Despite smaller effect sizes, interventions delivered at population level to prevent non-communicable diseases generally have greater reach, impact and equity than those delivered to high-risk groups. Nevertheless, how to shift population behaviour patterns in this way remains one of the greatest uncertainties for research and policy. Evidence about behaviour change interventions that are easier to evaluate tends to overshadow that for population-wide and system-wide approaches that generate and sustain healthier behaviours. Population health interventions are often implemented as natural experiments, which makes their evaluation more complex and unpredictable than a typical randomised controlled trial (RCT). We discuss the growing importance of evaluating natural experiments and their distinctive contribution to the evidence for public health policy. We contrast the established evidence-based practice pathway, in which RCTs generate 'definitive' evidence for particular interventions, with a practice-based evidence pathway in which evaluation can help adjust the compass bearing of existing policy. We propose that intervention studies should focus on reducing critical uncertainties, that non-randomised study designs should be embraced rather than tolerated and that a more nuanced approach to appraising the utility of diverse types of evidence is required. The complex evidence needed to guide public health action is not necessarily the same as that which is needed to provide an unbiased effect size estimate. The practice-based evidence pathway is neither inferior nor merely the best available when all else fails. It is often the only way to generate meaningful evidence to address critical questions about investing in population health interventions.