Data processing
Statistical analysis
The main statistical analyses were based on the 'intention-to-treat' sample. The data contained repeated measurements of the response variables. In the case of informative dropouts the values of the outcome variables were completed using multiple imputation (Rubin 1987). Information on atypical therapy, resulting from incomplete therapy or use of other therapies during follow-up, was included as covariates in the model.
In the case of continuous response variables, the statistical analyses were based on linear mixed models (Verbeke and Molenberghs 1997) and in the case of binary responses on logistic regression models (Ling and Zeger 1986).
Model-adjusted statistics were calculated for different design points (Lee 1981). For continuous responses, absolute means and differences and for binary responses prevalences and relative risks were estimated. The delta method was used for calculation of confidence intervals (Migon and Gamerman 1999 p. 138). Statistical significance was tested with the Wald test.
Since compliance as well as changes in patient health may be dependent on various individual, possibly latent background variables, the statistical analyses were also carried out using Bayesian analysis (Kenward and Molenberghs 1999, Chickering and Pearl 1997).
In the analysis of quality control data analysis, the agreement between and repeatability of measurements were estimated as intraclass correlation coefficients (Winer 1971, Fleiss 1981).
Other statistical approaches included cost-effectiveness analysis (Drummond et al. 1997), which was performed using incremental cost-effectiveness ratios. The confidence intervals for these ratios were estimated using bootstrap methods. Multidimensional sensitivity analyses were also performed.
The main statistical analyses were carried out using the SAS software SAS/STAT (procedures MIXED, GENMOD, and PHREG) and SAS/IML (procedure IML) (SAS Institute Inc. 1999).
Quality control
Repeatability of results during the data collection period
The research data was collected over several years by several raters. This would potentially cause variation in the quality of the data. Therefore the following quality control actions were taken and the data collected was controlled using several separate designs:
- The consistency of the interviewers' ratings was evaluated by repeated control ratings of 28 selected interviews recorded on videotape over a 3-year period. Based on these ratings, both the agreement between interviewers and the long-term repeatability of the ratings were evaluated.
- The reliability of the Rorschach Inkblot Method was estimated based on 20 protocols using the Comprehensive System guidelines (Exner 1993).
- All laboratory determinations were made under standard quality control.
Validity of results
Some of the methodological choices required that the agreement with other measurement methods were evaluated:
- The validity of the diagnosis based on the semi-structured interview method, was evaluated by comparison with diagnoses based on the Structured Clinical Interview for DSM-IV axis I and axis II disorders (SCID) in a sample of 27 selected patients.
- The agreement between the symptom questionnaires and the symptom interviews was evaluated.
- The agreement between the information on the use of psychiatric medication given by the patient and from the public health register was evaluated.
- Patients had to wait for their treatment to start after they had been admitted to the study. Possible changes in their symptoms during that period were evaluated.