Information about the trial and a questionnaire were mailed in 1998 to a sample of 39 713 women aged 45-64 taken from the Estonian Population Registry and living in two Estonian counties. The recruitment questionnaire included questions about willingness to join a randomised trial and questions about health and social status. Of the 14 743 women who returned the questionnaire, 6606 respondents were interested in participating, of which 4295 women were found to be eligible according to the preliminary assessment, based on time since menopause and other health data from the questionnaires.
Randomisation was carried out before recruitment in permuted blocks, each of a size 16 and each block of the three trial clinics separately, at the National Research and Development Centre for Welfare and Health in Finland. The 125 women participating in the pilot study were not included in the present analysis. The 4170 eligible women who were willing to join the main trial were randomised into blind and non-blind arms and, at the same time, into treatment and control arms (two-by-two design). Hence, there were four study arms: 1) blind HT arm; 2) non-blind HT arm; 3) placebo arm and 4) non-treatment arm. Hereafter, blind HT and placebo arm combined together will be named blind trial, and non-blind HT and non-treatment arms combined together non-blind trial.
Detailed descriptions of recruitment, inclusion and exclusion criteria, trial treatment, adherence, follow-up and trial outcomes as well as the content of information leaflets and trial questionnaires have been published elsewhere [11
]. All participants gave written informed consent. The study protocol was approved by the Committee of Medical Ethics in Tallinn, Estonia and by the Ethics Committee of the University Clinic of Tampere, Finland. Data about randomised women who were not recruited to the trial were acquired with a special permission from the Estonian Data Protection Agency.
Recruitment, trial treatment and adherence
The 4170 randomly assigned women were mailed an invitation letter revealing whether they had been assigned to the blind or the non-blind trial. The treatment allocation was enclosed in a non-transparent sealed envelope with a woman's study number and name on it, and sent to the trial clinic. In the blind trial, the women were told that they would be using either hormone therapy or a placebo; in the non-blind trial, they were told that they would be receiving hormone therapy or non-treatment.
Final recruitment took place between January 1999 and December 2001. A total of 2323 women responded to the mailed invitation to visit the trial doctor. After this secondary assessment of eligibility, 1778 women proved to be eligible and were willing to join the trial, and their randomisation envelope was opened. The reasons for ineligibility after secondary assessment have been reported in detail elsewhere [12
As a result, 404 women were recruited into the blind HT arm, 373 into the placebo arm, 494 into the non-blind HT arm and 507 into the non-treatment arm (Figure ). None of the trial participants switched the trial arm after randomisation. The women in the non-blind HT arm were allocated to open-label HT, the women in the non-treatment arm did not receive any drugs. Adherence was assessed by the number of collected and returned drugs and by the information from annual questionnaires and weekly reports from the clinics. Women taking more than 80% of the allocated drugs were considered to be adherent, whilst women in the non-treatment arm were considered to be adherent if they were not taking hormone therapy for 80% of the time. Data about prescribed HT use in the non-treatment and placebo arms was obtained from the Estonian Health Insurance Fund.
During the treatment period, participating women and health care providers were blinded to the treatment assignment in the blind trial, but not in the non-blind trial. The persons doing the linkages in registries, data collectors and data analysts at research centres were unaware of the treatment allocation, and remained blinded also after the end of the trial. After publication of the results from the Women's Health Initiative (WHI) trial, trial treatment was stopped gradually between January and May 2004. Participants in the blind trial received a letter containing information on their treatment allocation within one month of their final visit.
Outcome measures, data acquisition and analyses performed
The randomised women were followed by annual linkages to the Estonian Health Insurance Fund database, the Estonian Cancer Registry database, and the Estonian Mortality database using a person's identity code. The outcomes studied were coronary heart disease (I20-I25 according to the 10th
revision of the International Classification of Diseases), [18
] cancer (C00-C97), cerebrovascular disease (I60-I69), bone fractures (S12, S22, S32, S42, S52, S62, S72, S82, S92), death from all causes and all these combined. As the database of the Estonian Health Insurance Fund was considered not fully complete for the years 1999 and 2000, the first follow-up date for all outcomes in the present analysis was January 1, 2001. The results of the analysis are reported here separately for December 31, 2004 and December 31, 2007, as the last follow-up dates.
The Estonian Health Insurance Fund is the only organization in Estonia dealing with compulsory health insurance [19
]. It stores information about all health care contacts, using an individual's personal identification code recorded in a centralised, computerised database. The Health Insurance Fund pays for all health care visits, diagnostic examinations, preventive and treatment procedures, hospital stays, surgeries, technical aids during or after surgery, and compensations for medicinal products. All participants in the trial were insured. As compensation to clinics depends on the transmission of data to the central electronic database and not on the diagnosis, we assume the probability of missing data in the database of the Estonian Health Insurance Registry to be minimal. The Estonian Cancer Registry database has been validated for completeness of registration, with the overall completeness of registration being 90.8% in 1998 [20
The linkage with the Population Registry showed that during the follow-up period nine women left from Estonia (all among recruited participants: four in the blind HT arm, one in the non-blind HT arm, one in the placebo arm, three in the non-treatment arm). Data about these women in the registries might have been incomplete (Figure ).
For analysing the effect of blinding on the selection bias (i.e. external validity), the numbers of women recruited to the blind and non-blind trial as well as the proportion of women with different background characteristics recruited to blind and non-blind trial were compared. In addition, hazard ratios for different outcome events among recruited versus non-recruited women were calculated in the placebo arm and in the non-treatment arm. The comparison was restricted to the non-HT arms and the hazard ratios between recruited and non-recruited women in the blind and non-blind HT arms were not calculated in order to exclude the effect of HT from the analysis of blinding.
For analysing the effect of blinding on reporting bias, observer bias, and placebo effect (i.e. internal validity), the hazard ratios of outcome events among recruited women in the blind and non-blind trial arms were compared, separately until the year 2004 and 2007. During the first period, participants in the HT arms were allocated to trial treatment, and the participants in trial arms as well as their physicians were blinded about their treatment allocation. Blinding and trial treatment was stopped in 2004. The analysis with the longer follow-up period until 2007 was done to check if the differences between the outcomes hold also after stopping blinding and trial treatment.
The effect of blinding on the recruitment process was first assessed by comparing the proportions of women recruited to the blind and non-blind trial. Next, the proportion of women with different background characteristics was calculated in blind and non-blind trial arms both among recruited and non-recruited women. The baseline covariates used in the analysis (also in the adjusted analysis for the outcomes) were age, education, living place (corresponding to the clinic of recruitment), smoking status, and time since menopause.
The outcome data were analysed by time-to-event methods. For each of the target disease groups, the number of days from recruitment to the first diagnosis in this group, as registered in the Health Insurance Fund database, the Estonian Cancer Registry database, or in the Estonian Mortality database was used as the outcome variable. Time to diagnosis was censored for women who did not have the corresponding diagnosis registered during follow-up. Person-years at risk were calculated from January 1, 2001 to December 31, 2007 (and from January 1, 2001 to December 31, 2004) or to the outcome event studied or to death whichever came first.
Cumulative hazard plots based on the Kaplan-Meier method were obtained for descriptive comparison of outcome distributions in blind and non-blind trial, separately for recruited and non recruited women.
Comparison of clinical outcomes between blind and non-blind trial for all randomised women was performed (the intention-to-treat analysis for the effect of blinding), using the Cox proportional hazards modelling. Clinical outcomes were then compared between blind and non-blind trial separately for recruited and non-recruited women. For recruited women, separate comparisons were done for HT and non-HT arms, as well as the combined analysis, stratified by HT assignment. Interaction for blinding and treatment among recruited women was tested for, but not included in the model. In addition, comparison of outcomes between recruited and non-recruited women in the placebo and non-treatment arms was carried out.
For each of the comparison, both crude and adjusted hazard ratios with 95% confidence intervals were obtained, using Cox proportional hazards modelling.
The software used for analyses was R for Windows, version 2.8.1 [21