The data for this retrospective analysis were collected from electronic medical records (EMRs) stored in the Nihon University School of Medicine (NUSM) Clinical Data Warehouse (CDW), which integrates clinical data from hospital information systems (HIS) at three hospitals affiliated to NUSM [²]. NUSM's CDW is a comprehensive data warehousing facility that provides data services to users across the clinical and research sectors of NUSM. The experimental protocol was approved by the Ethical Committee of Nihon University School of Medicine. The study subjects consisted of 483 Japanese patients with mild to moderate hypertension, aged 20 years or older who had been treated initially with candesartan cilexetil monotherapy (range: 1~12 mg/day, 93.5% of administration was in the range of 2~8 mg/day) for at least 4 weeks during the period from November 2004 to October 2009, as shown in Figure 1. Patients who had received antihyperlipidaemic agents were excluded from the study. Patients who had received other antihypertensive agents, such as an ARB other than candesartan, angiotensin-converting enzyme inhibitor (ACEI), calcium channel blocker, alpha-blocker, beta-blocker, alpha+beta-blocker, alpha-agonist or thiazide, during the 3 months before candesartan cilexetil administration were excluded from the study. In addition, patients with haemoglobin A1c (HbA1c) of 8.0% or higher were eliminated to exclude patients with very poor glycaemic control. Clinical data from the study subjects included sex, age at the start of treatment, diagnosis of diabetes mellitus (DM) according to the Committee for the Classification and Diagnosis of Diabetes Mellitus of the Japan Diabetes Society (defined as fasting plasma glucose level ≥126 mg/dl, casual plasma glucose level ≥200 mg/dl, plasma glucose 2 h after 75 g glucose load ≥200 mg/dl, or HbA1c level ≥6.5% [⁸]), duration of treatment, results of blood examinations including triglyceride (TG), TC, high density lipoprotein cholesterol (HDL-C) and LDL-C, which were determined at routine clinical visits, and date of examination. A total of 405 patients from this population were eligible for the study of TG, 440 for TC, 313 for HDL-C and 304 for LDL-C.

Our main explanatory variables included sex, age at start of treatment, diagnosis of DM, and "duration" defined as the timing of measurement in days since the start of treatment as follows; baseline (within 3 months before start of treatment), 0~3 M (>0, ≤3 months), 3~6 (>3, ≤6 months), 6~9M (>6, ≤9 months) and 9~12 M (>9, ≤12 months). Our main response variables were repeated measurements of TG, TC, HDL-C and LDL-C levels in blood before and after candesartan cilexetil monotherapy. These data were not randomized, and were inherently unbalanced because the number and timing of the repeated measurements were different among individuals. The composition of the patients and the number of examinations were not equal for each of the treatment durations, and may have been time-dependently affected by candesartan monotherapy itself. Marginal structural models using inverse probability of treatment weighting (IPTW) have been recently developed to solve this problem that the treatment effects are affected by time-dependent confounders that are themselves affected by the treatment [⁹]. Therefore, we used IPTW to balance the treatment durations so as to reduce bias in the patient background and obtain a better idea of the effect of treatment on the outcome of compliance. IPTW is calculated as the inverse of the propensity score. The propensity score method introduced by Rosenbaum and Rubin is an effective tool to reduce bias in nonrandomized studies including unbalanced data [¹⁰]. The traditional propensity score method such as matching, stratification and covariance adjustment is mostly used in binary-value treatment [^11-13]. Recently, there have been many reports trying to expand its application to more than two treatments [^14-16]. We referred to the method introduced by Leslie et al., and used an IPTW-linear mixed effect model to reduce bias in the duration of treatment [^15-18]. This method consisted of three steps as follows. As the first step, we used propensity score adjustment to account for potential selection bias in each treatment duration. We used a logistic regression model to calculate the propensity score as the probability of examinations with each treatment duration. The variables in this step included sex, age and diagnosis of DM. As the second step, IPTW was calculated as the inverse of the propensity score. By using this IPTW in the next step, bias in each treatment duration could be minimized. To give more weight to smaller treatment groups, a weight was created that reflects the sample size for each blood examination. As the third step, an IPTW-linear mixed effect model was fitted to analyse the relationship between these longitudinal data of blood examinations and all other covariates. We fitted repeated measurement analysis (covariance structure: Compound Symmetry) to the data, including sex, age, diagnosis of DM and a confounding factor of interaction of sex and duration of treatment as fixed effect covariates, and duration of treatment as a repeated effect covariate. We selected the Kenward-Roger method to compute the denominator degrees of freedom for the tests of fixed effects. A multiple-comparison test (Dunnett-Hsu post-hoc analysis) was used to compare the differences in means between "baseline" as a reference and other treatment duration periods. A result was considered statistically significant if the p value was less than 0.05. All statistical analysis was performed with SAS 9.1.3 (SAS Institute Inc., Cary, NC) statistical software.

Of the study population of 483 patients with hypertension, those who had been treated with candesartan cilexetil monotherapy but no antihyperlipidaemic agent were subjected to statistical analysis (Figure 1). The numbers of blood examinations were as follows; TG: total 1114 examinations (405 patients with 2.75 measurements per patient), TC: total 1204 examinations (440 patients with 2.74 measurements per patient), HDL-C: total 708 examinations (313 patients with 2.26 measurements per patient), LDL-C: total 781 examinations (304 patients with 2.57 measurements per patient) (Figure 1). Table 1 shows the details of patient information and data. Approximately 36% of patients who underwent blood examination were female and 64% were male. Approximately 47% of patients had DM. Mean age was approximately 61 years, and the age range was 20 to 91 years.

Table 2 shows the results of Type III test of propensity score-weighted linear mixed effect models fitted to plasma TG, TC, HDL-C and LDL-C data. There was a significant association between HDL-C level and sex (p = 0.0017), treatment duration (p = 0.0427), and interaction of sex and treatment duration (p = 0.0349), but no association between HDL-C and age or diagnosis of DM. There was a significant association between TG and sex (p = 0.0041), but no association between TG and age, diagnosis of DM, treatment duration or interaction of sex and treatment duration. There was a significant association between TC and sex (p = 0.0032), but no association between TC and age, diagnosis of DM, treatment duration or interaction of sex and treatment duration. There was no association between LDL-C and sex, age, diagnosis of DM, treatment duration or interaction of sex and treatment duration. Candesartan monotherapy was associated with a decrease in HDL-C level, but had little effect on other lipid parameters.

Table 3 shows the results of Dunnett's multiple-comparison test of propensity score-weighted linear mixed effect models fitted to plasma TG, TC, HDL-C and LDL-C data. HDL-C was slightly but significantly decreased in the '6~9M' period compared with baseline (1.40 vs. 1.47 nmol/L, p = 0.0218), but was not significantly changed in the other periods, '0~3M', '3~6M' and '9~12M', compared with baseline. There was no significant change in TG, TC or LDL-C in any treatment duration period.

We further analysed the data divided by sex, because HDL-C was also significantly associated with sex and interaction of sex and treatment duration. Figure 2 shows the change in HDL-C level in each treatment duration period, compared with baseline. HDL-C was significantly decreased in the '6~9M' period compared with baseline in female data (1.45 vs. 1.57 nmol/L, p = 0.0054), but was not significantly changed in male data (Figure 2 and Table 4). Candesartan monotherapy had an unfavorable effect on HDL-C in female subjects, but HDL-C level in female subjects was within the normal range during the study period.

TG: triglyceride, TC: total cholesterol, HDL-C: high density lipoprotein cholesterol, LDL-C: low density lipoprotein cholesterol, DM: diagnosis of diabetes mellitus.

DM: diagnosis of diabetes mellitus, Sex*Treatment duration: interaction of sex and duration of treatment, TG: triglyceride, TC: total cholesterol, HDL-C: high density lipoprotein cholesterol, LDL-C: low density lipoprotein cholesterol, DF: degrees of freedom, p value: p value of covariate, *: p < 0.05.

TG: triglyceride, TC: total cholesterol, HDL-C: high density lipoprotein cholesterol, LDL-C: low density lipoprotein cholesterol, LS mean: least squares mean, SE: standard error, CI: confidence interval, p value: p value of treatment duration (compared with baseline, multiple-comparison test: Dunnett-Hsu post-hoc analysis), *: p < 0.05.

HDL-C: high density lipoprotein cholesterol, Exam N: number of examinations, LS mean: least squares mean, SE: standard error, CI: confidence interval, p value: p value of treatment duration period (compared with baseline, multiple-comparison test: Dunnett-Hsu post-hoc analysis), *: p < 0.05.