Source
Institute of Health and Society, Newcastle University, 21 Claremont Place, Newcastle upon Tyne, NE2 4AA, UK. heather.dickinson@ncl.ac.uk
Abstract
BACKGROUND:
Studies included in a related systematic review used a variety of statistical methods to summarise clinical behaviour and to compare proxy (or indirect) and direct (observed) methods of measuring it. The objective of the present review was to assess the validity of these statistical methods and make appropriate recommendations.
METHODS:
Electronic bibliographic databases were searched to identify studies meeting specified inclusion criteria. Potentially relevant studies were screened for inclusion independently by two reviewers. This was followed by systematic abstraction and categorization of statistical methods, as well as critical assessment of these methods.
RESULTS:
Fifteen reports (of 11 studies) met the inclusion criteria. Thirteen analysed individual clinical actions separately and presented a variety of summary statistics: sensitivity was available in eight reports and specificity in six, but four reports treated different actions interchangeably. Seven reports combined several actions into summary measures of behaviour: five reports compared means on direct and proxy measures using analysis of variance or t-tests; four reported the Pearson correlation; none compared direct and proxy measures over the range of their values. Four reports comparing individual items used appropriate statistical methods, but reports that compared summary scores did not.
CONCLUSIONS:
We recommend sensitivity and positive predictive value as statistics to assess agreement of direct and proxy measures of individual clinical actions. Summary measures should be reliable, repeatable, capture a single underlying aspect of behaviour, and map that construct onto a valid measurement scale. The relationship between the direct and proxy measures should be evaluated over the entire range of the direct measure and describe not only the mean of the proxy measure for any specific value of the direct measure, but also the range of variability of the proxy measure. The evidence about the relationship between direct and proxy methods of assessing clinical behaviour is weak.