Types of studies

Randomised controlled trials (RCTs).

Types of participants

Ambulatory people with cancer (i.e., not hospital inpatients during the time of their participation in the trials) of any age (including children) with cancer with no standard indication for prophylactic anticoagulation (e.g. for acute illness, for central venous line placement, perioperatively) or for therapeutic anticoagulation (e.g. for the treatment of deep vein thrombosis (DVT) or pulmonary embolism). Typically, these people are undergoing chemotherapy, target therapy, immunotherapy, or radiotherapy.

Types of interventions

Intervention

• oral pharmacological thromboprophylaxis with VKA (e.g. warfarin)

• oral pharmacological thromboprophylaxis with DOAC (e.g. apixaban, rivaroxaban, edoxaban)

Control

• no pharmacological thromboprophylaxis

We included any comparison of a combination of the three management options listed above. The protocol from original studies should have planned to provide all other co‐interventions (e.g. chemotherapy) similarly.

Types of outcome measures

Electronic searches

The search strategy was part of a comprehensive search for studies of anticoagulation in people with cancer. We did not use language restrictions. We conducted comprehensive searches on 14 June 2021 following the original electronic searches performed in February 2016 (last major search). We electronically searched the following databases: the Cochrane Central Register of Controlled Trials (CENTRAL; 2021, Issue 6) in the Cochrane Library; MEDLINE via Ovid (1946 to June 14 2021 ); and Embase via Ovid (1980 to 2021 week 24). For each database, the search strategies combined MeSH terms and keywords for anticoagulants, terms for cancer, and a search filter for RCTs. An information specialist revised the search strategy in 2020 to ensure it was fit for purpose in the changing clinical environment. We report details on the full search strategies for each of the electronic databases in Appendix 2.

Living systematic review approach: we will update the searches using auto‐alerts on a monthly basis. We will incorporate new evidence rapidly after it is identified, and publish an update report every six months. We will report a minor update in the 'What's new' section, and a major update in the form of an update of the full review. This update of the systematic review is based on the findings of a literature search conducted on 14 June 2021. We will review search methods and strategies approximately yearly, to ensure they reflect any terminology changes in the topic area, or in the databases.

Searching other resources

We handsearched the conference proceedings of the American Society of Clinical Oncology (ASCO, starting with its first volume in 1982 up to June 2021) and of the American Society of Hematology (ASH, starting with its 2003 issue up to June 2021). We also searched ClinicalTrials.gov and WHO International Clinical Trials Registry Platform for ongoing studies. We reviewed the reference lists of studies included in this review and of other relevant systematic reviews. In addition, we contacted experts in the field to check for unpublished and ongoing trials.

Living systematic review approach: we will search on a regular basis the conference proceedings of ASCO and ASH soon after their publications, ClinicalTrials.gov, and WHO International Clinical Trials Registry Platform. As an additional step, we will contact corresponding authors of ongoing studies as they are identified and ask them to advise when results are available. We will continue to review the reference lists for any prospectively identified studies.

Selection of studies

Four pairs of review authors independently screened for eligibility the titles and abstracts of identified articles. We retrieved the full texts of articles judged as potentially eligible by at least one review author. The pairs then independently screened the full‐text articles for eligibility using a standardised form with explicit inclusion and exclusion criteria. We resolved any disagreements through discussion or by consulting another review author.

Living systematic review approach: for the monthly searches, we will immediately screen any new citations retrieved each month. As the first step of monthly screening, we will apply the machine learning classifier (RCT model) available in the Cochrane Register of Studies (CSR‐Web; Wallace 2017). The machine learning classifier currently has a specificity/recall of 99.98%  and it assigns a probability score (from 0 to 100) to each citation for being a true RCT. . Two review authors will independently screen any citations assigned a score from 10 to 100. Citations that score nine or less will be screened by Cochrane Crowd (Cochrane Crowd). Any citations deemed to be potential RCTs by Cochrane Crowd (i.e., scored 10 and above) will be returned to the authors for screening.

Data extraction and management

Pairs of review authors independently extracted data from each included study and resolved any disagreements through discussion. We aimed to collect data related to the following.

Assessment of risk of bias in included studies

We assessed the risk of bias at the study level using Cochrane's risk of bias tool (Higgins 2011). Two review authors independently assessed the methodological quality of each included study and resolved any disagreements through discussion.  In some instances, when details for certain criteria were not clearly reported, we made the following risk of bias judgments:
 

• adequate sequence generation: if sufficient details of sequence generation were not provided, we assumed that a trial is probably randomized if it explicitly states that it has been randomized; hence we judged the risk of bias to be low;

• allocation concealment: if sufficient details of the allocation concealment were not provided, we assumed that a trial is probably not concealed; hence we judged the risk of bias to be high;

• blinding of participants and personnel: we followed specific instructions for estimating unclearly reported blinding status (Akl 2012). In brief, when sufficient details of blinding participants and personnel were not provided, we assumed that they were probably not blinded. However, since the knowledge of the assigned intervention may have led to differential behaviours across intervention groups (e.g. differential dropout, differential cross‐over to an alternative intervention or differential administration of co‐interventions), we judged the risk of bias to be high

• blinding of outcome assessment: we followed specific instructions for estimating unclearly reported blinding status (Akl 2012). In brief, when sufficient details of blinding outcome assessors were not provided, we assumed that they were probably not blinded. However, since the knowledge of the assigned intervention is unlikely to have impacted the assessment of the hard outcomes studied in this review (e.g., mortality, VTE, and bleeding), we judged the risk of bias to be low

• incomplete outcome data: our judgment was based on the comparison between rate of participants with missing data and event rate for the main outcome in each arm. If the rate of participants with missing data was substantively higher than the event rate, then we judged risk of bias to be high. To assess the risk of bias associated with missing data at the meta‐analysis level, please refer to Dealing with missing data section;

• selective reporting: we checked whether the outcomes reported in any available protocol or registration record were all reported in the results section of the main paper. When trial was not registered or had no published protocol, we checked whether all outcomes listed in the methods section were reported in the results section. In that case, we judged risk of bias to be low

• other bias: when trial was stopped early for benefit, we assumed risk of bias to be high

Measures of treatment effect

We analysed risk ratios (RRs) for dichotomous outcomes and mean differences (MD) for continuous data, with 95% confidence intervals (CI).

Unit of analysis issues

The unit of analysis was the participant.

Dealing with missing data

Assessment of heterogeneity

We assessed heterogeneity between trials by visual inspection of forest plots, by estimation of the percentage heterogeneity between trials that could not be ascribed to sampling variation (I² test; Higgins 2003), and by a formal statistical test of the significance of the heterogeneity (Deeks 2001). If there was evidence of substantial heterogeneity, we attempted to investigate the possible reasons for this (see Subgroup analysis and investigation of heterogeneity).

Assessment of reporting biases

We explored whether the study was included in a trial registry and whether a protocol was available. We planned to create funnel plots for outcomes including 10 or more trials.

Data synthesis

For dichotomous data, we calculated the RR separately for each study (DerSimonian 1986; Review Manager 2014). As noted earlier, in the primary meta‐analysis, we used a complete case analysis approach; that is, we excluded participants considered to have missing data (Guyatt 2017). When analysing data related to participants who were reported as not compliant, we attempted to adhere to the principles of intention‐to‐treat (ITT) analysis. We approached the issue of non‐compliance independently from that of missing data (Alshurafa 2012). We then pooled the results of the different studies using a random‐effects model.

Living systematic review approach: whenever new evidence (studies, data or information) that meets the review inclusion criteria is identified, we will immediately assess risk of bias, extract the data, and incorporate it in the synthesis, as appropriate. We will not adjust the meta‐analyses to account for multiple testing, given that the methods related to frequent updating of meta‐analyses are under development (Simmonds 2017).

Subgroup analysis and investigation of heterogeneity

We planned to explore substantial heterogeneity by conducting subgroup analyses based on the type of oral anticoagulant and the characteristics of participants (type and stage of cancer, and whether participants were on cancer treatment or not). In particular, we conducted subgroup analyses for participants with lung cancer (either SCLC or NSCLC) versus participants with non‐lung cancer. We included in the lung versus non‐lung subgroup analysis data from:

• studies that recruited only participants with lung cancer (either SCLC or NSCLC) and studies that recruited only participants with non‐lung cancer;

• studies that recruited both lung and non‐lung cancer participants if they provided data for subgroups of participants with lung cancer AND data for subgroups of participants with non‐lung cancer;

• studies that recruited both participants with lung and non‐lung cancer but did not provide subgroup data, if more than 75% of participants had lung cancer or more than 75% of participants had non‐lung cancer.

Sensitivity analysis

We decided a priori to consider abstracts and completed studies published exclusively on ClinicalTrials.gov in the main analysis only if study authors supplied us with full reports of their methods and results; otherwise, abstracts were included only in the sensitivity analysis. As described earlier, we also planned for sensitivity meta‐analyses to assess the risk of bias associated with missing outcome data when the primary meta‐analysis of a specific outcome found a statistically significant effect.

Summary of findings and assessment of the certainty of the evidence

We assessed the certainty of evidence at the outcome level using the GRADE approach (GRADE handbook). We followed the guidance developed by the GRADE working group to communicate the findings of the systematic review (Santesso 2020).