Breast cancer is the most common cancer and the fifth cause of cancer-related death in women worldwide [1]. A systematic review and meta-analysis reported better survival for those with the highest compared to the lowest level of pre-diagnostic physical activity (HRoverall survival = 0.82, 95% CI: 0.76–0.86; HRbreast cancer-specific survival = 0.86, 95% CI: 0.78–0.94) [2]. Other systematic reviews are consistent [3], [4], [5]. However, studies that attempt to estimate the effect of pre-diagnostic exposures on cancer survival are susceptible to several biases which are seldom considered.

Systematic reviews examining pre-diagnostic physical activity and breast cancer survival include, and sometimes reward, studies that may be affected by bias due to confounding and inappropriate adjustment for post-exposure variables (i.e., variables occurring after the exposure). For example, in one review, studies were downgraded for confounding bias in quality assessments if they did not adjust for cancer-related prognostic factors such as cancer stage and treatment [3]. However, these factors are potential mediators on a causal pathway from pre-diagnostic physical activity to cancer survival (Fig. 1). Physical activity before breast cancer diagnosis may influence the cancer stage at diagnosis and hence the treatment received; both are strong determinants of breast cancer survival [1], [6]. Adjustment for such variables may bias effect estimates towards the null [7], [8]. Additional bias may arise if these variables share common causes with breast cancer survival (V in Fig. 1) [7], [9]. This bias is known as collider stratification bias [7], [9]. Many studies also adjust for adiposity. The temporal relationship between physical activity and adiposity usually cannot be determined if they are measured at the same time [1]. As such, adiposity could either be a confounder (i.e., a member of C in Fig. 1) or a mediator (i.e., a member of L in Fig. 1 that is affected by the exposure) [1], [10], [11], [12]. This highlights the importance of using causal diagrams to identify and distinguish between variables that may be confounders, mediators and/or colliders [7], [8], [9], [13]. In contrast, confounder selection methods that rely on p-values are agnostic to the underlying causal structure [7], [8], [13], [14]. Consequently, their use can lead to inappropriate retention of mediators and colliders, inadequate control for confounding, and ultimately, erroneous effect estimates [7], [8], [13], [14]. In addition, regression – the most common method of confounder adjustment – is inappropriate to control for post-exposure variables affected by the exposure (including cancer-related prognostic factors and members of L in Fig. 1 that may be affected by pre-diagnostic physical activity) if the investigator believes that accounting for such variables is required to obtain valid effect estimates [13]. Previous systematic reviews [2], [3], [4], [5] have not penalised studies that employed inappropriate methods of confounder selection or control.

Studies of pre-diagnostic exposures and cancer survival are susceptible to an in-built selection bias as well as immortal time bias, made clear when applying the target trial approach [15], [16]. This framework involves designing and analyzing an observational study to emulate a pragmatic randomized controlled trial (RCT) [13], [15]. Target trials require the temporal alignment of treatment assignment, eligibility, and time zero of follow-up time, which is not possible for any observational study of a pre-diagnostic exposure and cancer survival [15], [16]. This makes a hypothetical target trial challenging to specify and bias difficult to avoid.

For example, consider a cohort of cancer-free participants in which the pre-diagnostic exposure (in this case, physical activity) is measured at recruitment. This study design is depicted in Fig. 1. Participants become eligible for study of cancer survival at diagnosis. Treatment assignment (which corresponds to the commencement of physical activity before diagnosis) and eligibility are misaligned because eligible participants have survived to a post-exposure event (in this case, to diagnosis) [16]. This introduces selection bias because the likelihood of selection may be affected by the pre-diagnostic exposure and other pre- and post-exposure variables (including members of C and L in Fig. 1) [16], [17]. Physical activity reduces the risks of breast cancer and death from other causes, and women diagnosed with breast cancer who are physically active may be more likely to possess other factors that brought about their cancer despite their physical activity [1], [6], [17], [18]. Additional bias can be introduced in the handling of follow-up time. The approach at least risk of additional bias is to set time zero of follow-up when eligibility criteria are met by participants — that is, to left-truncate follow-up time at cancer diagnosis (Fig. 2a) [16], [19]. If follow-up time commenced at the pre-diagnostic physical activity assessment instead, the analysis is at risk of immortal time bias (Fig. 2b) [16]. This is because all participants are essentially immortal (i.e., their risk of dying is zero) between treatment assignment and eligibility [16].

A common alternative study design at greater risk of selection bias is a longitudinal study that follows from a previous case-control study or RCT. This study design is depicted in Supplementary Fig. 1. Participants are recruited post-diagnosis and pre-diagnostic physical activity is retrospectively assessed. Eligible participants must survive not only to cancer diagnosis, but to the post-diagnostic assessment of pre-diagnostic physical activity. To avoid additional bias, follow-up time should commence at the post-diagnostic assessment (when all eligibility criteria are met), but this means that follow-up time is left-truncated further (Fig. 2c) [16], [17], [19]. If time zero of follow-up aligns with cancer diagnosis instead, participants become immortal between diagnosis and the post-diagnostic assessment of pre-diagnostic physical activity [16] (Fig. 2d).

The objective of this systematic review and meta-analysis is to illustrate and assess the potential impact of different biases in studies of pre-diagnostic physical activity and overall or breast cancer-specific survival in women with breast cancer. These biases include confounding bias, bias due to inappropriate adjustment for post-exposure variables such as cancer-related prognostic factors, selection bias due to participants becoming eligible for study when they have survived to post-exposure events, and immortal time bias due to inappropriate handling of follow-up time. Herein, ‘survival’ refers to survival after cancer diagnosis and beneficial survival outcomes are implied by HRs < 1.