This two-part review is intended principally for practising clinicians who want to know why some types of evidence about the effects of treatment on survival, and on other major aspects of chronic disease outcome, are much more reliable than others. Although there are a few striking examples of treatments for serious disease which really do work extremely well, most claims for big improvements turn out to be evanescent. Unrealistic expectations about the chances of discovering large treatment effects could misleadingly suggest that evidence from small randomised trials or from non-randomised studies will suffice. By contrast, the reliable assessment of any more moderate effects of treatment on major outcomes--which are usually all that can realistically be expected from most treatments for most common serious conditions--requires studies that guarantee both strict control of bias (which, in general, requires proper randomisation and appropriate analysis, with no unduly data-dependent emphasis on specific parts of the overall evidence) and strict control of random error (which, in general, requires large numbers of deaths or of some other relevant outcome). Past failures to produce such evidence, and to interpret it appropriately, have already led to many premature deaths and much unnecessary suffering.