AON's test for Job security behaviour doesn't fit with accident rates

I'm working in a consulting company as an intern, and I was given the task to check if the AON test for worker security and trustworthiness was a good predictor for accident rates within the same population. I found that none of the provided factors were strongly coorrelated with the empirical data that the company gave us. I found a strong heteroskedacity problem in the population and an estructural break tested with Chow. We are everyday more tempted to dismiss the problem as an internal test issue, but i would love some insight on what might be causing this. The test is meant to evaluate the insecure tendencies of the employees but it doesn't correlate with the data.