Checklist for the reporting of observational studies
The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) checklist for the reporting of observational studies has been used while writing this article(12).
Study design
This is a multicentre cohort study with eleven participating centres in Sweden. Patients operated between 2014 and 2018 were identified from theatre lists at each hospital. Chart review was conducted at each site and relevant data such as perioperative variables, postoperative highest CRP levels and postoperative anastomotic leakage were entered into a clinical report form in a REDCap database (Vanderbilt University, Tennessee, USA). The REDCap system is a secure web application for building and managing online surveys and databases(13).
The REDCap database and the Swedish Colorectal Cancer Registry (SCRCR) were linked to ensure that all patients included were operated for rectal cancer and to provide additional information such as demographics, tumour stage and height, and additional perioperative data. The SCRCR was established in 1995 and includes all hospitals in Sweden that operate rectal cancer(14). The SCRCR has been validated several times with excellent results, demonstrating an average completeness of 99% and overall agreement between registry and re-abstracted variables at 90% between 2008–2015(15). However, some variables such as anastomotic leakage are known to be underreported(16).
Inclusion and exclusion criteria
All patients having undergone an anterior resection with anastomosis for rectal cancer were included. Exclusion criteria included patients with disseminated disease at surgery and missing data regarding postoperative CRP, anastomotic leakage, T or N stage, or recurrence.
Exposure
Any anastomotic leakage within 12 months after the anterior resection was considered the main exposure. Leakage was defined according to the consensus criteria by the International Study Group for Rectal Cancer (ISREC)(17), where any compromised suture or staple line as well as fistulisation to or from the rectum are considered leaks; isolated pelvic abscesses close to the anastomosis are also included. The leaks were subdivided into grade A (leak without need for any treatment, typically asymptomatic), grade B (leak requiring treatment, excluding laparoscopy or laparotomy), and grade C (leak necessitating laparoscopy or laparotomy). Day of leakage was registered as well, where leaks occurring later than 14 days after surgery were assumed to have taken place during the early postoperative period, even though they wer diagnosed only later. This assumption allowed the highest postoperative CRP level to be ascribed to initially undetected leaks, which is plausible because many patients, especially those with a defunctioned stoma, are diagnosed with leaks several weeks and even months after surgery(18).
Mediator
The mediator of interest in this study is the highest CRP level, measured in mg per L, during the first 14 postoperative days. This was assumed to reflect the systemic inflammatory response of the patient.
Outcome
The primary outcome was recurrence-free survival according to chart review. Secondary outcomes were local and distance recurrence as well as overall survival. Local recurrence was defined as any recurrent tumour in the pelvis, where a strong radiological suspicion was deemed sufficient (a positive biopsy or histopathological specimen was not required). Distant recurrence denoted any tumour spread beyond the pelvis not known before surgery.
Statistical analysis
Frequency tables concerning patient and tumour characteristics were constructed. Continuous variables were described using the median along with the interquartile range (IQR). The association between leak status and recurrence-free as well as overall survival were visualized using Kaplan-Meier curves, where a log rank test was used to test for differences.
Associations between exposure (anastomotic leakage), mediator (postoperative highest CRP) and outcome (5-year recurrence free survival; 5-year overall survival) were estimated by fitting three separate models. The exposure-mediator relationship was estimated by a linear model regressing exposure on the mediator while the exposure-outcome relationship was estimated by a Cox model regressing exposure on outcome. Baseline confounders were included in both models. These confounders were selected with the use of a causal diagram(19) (Supp Fig. 1), and included age (continuous), sex (male or female), American Society of Anesthesiologists’ (ASA) fitness grade (I, II, or III–IV), current smoking (yes or no), neoadjuvant treatment (none, radiotherapy, or chemoradiotherapy), pathological tumour category (T0, T1–2, T3, or T4), pathological node category (N0, or N1–2), surgical technique (open, or minimally invasive), type of mesorectal excision (total, or partial), defunctioning stoma (yes, or no), blood loss (continuous), year of surgery, and annual hospital volume (continuous). A Cox model including both exposure and confounders were used to estimate the association between mediator and outcome.
A mediation analysis was performed to explore to what extent the effect of anastomotic leakage on survival was mediated by postoperative highest CRP. The direct, indirect and total effects were estimated using a natural effects Cox model(20). The direct effect reflects the part of the association between exposure and outcome that is not affected by the mediator, while the indirect effect reflects the part of the association that can be explained by the mediator. As a first step, a parametric survival model, including all confounders and the mediator, was used to model the relationship between exposure and outcome. A duplicate of the exposure variable, exposure*, was then added to the data and a copy of the data was created in which the values in the variable exposure* were set to be counterfactual to the values in the original exposure variable. The data copy was appended to the original data so that the new dataset consisted of twice as many observations as the original data. Then the first step survival model was used to impute survival times conditional on the exposure* values and observed mediator and confounder values. Finally, a Cox model was fitted to the imputed survival times using the original exposure variable, exposure* and baseline confounders. The coefficients of exposure* and exposure are estimates of the natural direct and indirect log hazard ratios. This was done 10 times, and the Cox model fits were pooled using Rubin's rules(21). Confidence intervals of the natural effects were based on 1,000 bootstrap samples(22).
A dataset where the missing values were imputed by multiple imputation by chained equations was used in the analyses(23).
The above analyses were subsequently repeated for a subset of the cohort, where grade A and B leaks were excluded, retaining only patients without anastomotic leakage and a grade C leak group; this was done to evaluate whether severe leakage imparts a more pronounced effect, and potentially also to alleviate diagnostic bias between centres as such leaks were considered straightforward to detect. Finally, a sensitivity analysis was conducted to evaluate whether chronic inflammation, using the proxy of leakage with and without the anastomosis in situ as exposure, would impact recurrence-free survival.
All analyses were performed using R 4.2.3 statistical software(24). The R-package mice(25) was used for imputation.