Only variables occurring in >2.5% of cases or controls were analysed. Differences between cases and controls were analysed using conditional logistic regression. Variables associated with cancer in univariable analyses, with a P-value <0.1 entered the multivariable analysis. In the multivariable analyses, a P-value <0.05 was used as a significance threshold.
How this fits in
Most prostate cancers in the UK present with symptoms. The risk of cancer posed by these symptoms is largely unknown, so it is difficult for a GP to advise on the need for further testing. Most lower urinary tract symptoms had a risk of cancer in the order of 3% suggesting that further testing for prostate cancer is warranted. A new finding was of the link between previous impotence and later prostate cancer.
The results from PSA testing were not used in the multivariable modelling. PSA testing in this study was largely undertaken after presentation with symptoms, and so after prostatic cancer was suspected. Furthermore, by excluding PSA results from the main analysis, the resulting model can act as a guide as to whether to measure a PSA.
Modelling was performed in stages, first collecting similar variables together, such as those which could represent urinary obstruction. These were then analysed to identify variables to progress to the second stage. These variables were re-grouped into symptoms, signs and investigations. Further multivariable analyses were then performed. Using this approach, a final model was derived including all the variables independently associated with prostate cancer. All discarded variables were then checked against the final model. Finally, seven clinically plausible interactions were tested. Analyses were repeated excluding data from the last 180 days of the 730-day period studied. A third analysis of the effect of verification bias excluded the 39 patients who had clinically unsuspected cancer found solely by histology of material from a prostatectomy.11
