WP3: Quality assurance
WP leader: Hans Kromhout, UU
The main approach to quality control during the project is by peer-review within the project as the Consortium is composed of well-qualified scientists with considerable experience in designing and conducting epidemiological studies and statistical analyses. The ultimate quality check of the study will be the peer-review process when manuscripts are submitted to scientific journals for publication.
To ensure the highest quality and completeness of information collected and developed, this work package will: verify the brain tumour diagnosis; verify the location of the brain tumour; validate the questionnaire data; and assess the quality of self-reported phone use; validate exposure assessment models as well as individual exposure estimates; and conduct quality control on the data analyses.
Validation of study data analyses
We will compare self-reported and operator-recorded number and duration of calls. The level of agreement between self-reported and operator-recorded phone use will be assessed on a continuous scale as well as by categories.
Multivariate linear regression will be used to test for differences in the log-ratio between cases and controls, adjusting for the effects of country, age, and sex. Regression analyses will be conducted including explanatory variables from the interview. The regression analyses will include random effects accounting for multiple, correlated periods within one subject.
The graphical method of Bland and Altman will be used to assess the relationship between level of use and difference or ratio of measurements obtained by the different approaches.
Impact of potential biases and errors on risk estimate
Sensitivity analyses will be conducted to provide an idea of the uncertainty of the main outcomes of the study due to the quality of the data collection, exposure models, and exposure assignment.
The potential impact of differential and non-differential recall errors and biases on risk estimates will be evaluated based on the results of the validation study. The potential impact of selection bias on risk estimates will also be evaluated, using the information from non-response questionnaires and making assumptions about phone use by other non-participants. As in INTEPRHONE, “bias factors” will be calculated under these different scenarios following the method described by Greenland and others.
A thorough bias analysis, simultaneously accounting for selection and recall bias, other misclassification, and random error will also be attempted, using a Bayesian approach and asserting reasonable prior distributions and simulations to include the random error. Finally, the impact of uncertainties in the RF and ELF exposure estimates (from WP3) will be evaluated either through Bayesian analyses or with Monte-Carlo Maximum-Likelihood approaches.