Types of studies

Randomised controlled trials (RCTs). We will only include RCTs when the agent evaluated is mechanistically described as a PARP inhibitor, used alone or in combination with chemotherapy or other biologic agents (or a combination of these treatment modalities).

Types of participants

Women with locally advanced or metastatic breast cancer treated with PARP inhibitors.

Types of interventions

The intervention will be the use of PARP inhibitors for locally advanced or metastatic breast cancer treatment. The comparator will involve treatment with chemotherapy without PARP inhibitors.

We will include trials involving:

• treatment for progression after multiple lines of chemotherapy in combination with PARP inhibitors, compared to the same chemotherapy without PARP inhibitors;

• trials with placebo groups and trials with open control groups (no treatment or best supportive care controls).

Selected studies will be further assessed if the information given suggests that the study compares chemotherapy in combination with PARP inhibitors, to the same chemotherapy without PARP inhibitors.

Types of outcome measures

Electronic searches

We will conduct systematic literature searches to identify published and unpublished RCTs. Due to the relatively recent availability of PARP inhibitors, the literature search will incorporate a start‐date of year 2008, which will be considered sufficient for the purpose of this review. We will identify RCTs using the following key words: including but not restricted to PARP inhibitors (specific drugs: veliparib, olaparib), and breast cancer.

We will not apply any language restriction in the searches.

We will search the following databases:

1. the Cochrane Breast Cancer Group Specialised Register;

2. the Cochrane Central Register of Controlled Trials (CENTRAL) (The Cochrane Library) (Appendix 1);

3. MEDLINE via OvidSP (Appendix 2);

4. EMBASE via EMBASE.com (Appendix 3);

5. WHO International Clinical Trials Registry Platform (ICTRP) search portal for all prospectively registered and ongoing trials (http://apps.who.int/trialsearch/) (Appendix 4);

6. ClinicalTrials.gov (http://clinicaltrials.gov/) (Appendix 5).

Searching other resources

We will screen reference lists from trial publications selected by electronic searching in order to identify further relevant trials. We will also search published conference abstracts from the following organisations:

• European Society for Medical Oncology (published in the Annals of Oncology);

• European Council for Clinical Oncology (published in the European Journal of Cancer);

• San Antonio Breast Cancer Symposium of the American Association of Cancer Research;

• St. Gallen International Breast Cancer Conference;

• American Society for Clinical Oncology;

• American Society for Clinical Oncology Breast Cancer Symposium.

The search strategy will be constructed by using a combination of subject headings and text words relating to PARP inhibitors in breast cancer. In addition, we will contact members of the relevant cancer research groups, experts in the field and manufacturers of relevant drugs for details of outstanding clinical trials and any relevant unpublished material.

Selection of studies

Two review authors (MK, TT or MR) will independently assess the titles and abstracts retrieved by the search strategy for potential eligibility. We will obtain full‐text articles of potentially eligible studies for further assessment (refer to Types of interventions). The two authors will assess these full articles for quality independently and in a blinded fashion (to authors, journal, drug company, institutions and results), with disagreement resolved by consensus with a third review author (SMP).

We will include abstracts or unpublished data only if sufficient information on the study design, characteristics of participants, interventions and outcomes is available. We will attempt to obtain further information or final results from the primary trial author.

Data extraction and management

Two authors (MK, TT or MR) will perform data extraction independently. We will enter data into the Cochrane Collaboration statistical software, Review Manager 2014. For each eligible trial, we will record the following study characteristics: study design, participants, setting, interventions, quality components, duration of follow‐up, efficacy outcomes, biomarker analyses and side‐effects.

For studies with more than one publication, we will extract data on all outcomes from the most recent publication.

Two review authors (MK, TT or MR) will independently extract details of study population, interventions and outcomes by using a standardised data extraction form, which will be tested in a pilot study. We will resolve differences in data extraction by consensus with a third author (SMP), referring back to the original article.

Our data extraction form will include at least the following items.

• General information: title, authors, source, contact address, country, published/unpublished, language and year of publication, sponsoring of trial.

• Trial characteristics: study design, duration/follow‐up, quality assessment as specified above.

• Patients: inclusion and exclusion criteria, sample size, baseline characteristics, similarity of groups at baseline, withdrawals and losses to follow up.

• Interventions: dose, route and timing of chemotherapy, PARP inhibitors therapy and comparison intervention.

• Outcomes: hazard ratio (HRs) and 95% confidence intervals (CIs), log rank chi square, log rank P values, number of events, number of patients per group, median‐, one, two, three and five‐year survival rates.

Assessment of risk of bias in included studies

Two independent authors (MK, MR) will assess all studies that meet the inclusion criteria for quality, with disagreement resolved by the third review author (SMP). We will assess the risk of bias for every included study using the Cochrane Collaboration's risk of bias tool (Higgins 2011). The Cochrane Collaboration's tool for assessing risk of bias encompasses seven domains, and our judgements on these domains for each trial will be reported in the 'Risk of bias' table. The seven domains are:

1. sequence generation;

2. allocation concealment;

3. blinding of participants, personnel;

4. blinding of outcome assessment;

5. incomplete outcome data;

6. selective outcome reporting (trial protocols and trial result publications will be cross‐checked); and

7. other sources of bias.

In addition, we will also include trials which permit a cross‐over for patients after disease progression. However, in these trials, the number of patients who crossed over has to be considered in the interpretation of the results for overall survival. We will consider the following criteria.

• Was the allocation truly random?

• Were groups similar at baseline regarding the most important prognostic factors?

• Were the number of withdrawals, dropouts and losses to follow up in each group completely described?

• Was the analysis done by intention‐to‐treat?

• Were type and schedule of the follow‐up similar in the comparison group?

We will categorise our answers to the above questions as "low risk", "unclear risk" or "high risk".

Measures of treatment effect

We will perform meta‐analysis on the basis of published data; unpublished or updated data provided will be used when available.

The summary statistics for time‐to‐event outcomes (i.e. overall survival and progression‐free survival) will be hazard ratios (HRs) and their 95% confidence intervals (CIs) (Cox 1972). We will estimate HRs and their 95% CIs directly or indirectly from the published data (Altman 2001). HRs can be estimated (under some assumptions) from log rank Chi square values, from log rank P values, from observed to expected event ratios, from ratios of median survival times or time point survival rates (Machin 1997; Parmar 1998). For overall survival and progression‐free survival, we will calculate a statistical summary of the results of all individual trials performed with one PARP inhibitor, that will be used to compare the results for different drugs. We will use the fixed‐effect model for meta‐analyses to compare treatment differences. We will perform statistical analysis of summary data using Review Manager 2014.

For dichotomous outcomes (e.g. response rates, adverse events), we will express the treatment effect as a risk ratio (RR) with 95% CIs.

For continuous outcomes such as quality of life, we will present data as percent change. If data are not presented in the same format across studies, we will express the treatment effect as the standardised mean difference (SMD) if different scales have been used.

We will carry out intention‐to‐treat analyses for all outcomes.

Unit of analysis issues

There are no unit of analysis issues anticipated in this review.

Dealing with missing data

We will contact study authors to request missing data where possible. Studies with insufficient data for a particular outcome will not be included for analysis of that outcome.

Assessment of heterogeneity

We will first inspect heterogeneity graphically using forest plots displaying effects of individual studies with 95% CIs. We will also assess heterogeneity of effects between studies using the Chi‐squared test and the degree of inconsistency among results of included studies using the I² statistic (I² value > 50% will be considered as substantial heterogeneity and > 75% will represent considerable heterogeneity) (Higgins 2011).

If there is evidence of substantial heterogeneity between studies, we will first investigate whether this heterogeneity can be reduced by examining subgroups of the studies. After review of these analyses, we will make a decision on whether to use a random‐effects model to account for heterogeneity. Otherwise, we will use a fixed‐effect model for the meta‐analysis.

Assessment of reporting biases

We will explore the possibility of publication bias using funnel plots as described in the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011).

Data synthesis

We will extract and enter the number of participants experiencing each outcome and total number of patients randomised to each study arm into Review Manager 2014 for statistical analysis. The data analysis will adhere to the intention‐to‐treat principle.

For time‐to‐event outcomes, we will conduct a fixed‐effect (inverse‐variance method) analysis, if appropriate. If substantial heterogeneity is observed, we will employ a random‐effects (inverse‐variance method) analysis.

For dichotomous outcomes, we will use the fixed‐effect model (Mantel‐Haenszel method) to calculate pooled results if there is no substantial heterogeneity or the random‐effects model (DerSimonian and Laird method) if substantial heterogeneity is identified.

For continuous outcomes, we will conduct a fixed‐effect (inverse‐variance method) analysis, if appropriate. If substantial heterogeneity is observed, we will employ a random‐effects (inverse‐variance method) analysis.

We will perform all analyses using Review Manager 2014 in accordance with the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011).

We will describe the quality of the available evidence in 'Summary of findings' tables in line with recommendations from the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011). We will use the GRADEprofiler (GRADEpro) software to develop the tables (GRADEpro 2014).

Subgroup analysis and investigation of heterogeneity

The specific a priori subgroups include:

1. proportion of triple‐negative breast cancer/basal‐like breast cancer patients;

2. proportion of BRCA mutation carriers;

3. distribution of ER, PR, and HER2;

4. line of therapy;

5. performance status;

6. PARP inhibitors as monotherapy, maintenance therapy or in combination with chemotherapy (or chemo‐radiotherapy) and by their position in the breast cancer treatment paradigm (first line, second line or third line);

7. different PARP inhibitors.

Sensitivity analysis

We will perform a sensitivity analysis excluding studies at high risk of bias.