged to be the MPR for that quarter. For comparison to traditional analyses, two versions of a single-index MPR measure were generated: a fixed-length single index MPR (adherence over 1 year) and variable-length single index MPR (adherence over the entire follow-up period up to the event occurrence or censoring point defined by averaging MPRs of all followed quarters). Adherence level is dichotomized based on a traditional threshold of an 80% MPR value [2,17,18]. Sensitivity analyses of different thresholds for the adherence definition were conducted with 10% point increments from 40% to 90%.

Patient-specific baseline variables included age in years at index date, gender, ethnicity (Asian, Black, Hispanic, Caucasian, and other), index year (calendar year of the index date), length of preindex eligibility in months, Charlson Comorbidity Index measured over the 6-month preindex period, and hypoglycemic regimen during the first quarter after the index date. Regimens with relatively high frequency are reported, and the rest are combined together as “other regimens.”

The following covariates were measured quarterly as time-varying confounders considering their potential interaction effect with adherence: 1) number of office/outpatient visits; 2) any emergency visit (dichotomous); 3) any hospitalization (dichotomous); 4) number of unique drugs measured at generic active ingredient combination level regardless whether the drug is diabetes related or not; 5) any use of self-monitoring blood glucose monitors (SMBG), lancets and test strips use (as identified by Healthcare Common Procedure Coding System codes, Medi-Cal supply codes, or national drug code codes as recorded in Medi-Cal paid claims); 6) any diagnosis of hypoglycemia; 7) any diagnosis of coronary heart disease (CHD), congestive heart failure (CHF), depression, hypertension, dyslipidemia, valvular disease, obesity, or hypothyroidism; 8) any use of ACE inhibitors (ACEI) or angiotensin receptor blockers; 9) any diagnosis of uncontrolled status of diabetes (250.x2); and 10) any diagnosis of ketoacidosis or hyperosmolarity.

Descriptive Analysis

Patient baseline characteristics and their initial hypoglycemic regimens are described with mean and standard deviation for continuous variables and number and percentage for categorical variables. The Wilcoxon test is reported for comparing continuous variables, and the chi square test for categorical variables.

Cox Proportional Hazard Models

We first applied the conventional methodology to estimate effect of adherence. The Cox proportional hazard model is a semiparametric model with time-invariant proportionality assumption. To show the different modeling strategies with alterative specifications of Cox models and MSMs, we constructed five models with increasingly weaker assumptions (see Fig. 2). Model 1 includes a fixed-length single-index adherence variable C without any other covariates, which is the overall adherence measure for the first year. Model 2 includes two models with baseline covariates B and two measures of adherence: fixed-length single-index adherence and variable-length single-index adherence. Model 3 includes baseline covariates B and time-varying adherence C(t) which was measured each quarter until the end of the follow-up period. Model 4 includes baseline covariates B, all time-varying covariates L(t), and time-varying adherence measure C(t). In model 4, time-varying covariates were entered as exogenous variables (not determined by past treatment history) and their dynamic interactions with time-varying adherence were ignored. Lastly, model 5 was an MSM with IPTW estimation which modeled time-varying covariates as confounders independent of time-varying adherence.


Figure 2. Diagrams of study models with (increasingly weaker) assumptions required for unbiased estimates. C, single-index adherence; B, baseline covariates; C(t), time-varying adherence; L(t), other time-varying covariates; U, unobservable factors; Y, the outcome variable. Model 1: random clinical trial; Model 2–4: Cox models with different assumptions; Model 5: MSM (with weakest assumption).

Download figure to PowerPoint


In order to implement the MSMs, we fitted weighted pooled logistic regressions with each person-quarter as an observation for weight estimation [8]. To obtain more robust error estimates, we performed nonparametric bootstraps of random sampling with replacement of 500 times to estimate 95% confidence interval (CI) of adherence effect on outcomes. All analyses were performed by SAS version 9.1 (SAS Institute Inc., Cary, NC, 2003).


  1. Top of page
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. References

Patient Characteristics

A total of 4708 patients with type 2 diabetes were selected as our study cohort, with an average 8.9 and median 6.4 quarters of follow-up. These patients contributed a total of 44,464 patient-quarter observations. The mean age was 61.0 (±14.8) years, and 40.0% were male (Table 1). At baseline, 21.2% of the patients had a record of diagnosis for hypertension, 6.7% with dyslipidemia, 7.1% with CHD, and 3.6% with CHF. During the first quarter, 2544 (54.0%) patients received sulfonylurea monotherapy (SUL) and 1190 (25.3%) received metformin monotherapy (MET). Sulfonylureas plus metformin (SUL + MET) was the predominant combination therapy used as the initial therapy by 520 (11.1%) patients. By the end of the follow-up period, 2644 (56.2%) patients had received at least one diagnosis of microvascular complications, and the median complication-free survival time was seven quarters.

Table 1.  Baseline characteristics of the study population (n = 4708)
VariableMean ± SD n (%)
  1. SD, standard deviation; CHD, coronary heart disease; CHF, congestive heart failure; SUL, sulfonylureas; MET, metformin; ACEI, ACE inhibitors; ARB, angiotensin receptor blockers.

Age61.0 ± 14.8
Male1882 (40.0)
 Asian1261 (26.8)
 Black434 (9.2)
 Hispanic637 (13.5)
 Other881 (18.7)
 Caucasian1495 (31.8)
Index year 
 1995571 (12.1)
 19961009 (21.4)
 1997616 (13.1)
 1998530 (11.3)
 1999525 (11.2)
 2000673 (14.3)
 2001784 (16.7)
Prior eligibility length in quarter28.3 ± 20.9
Charlson comorbidity index at baseline1.1 ± 1.1
Number of different drugs4.2 ± 3.8
Number of office/outpatient visits2.2 ± 3.4
Comorbidities at baseline 
 CHD336 (7.1)
 CHF169 (3.6)
 Depression88 (1.9)
 Hypertension997 (21.2)
 Dyslipidemia314 (6.7)
 Valvular Disease62 (1.3)
 Obesity85 (1.8)
 Hypothyroidism66 (1.4)
Initial hypoglycemics regimen 
 SUL2544 (54.0)
 MET1190 (25.3)
 SUL + MET520 (11.1)
 TZD97 (2.1)
 INS + SUL48 (1.0)
 INS + MET43 (0.9)
 SUL + TZD45 (1.0)
 MET + TZD40 (0.9)
 NAT32 (0.7)
 All Other149 (3.2)
Other time-varying covariates at baseline 
 Any ER visit184 (3.9)
 Any hospitalization501 (10.6)
 Use of ACEI or ARB869 (18.5)
 Uncontrolled diabetic condition269 (5.7)
 Hypoglycemia57 (1.2)
 Ketoacidosis or hyperosmolarity43 (0.9)

Adherence Changes over Time

For the entire 44,464 included patient-quarters, the mean of MPR of all quarters was 48.4% (±39.1%) and the median was 54.4%. The distribution was not normal, with 13,500 (30.4%) of the quarters having zero MPR, corresponding to no drug-covered days during the quarter, and 4028 (9.5%) quarters having 100% MPR. To examine the longitudinal adherence pattern, we first plotted the population adherence over time using a box plot to summarize mean, median, and interquartile range for each quarter (available upon request). The adherence patterns appeared to be smooth and stable at the population level over time despite of a sharp drop during the first year. Nevertheless, the population mean did not provide insights regarding individual changes over time in adherence. Figure 3 depicts the plotted quarterly measured MPRs of 10 randomly selected patients revealing an erratic pattern of patients' longitudinal adherence.


Figure 3. Quarterly adherence of 10 randomly selected patients. MPR, medicine possession ratio.

Download figure to PowerPoint

Survival Analysis and Cox Regression Models

Without adjustment of any covariate in model 1, the 1-year single-index adherence was associated with higher risk (hazard ratio, HR = 1.08) of developing at least one microvascular complication, with statistical significance of P = 0.092 (Table 2). When adjusting for baseline covariates with Cox regression model in model 2A, the 1-year single index adherence was still associated with higher risk (HR = 1.06), but as with model 1, the adherence variable was not statistically significant. model 2B is the same as model 2A, except that a variable-length single-index measure of adherence over the entire follow-up period is used. The estimated HR of this adherence measure was 1.53 (P < 0.0001), presenting a very strong detrimental effect of higher adherence. Model 3 included time-varying adherence measured quarterly up to the end of the follow-up period with baseline covariates, and estimated the HR of adherence as 1.09 with statistical significance (P = 0.0449). Including all time-varying covariates in model 4, the selection bias appeared to be alleviated, with estimates indicating a trivial effect of adherence (HR = 0.96, P = 0.3089).

Table 2.  Comparison of adherence effect estimates of model 1–5
Model no.ModelEstimateHazard ratio95% CIP-value
  1. Notes: No. 1—a Cox model only including fixed length single index adherence C1, No. 2A—a Cox model including fixed length single index adherence C1 and baseline covariate(s) B, No. 2B—a Cox model including variable length single index adherence C1 and baseline covariate(s) B, No. 3—a Cox model including time-varying adherence and baseline covariate(s) B, No. 4—a Cox model including time-varying adherence and time-varying covariate(s) L(t), No. 5—a Marginal structural model with IPTW estimation and Bootstrap estimation for confidence interval.

  2. MSM, marginal structural model; IPTW, inverse probability of treatment weight.

1C10.07371.08(0.99, 1.17)0.0923
2AC1 + B0.05321.06(0.97, 1.15)0.235
2BC2 + B0.42661.53(1.40, 1.67)<0.0001
3C(t) + B0.08511.09(1.00, 1.18)0.0449
4C(t) + B + L(t)−0.04470.96(0.88, 1.04)0.3089
5MSM−0.31510.73(0.54, 0.99)0.0441
(Bootstrap)  (0.60, 0.92) 

Adherence Effects Estimated by MSM with IPTW

Combining all patient quarters, the mean of the stabilized weights was 1.37 (SD = 19.57) and median 0.812 (interquartile range = 0.592−1.070). Summarizing the weights by quarter, the box plot shows the temporal distribution of the log of stabilized weights (Fig. 4). This plot shows that the mean stabilized weights are decreasing over time, yet being “stable” and not too far from the zero line, and the variation of stabilized weights increases over time.


Figure 4. Log of stabilized weights distribution by quarter. The width of the box indicates the number of patients remaining in each quarter, and the upper and lower end of the box represents interquartile range. Whiskers correspond to 1.5 times of interquartile range above and below the 75th and 25th percentile. The squares denote outliers.

Download figure to PowerPoint

Applying the estimated stabilized weights, the effect of adherence on the risk of microvascular onset was estimated by generalized estimating equations (GEE) with binomial distribution and logit link function. The estimated hazard risk was 0.73 (Table 2), with a bootstrap 95% CI (0.60, 0.92), showing a statically significant benefit of adherence. The CI estimated by the robust estimator of standard error of GEE as (0.54, 0.99) and P-value of 0.0441, which is more conservative. This estimate illustrates a reasonable effect estimate by addressing time-varying confounding. Comparing all estimates using different methods, only the adherence effect estimated by MSMs indicates a beneficial effect of adherence on outcomes, whereas all other estimates indicate that improved adherence is associated with increased risk of developing microvascular complications.

Sensitivity Analysis

We repeated the analysis for all specified modeling strategies (model 1–5) by various MPR thresholds using 10% point increments from 40% to 90%. The results of the estimated HRs of adherence (not reported here) reveal that the benefit of adherence estimated by MSMs is stable across different adherence cutoff points, with the effects more pronounced at smaller threshold values (e.g., 40% and 50%).


  1. Top of page
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. References

Past efforts of investigating the relationship between adherence and outcomes often establish an associational relationship using conventional models. Because patients who experience worse intermediate outcomes over time tend to increase the subsequent drug adherence, relative sickness and adherence tend to be positively correlated and vary over time. Therefore, conventional regression models without appropriate accounting for dynamic interaction between time-varying adherence and intermediate variables are prone to negative effects of higher medication adherence, as evidenced in models 1 to 4 in this study. This could partially explain the relative scarcity of publication in estimating the effect of adherence level on outcomes, despite the voluminous literature in other aspects of medication adherence (e.g., measuring, predicting, or developing methods to improve adherence). Since 1975, over 32,000 articles were classified under the medical subject heading “patient compliance,” yet the benefit of higher medication adherence on health outcomes remains an assumption.

In contrast with past research, this study applied a longitudinal MSM that accounted for time-varying confounding, and demonstrated that hypoglycemic adherence significantly reduced the risk of microvascular complications (HR = 0.73, 95% bootstrap CI: 0.60, 0.92). It should be noted that this result is consistent with the United Kingdom Prospective Diabetes Study (UKPDS) clinical trial findings that glycemic control by SULs or insulin compared to a conventional diet reduced microvascular complications [19].

There are several elements of this study that require further exploration in subsequent studies. We dichotomized adherence partially for the analytical and interpretational convenience when MSMs were applied, although MSMs can also be used to estimate the effect of multilevel categorical or continuous exposure variables. Although we applied sensitivity analysis to test the impact of different thresholds on dichotomizing adherence, future studies with finer levels of adherence may be needed to explore the nonlinear effect of adherence on outcomes. In addition, the MSM estimate is based on the weaker assumption that the treatment decision is independent of unobserved confounders at each time conditional on the past observed covariates and intermediate outcomes. Despite controlling multiple observed time-varying confounders, many parameters are unobservable or immeasurable because of the limitations of administrative claims databases. For example, it would be preferable to have specific clinical and behavior factors such as fasting blood glucose, hemoglobin A1c (HbA1c), blood pressure, lipid level, body weight, diet, and exercise over time. These parameters also change and may interact with adherence over time, and thus may be considered as time-varying confounders. Future studies of this nature should incorporate such measures. Finally, it should be noted that the study outcomes are based on presence of an ICD-9 diagnosis code of microvascular complications. The timing of such microvascular complications is unknown, and patients with microvascular complications may be undiagnosed until long after their onset.

Additionally, this study evidenced a strong beneficial effect of adherence to medication, but it provides no insight regarding the mechanisms by which medication adherence produces this effect. The effect could be partially owing to the therapeutic effect of the drug therapy, and can also contain the psychological effect of belief about the drug (e.g., the placebo effect, or the effect of positive thinking) that is independent of the therapeutic effect. In addition, the estimated effect could include the benefit of adherence to other behavioral modifications, such as diet, exercise, self-care of eye/foot, adherence to SMBG, and even better compliance to office appointments, which collectively may account for a major portion of the overall estimated benefit. These effects could be largely immeasurable and inseparable to that of medication adherence, even in a well-controlled setting. Nevertheless, this composite effect of multiple inseparable components, albeit labeled as “adherence effect,” may be more meaningful than a therapeutic effect alone in real practice. For example, patient education programs aimed at improving overall patient adherence to healthy behaviors could simultaneously improve medication adherence and positive beliefs about drug efficacy.

Longitudinal data, where treatment and other factors influencing treatment decision change over time, provide a rich source of information. This study addressed an important question in medication adherence research that is complicated by the dynamic interactions between time-varying adherence and confounders. Unlike conventional research methods correlating adherence level and outcomes, the MSM with time-varying confounding demonstrated a significantly beneficial effect of higher medication adherence on the risk of developing microvascular complication in type 2 diabetes.

Medication adherence is predicated on the expectation that patients make individual choices in treatment decision-making and self-select the adherence level after trading off multiple factors such as perceived clinical benefit, side effects, and cost [20,21]. Strategies intended to enhance adherence have been found to be surprisingly weak in their effects, as even the most effective interventions had only modest effects on adherence and outcomes [22,23]. Understanding the relationship between adherence and clinically important outcomes, and the factors that moderate the relationship between adherence and these outcomes, may guide future evaluations of interventions designed to improve adherence [2].

Source of financial support: No funding was received for this study and the authors have no conflicts of interest which are relevant to the contents of this manuscript. The manuscript was prepared without a contract or funding from a sponsor. The publication of study results is not contingent on the sponsor's approval.


  1. Top of page
  3. Introduction
  4. Methods
  5. Results
  6. Discussion
  7. References