Screening biases: The case of mammography for breast cancer

Biases can distort the perceived benefits of screening programs. Understanding biases affecting screening research is essential for accurately interpreting evidence and supporting well-informed public health guidelines and decisions. The “Big 3” screening research biases-lead-time bias, length bias, and overdiagnosis-can significantly impact estimates of screening benefits, making it vital to identify them and implement strategies for their prevention, estimation, and correction. Using breast cancer screening mammography as a case study, we demonstrated how bias analysis applies to real-world data, evaluated the extent to which these biases are addressed in the literature, and identified the most common methods used for their evaluation.After defining the “Big 3” screening biases, we performed a systematic PubMed search using adapted algorithms from previous reviews addressing similar research questions. MeSH terms and free-text keywords were chosen to include studies on breast cancer, diagnosis/screening, mammography, randomized clinical trials/observational studies, and bias. We restricted the search to studies of adult women published since 2000 in English or Spanish. We then analyzed how the Big 3 biases are addressed in the literature and the most common methods used for their estimation and correction.We found 3,660 records of studies evaluating mammography for breast cancer screening, of which 163 (<5%) explicitly mentioned bias and were included in the analysis. Among these, 69% assessed bias in the context of mammography as a screening modality. Overdiagnosis was the most frequently evaluated bias (82%), followed by lead-time bias (18%). Only 27% of this latter pool of studies conducted a formal bias analysis.Despite its importance, few mammography studies explicitly address bias. Lead-time bias can overestimate screening's benefits because survival appears longer when the disease is detected earlier. Length bias can skew survival rates in screened groups by overrepresenting long-duration diseases. Overdiagnosis (an unavoidable consequence of secondary prevention programs) can lead to unnecessary diagnoses and treatments. Strategies to address these biases include causal tools like directed acyclic graphs to evaluate lead-time bias and related forms of immortal time bias. Quantitative methods are available to assess length or lead-time bias, which can be summarized for systematic bias correction (e.g., stage-specific proportions). Recent years have seen an increase in studies focusing on overdiagnosis, often employing statistical models such as excess-incidence models or progressive-indolent mixture models to measure it. Based on the evidence, we argue that a comprehensive assessment of screening effectiveness should integrate a harm-benefit analysis while considering bias, therapeutic advances, and broader social determinants of health.