Establishing the protocol validity of an electronic standardised measuring instrument
MetadataShow full item record
Over the past few decades, the nature of work has undergone remarkable changes, resulting in a shift from manual demands to mental and emotional demands on employees. In order to manage these demands and optimise employee performance, organisations use well-being surveys to guide their interventions. Because these interventions have a drastic financial implication it is important to ensure the validity and reliability of the results. However, even if a validated measuring instrument is used, the problem remains that wellness audits might be reliable, valid and equivalent when the results of a group of people are analysed, but cannot be guaranteed for each individual. It is therefore important to determine the validity and reliability of individual measurements (i.e. protocol validity). However, little information exists concerning the efficiency of different methods to evaluate protocol validity. The general objective of this study was to establish an efficient, real-time method/indicator for determining protocol validity in web-based instruments. The study sample consisted of 14 592 participants from several industries in South Africa and was extracted from a work-related well-being survey archive. A protocol validity indicator that detects random responses was developed and evaluated. It was also investigated whether Item Response Theory (IRT) fit statistics have the potential to serve as protocol validity indicators and this was compared to the newly developed protocol validity indicator. The developed protocol validity indicator makes use of neural networks to predict whether cases have protocol validity. A neural network was trained on a large non-random sample and a computer-generated random sample. The neural network was then cross-validated to see whether posterior cases can be accurately classified as belonging to the random or non-random sample. The neural network proved to be effective in detecting 86,39% of the random responses and 85,85% of the non-random responses correctly. Analyses on the misclassified cases demonstrated that the neural network was accurate because non-random classified cases were in fact valid and reliable, while random classified cases showed a problematic factor structure and low internal consistency. Neural networks proved to be an effective technique for the detection of potential invalid and unreliable cases in electronic well-being surveys. Subsequently, the protocol validity detection capability of IRT fit statistics was investigated. The fit statistics were calculated for the study population and for random generated data with a uniform distribution. In both the study population and the random data, cases with higher outfit statistics showed problems with validity and reliability. When compared to the neural network technique, the fit statistics suggested that the neural network was more effective in classifying non-random cases than it was in classifying random cases. Overall, the fit statistics proved to be effective indicators of protocol invalidity (rather than validity) provided that some additional measures be imposed. Recommendations were made for the organisation as well as with a view to future research.