This protocol of a systematic review and meta-analysis follows the Preferred Reporting Items for Systematic Review and the Meta-Analysis Protocols (PRISMA-P) guideline [27].
Eligibility criteria
Studies of female breast cancer patients aged over 18 and diagnosed with breast cancer stage I to III between 1995 and 2019 will be eligible for inclusion. Included studies must be cohort studies of breast cancer patients who have received at least one systematic modality (e.g., chemotherapy) or local therapy (e.g., surgery). Studies addressing advanced or metastatic breast cancer patients will be excluded, as treatment in this group of patients mainly involves palliative care.
To be included, a study must include a measure of race/ethnicity, and the measure must be included in the statistical analysis. Peer-reviewed publications of original cohort study findings that appeared in English-language academic journals will be eligible for inclusion. Editorials, letters, reviews, and preprints will be excluded. Moreover, studies covering data on breast cancer patients during the COVID-19 pandemic will also be excluded, because the treatment modalities and management of breast cancer patients changed based on both individual and health system factors.
In terms of further eligibility criteria for the potential meta-analysis, at least one of the outcomes, either survival related outcomes or TTI (see Outcome measure), must be reported in the findings of included studies. Studies reporting hazard ratio will be eligible for inclusion in the meta-analysis. Only studies with low or moderate risk of bias will be included in the meta-analysis.
Outcome measure
We will consider two outcome measures: survival, and time to treatment initiation. The primary outcome will be survival in breast cancer patients, including overall survival and progression free survival. Survival time can be measured using the difference in time interval between time to treatments and either time to death due to breast cancer or time to follow-up of survivors. Within time-to-event data analysis literature, survival has been mainly presented as 5-year survival or 10-year survival. Studies identified in our preliminary search suggested that overall survival, rather than progression-free survival, was measured as an outcome; this is because in most administrative databases, information on cause of death does not have high validity or has a high frequency of missingness [28, 29]. While overall survival was the most common outcome observed in our preliminary search, some studies did examine progression-free survival or 5-year recurrence-free survival in the context of racial disparity and delay in beginning treatments [28, 15].
Time to treatment initiation (TTI) will be the secondary outcome of the study. The definition of TTI varies across studies; in this review, TTI will be defined as the difference between time to diagnosis and time to initiation of a treatment (e.g., surgery, chemotherapy) [15, 30]. According to the studies’ methods for measuring TTI, cut-off points for determining that there has been a delay in accessing systemic therapies have been established based on stage and hormonal receptors in patients younger than 70. Based on the measures for stage II and III chemotherapy patients, the cut-off is less than 120 days. Although there is no established categorization for TTI, this review will examine three categories of intervals between diagnosis and treatment initiation including less than 30 days, 60 days, and 90 or more days, after diagnosis [14, 15]. However, because previous studies addressed various subjectively-coded categories including binary or multiple categories of time to treatment, this study will also classify binary categories for delay to treatment such as less than 30 days and over 31 days.
Race/ethnicity measure
Self-reported race/ethnicity is considered as the gold standard measurement, because there is a low likelihood of misclassification error [17]. In the U.S, the Centers for Disease Control and Prevention’s National Center for Health Statistics (NCHS) categories for race include: White, Black/African American, American Indian/Alaska Native, Asian Indian, Chinese, Japanese, Korean, Filipino, Vietnamese, Other Asian, Native Hawaiian, Guamanian/Chamorro, Samoan, and Other Pacific Islander [31]. However, as there is no standardized classification of race/ethnicity in studies, the likelihood of misclassification of race and ethnicity can be expected. To avoid this error, narrowing race/ethnicity classification into three categories including Black, white and others may be helpful [10, 11, 16]; however, this review will attempt to maintain more information related to race/ethnicity where possible.
Search strategy
Eligible studies will be identified through searching PubMed, Ovid, Web of science, and the Cochran library. Additionally, a search filtered on the title will be carried out through Google Scholar using keywords ‘breast cancer’ and ‘time to treatment’. An extra search will be manually performed to check the reference lists of included studies. Because the term ‘race’ is not likely to appear in the title/abstract even when a study is relevant to the review, two main keywords will be searched: ‘breast cancer’ and ‘time to treatment’. All databases will be searched using the combination of keywords ‘breast cancer’ and ‘time to treatment’ with their MeSH terms. As an example, search strings in PubMed were as follows: (((((((((((breast cancer[MeSH Terms]) AND (“time to treatment”[Title/Abstract]) OR (“time to surgery”[Title/Abstract]) OR (“time to chemotherapy”[Title/Abstract]) OR (“time to radiotherapy”[Title/Abstract]) OR (“time to hormone therapy”[Title/Abstract]) OR (“time to endocrine therapy”[Title/Abstract]) OR (“delayed treatment”[Title/Abstract] OR (“delayed treatments”[Title/Abstract])). This review will use the Preferred reporting items for systematic reviews and meta-analyses (PRISMA) guideline [32].
Study records
First, duplicated records will be removed from the pooled records using reference manager software (e.g., Mendeley). The remaining studies will be screened based on the title and abstract. Two independent reviewers will extract data. Other reviewers will convene to discuss any uncertainties, including any errors in data collection process or uncertainty regarding inclusion of a study. Finally, in-depth analysis of full texts will be performed to determine the total number of included studies following eligibility criteria. In addition, studies will be searched for correction or retraction notices. Of these, eligible studies will be selected to conduct a meta-analysis. The selection process of studies in the systematic review is shown in Fig. 1.
Risk of bias assessment
The Risk Of Bias In Non-randomised Studies – of Interventions (ROBINS-I) will be used to assess risk of bias for included studies. ROBINS-I is a powerful tool which requires methodological and content knowledge and addresses internal validity. The tool assigns each paper one of five judgments: low risk of bias, moderate risk of bias, serious risk of bias, critical risk of bias and no information [33]. If a study fails to meet low or moderate risk of bias, it will be excluded from the meta-analysis. The risk of bias assessment of included studies will be done by two reviewers and in case of controversy, the third reviewer will review the assessment.
A brief on the ESC-DAGs method
This review will use the Evidence synthesis for constructing directed acyclic graphs (ESC-DAGs) guideline on evidence synthesis for constructing directed acyclic graphs (DAGs) to develop an integrated causal graph that summarizes the causal relationships across the study papers from our systematic review [26]. There is evidence that cohort studies frequently suffer from unmeasured factors (e.g., confounders), and DAGs are often used as tools to aid identification of confounders, as well as to help capture the effects of mediators [34]. The ESC-DAGs method uses three major steps to develop an integrated DAG (Fig. 2). The first step is mapping, which begins by drawing an “implied graph” for each study, based on that study’s results and conclusions, which may include directed and undirected edges. Then, directed edges corresponding to other known factors and structures (e.g., mediators, confounders etc.) affecting the exposure-outcome association are added to the graph based on relationships implied by the findings of that study. The implied graph will be saturated and complex for each study, as it combines all relationships observed in the implied graph from the first stage. Translation of each implied graph into a DAG is the next step that can be applied through various approaches including causal theory and counterfactual thought experiments. In the ESC-DAGs protocol, three criteria of Hill’s guideline, temporality, face validity and resource theory were suggested. The final stage is integration, which includes synthesis and recombination. This stage creates a single DAG by combining the translated DAGs. This is done by combining nodes with the goal of creating consensus and reducing complexity in the final DAG.
The result will be the integrated DAGs which describe the evidence on racial disparity for U.S. adult populations diagnosed with a primary breast cancer.
A workflow to apply ESC-DAG to cohort studies addressing racial disparity
The original ESC-DAG method is presented in general, without regard to study design. Here, we describe in detail our checklist for applying the method to cohort studies, which addresses two issues specific to applying ESC-DAG to cohort studies. First, authors might not specify the causal role of the covariates that are included, for example whether they are considered to be confounder(s) or mediator(s). Because of this, we might not be able to directly draw the causal graph for a study if we do not know the authors’ causal assumptions. Second, we expect that studies will be of variable quality in terms of their risk of bias, and we only wish to include studies in the ESC-DAGs process if their risk of bias is low or moderate.
As part of this protocol, a workflow was created to describe the steps to apply the ESC-DAGs method to the included retrospective studies: 1) assess each study and create a study-specific DAG, 2) modify and/or simplify the DAGs using established causal criteria, 3) build a combined DAG which describes the existing literature and includes prior knowledge about causal theory, and 4) visualize the resulting DAGs (Fig. 3).
Figure 4 illustrates the steps. The checklist begins with determining the PICO (population, intervention, comparison group and the outcome) and assessing risk of bias. Studies will be removed if either PICOs cannot be determined, or if the study is not determined to have low or moderate risk of bias according to the ROBINS-I tool. For the included studies, the checklist follows the three ESC-DAGs stages: i) The mapping stage includes drawing the exposure-outcome arrow, adding other covariates that are connected to the outcome and the exposure as nodes and then drawing related arrows from those control variables in the model towards them and vice versa. ii) The translation stage applies causal theory and counterfactual thought experiments to augment the implied graph resulting from mapping stage. For example, just as the exposure proceeds the outcome of interest, based on temporality, we will recognize which causes occur first and direct arrows accordingly. Additionally, based on causal logic, the impact of an exposure such as race, for instance, can be directly related to income or socioeconomic status (SES), and then to initiation of treatment not directly related to TTI. Finally, unmeasured confounders and potential latent variables can be found in the studies’ limitation and then they can be added as unobserved covariates into the DAGs. iii) The integration stage combines and simplifies all of the translated DAGs by grouping similar nodes to a single node. For example, nodes for income, occupation, and education might be combined into a single node that denotes SES. The final stage of our checklist, which is added to ESC-DAGs, will create integrated DAGs for each outcome of interest in the review. In our review we are considering two outcomes, TTI and survival, so we will create two different integrated DAGs.
The following two examples show how the mapping and translation checklist stages are applied to studies on racial disparity and TTI effects on survival.
Example 1
racial disparity and time to surgery [22]
In a large retrospective cohort study, Jackson et al. examined the differences in time to surgery and low-value care in early-stage non-Hispanic Black and non-Hispanic white breast cancer patients. This study found that time to initiation of surgery (i.e., > two months) was significantly higher among Blacks than in whites. Conduct the mapping stage, based on the variables in the study, a DAG with 29 nodes was created, with the ‘time to surgery’ node and the ‘race’ node considered as the outcome and the exposure, respectively, in the graph. A directed edge was drawn from the outcome node to the exposure node, and edges were drawn from each node corresponding to a variable that was controlled for in the analysis to the exposure node and to the outcome node. We considered genotype as an unmeasured covariate because the genetic information was not measured and was identified by the authors as a limitation. The role of each variable in this study were assessed and as a result, Fig. 5 depicts the incorporation of causal theory into the graph.
Example 2: racial disparity and survival [23]
A retrospective cohort study with large administrative data from National Cancer Database addressed the effect of delay in surgery on the likelihood of upstaging and overall survival in primary breast cancer patients. The results showed a significantly lower overall survival in Black patients compared to white patients, with a hazard ratio of 1.33. As time to surgery goes up, the likelihood of upstaging is likely to occur in patients. This study used 21 covariates in the Cox model. The main exposure was race/ethnicity, and the outcome was survival. All covariates as nodes were included in the graph and a directed edge originated from them was drawn to race and time to death (e.g., survival). The associated implied DAG is shown in Fig. 6.
Challenges of causal effects of race in the context of disparity
In the context of health disparity, the interpretation of “effects of race” can be challenging [35]. Based on counterfactual theory, it is not causally supported to evaluate the effect of race on the health outcomes. Vanderweele and Robinson identify two methods for interpreting race effects in the presence of control variables, depending on whether or not the control variables mediate the effect of race on the outcome [35]. If the included control variables do not mediate the effect of race on the outcome, then the race/ethnicity coefficient in the regression model estimates the direct effect of race, assuming no unmeasured confounders. However, if the control variables mediate the effect of race on the outcome, then the interpretation of the race/ethnicity coefficient will subject to whether interaction between race and mediator is considered in the model or not. If there is no interaction, the interpretation of the race/ethnicity coefficient is the direct effect of race on the outcome, and the mediated effect can be captured using the difference in the coefficients in models with and without mediators [35]. If an interaction between race and the mediator is included in the model, the interpretation of the effect of the race coefficient corresponds to the natural indirect effect and controlled direct effect. Therefore, the interpretation of race/ethnicity effect depends on what controls have included in the regression model and their causal roles. In fact, multiple regression models can be considered to assess the effect of race/ethnicity as different controls combination can capture different aspects of race/ethnicity. This review will consider both methods as studies might include different control variables in the regression model.
Data synthesis
Data obtained from each single study will be synthesized by providing descriptive tables reporting authors’ names, publication year, main objective, sample size and period, defined race/ethnicity group, type of treatment(s), minimum causal components (e.g., covariates included in the multivariable model), and limitations related to internal validity (i.e., selection bias, information bias, and confounding). The findings of included studies will be presented chronologically. Because we expect effect measures to mainly be expressed as hazard ratios, the estimate of the hazard ratio, its 95% confidence interval, number of people in each race/ethnicity category, and any related information will be extracted from studies.
To perform the meta-analysis, based on the results of heterogeneity (i.e., I2 statistic), either a random effect (e.g., if I2 > 50%) or fixed effect model will be used. The potential effect measure will be the hazard ratio (e.g., hazard ratio for all-cause mortality) which will be presented for each single study as well as its pooled estimate with 95% confidence interval. Publication bias of included studies will be assessed using a funnel plot. In addition, 5-or 10-year survival probabilities will be reported from included studies if they are classified by race/ethnicity categories. Finally, as some studies might report odds ratios or regression coefficients, we will convert them to hazard ratios as much as possible.
Two additional analyses are proposed. First, a sensitivity analysis will be performed to evaluate the robustness of primary findings. This will help to identify and exclude low-quality studies if they have a significant impact on the results. As there is a likelihood of different included studies using the same or overlapping data sources, sensitivity analysis can aid to assess whether the proposed systematic review can rely on the results of pooled data or reduced data (i.e., excluding few studies which lead to less robust results). Moreover, subgroup analyses defined by essential risk factors such as stage of disease, hormone receptors (e.g., estrogen receptor, progesterone receptor and HER-2) and age will be conducted if studies provide information on same subcategories of mentioned risk factors. All statistical analyses will be performed using R software.