It reduces these studies' power to detect results. If you thought you had 100 cases and 100 controls, it turns out perhaps you really had 30 cases and 100 controls, and 70 controls misclassified as cases. Power is going to be substantially hurt.
I could see it having more implications than just measurement error reducing power. It's entirely plausible that a drug which has an effect on Alzheimers would come with side-effects that would damage those who just have senile dementia; imagine a drug which dissolves amyloid plaques while doing some moderate damage, in a trial with a mix of dementia and Alzheimers, the Alzheimers getting better is masked by the dementia ones getting worse, for a net average zero compared to the placebo group.
However, it may significantly increase the size of measured effect if you can go back and re-classify the patients - in your example, it may turn an earlier study where only 30 of 100 cases showed improvement to a study where 30 of 30 cases responded to treatment well.
