Statistical assessment of the learning curves of health technologies

Abstract

OBJECTIVES: (1) To describe systematically studies that directly assessed the learning curve effect of health technologies. (2) Systematically to identify 'novel' statistical techniques applied to learning curve data in other fields, such as psychology and manufacturing. (3) To test these statistical techniques in data sets from studies of varying designs to assess health technologies in which learning curve effects are known to exist. METHODS - STUDY SELECTION (HEALTH TECHNOLOGY ASSESSMENT LITERATURE REVIEW): For a study to be included, it had to include a formal analysis of the learning curve of a health technology using a graphical, tabular or statistical technique. METHODS - STUDY SELECTION (NON-HEALTH TECHNOLOGY ASSESSMENT LITERATURE SEARCH): For a study to be included, it had to include a formal assessment of a learning curve using a statistical technique that had not been identified in the previous search. METHODS - DATA SOURCES: Six clinical and 16 non-clinical biomedical databases were searched. A limited amount of handsearching and scanning of reference lists was also undertaken. METHODS - DATA EXTRACTION (HEALTH TECHNOLOGY ASSESSMENT LITERATURE REVIEW): A number of study characteristics were abstracted from the papers such as study design, study size, number of operators and the statistical method used. METHODS - DATA EXTRACTION (NON-HEALTH TECHNOLOGY ASSESSMENT LITERATURE SEARCH): The new statistical techniques identified were categorised into four subgroups of increasing complexity: exploratory data analysis; simple series data analysis; complex data structure analysis, generic techniques. METHODS - TESTING OF STATISTICAL METHODS: Some of the statistical methods identified in the systematic searches for single (simple) operator series data and for multiple (complex) operator series data were illustrated and explored using three data sets. The first was a case series of 190 consecutive laparoscopic fundoplication procedures performed by a single surgeon; the second was a case series of consecutive laparoscopic cholecystectomy procedures performed by ten surgeons; the third was randomised trial data derived from the laparoscopic procedure arm of a multicentre trial of groin hernia repair, supplemented by data from non-randomised operations performed during the trial. RESULTS - HEALTH TECHNOLOGY ASSESSMENT LITERATURE REVIEW: Of 4571 abstracts identified, 272 (6%) were later included in the study after review of the full paper. Some 51% of studies assessed a surgical minimal access technique and 95% were case series. The statistical method used most often (60%) was splitting the data into consecutive parts (such as halves or thirds), with only 14% attempting a more formal statistical analysis. The reporting of the studies was poor, with 31% giving no details of data collection methods. RESULTS - NON-HEALTH TECHNOLOGY ASSESSMENT LITERATURE SEARCH: Of 9431 abstracts assessed, 115 (1%) were deemed appropriate for further investigation and, of these, 18 were included in the study. All of the methods for complex data sets were identified in the non-clinical literature. These were discriminant analysis, two-stage estimation of learning rates, generalised estimating equations, multilevel models, latent curve models, time series models and stochastic parameter models. In addition, eight new shapes of learning curves were identified. RESULTS - TESTING OF STATISTICAL METHODS: No one particular shape of learning curve performed significantly better than another. The performance of 'operation time' as a proxy for learning differed between the three procedures. Multilevel modelling using the laparoscopic cholecystectomy data demonstrated and measured surgeon-specific and confounding effects. The inclusion of non-randomised cases, despite the possible limitations of the method, enhanced the interpretation of learning effects. CONCLUSIONS - HEALTH TECHNOLOGY ASSESSMENT LITERATURE REVIEW: The statistical methods used for assessing learning effects in health technology assessment have been crude and the reporting of studies poor. CONCLUSIONS - NON-HEALTH TECHNOLOGY ASSESSMENT LITERATURE SEARCH: A number of statistical methods for assessing learning effects were identified that had not hitherto been used in health technology assessment. There was a hierarchy of methods for the identification and measurement of learning, and the more sophisticated methods for both have had little if any use in health technology assessment. This demonstrated the value of considering fields outside clinical research when addressing methodological issues in health technology assessment. CONCLUSIONS - TESTING OF STATISTICAL METHODS: It has been demonstrated that the portfolio of techniques identified can enhance investigations of learning curve effects. (ABSTRACT TRUNCATED)