Protocol development and registration
This protocol has been registered in PROSPERO (CRD42017068299) and is reported in accordance with the Preferred Reporting Items for Systematic review and Meta-analysis Protocols (PRISMA-P) 2015 statement [17]. Representatives of all trials identified during project development (Table 1) have been invited to comment on and contribute to the development of this protocol. For transparency and to safeguard against perception of academic bias, a record is kept of trial investigator comment and the IPD-MA research team is responsible for making methodological decisions. Feedback on project design has also been obtained from consumer representatives.
Inclusion and exclusion criteria
We aim to include all relevant trials irrespective of whether they are published or unpublished, where trials have been carried out, or which language they have been managed and reported in. We aim to include any trial that completed recruitment before July 2016 (allowing 1 year between completion and EPPPIC’s data collection for trial investigators to complete their own analyses). Such trials will be included in any future updates.
Population
Trials including asymptomatic women at increased risk of preterm birth including women who have experienced previous spontaneous preterm birth, a multiple gestation pregnancy, a short cervical length, or a positive fetal fibronectin test at randomization. Trials of progestogen given to women to prevent miscarriage will be excluded. Trials of progestogen administered for immediately threatened preterm birth including premature preterm rupture of membrane or uterine contractions will be excluded. Singleton, twin, and triplet pregnancies will be considered separately.
Intervention
Trials evaluating any form of progestogen are eligible, including natural progesterone and synthetic 17 alpha-hydroxyprogesterone caproate (17-OHPC), delivered by any route including vaginal gels, capsules and suppositories, intramuscular injection, intravenous injection, and oral administration. Different types of progestogen will be analysed separately. Planned and unplanned co-treatments are permitted provided that co-treatments are equally permitted on each intervention arm, or that trials with planned co-interventions make un-confounded comparison (e.g., progestogen + cerclage versus cerclage alone). Trials where administration of progestogen does not continue beyond the 16th week of pregnancy will be excluded.
Comparators
Trials that compare progestogen with placebo or with non-intervention will be included. Trials that compare progestogen with other active interventions such as cerclage will be excluded. Trials that compare different types of progestogen will be included in a secondary network meta-analysis exploring particularly the comparative effectiveness of vaginal progesterone, 17-OHPC, and oral dedroxyprogesterone acetate. A future project may extend to an IPD evaluation of all active interventions for prevention of preterm birth.
Outcomes
All trials that meet the above criteria will be included and contribute to the IPD-MA prespecified outcomes for which they collected data.
Study design
To limit potential for bias, only randomized controlled trials will be included. Quasi-randomized studies will be excluded. Cluster randomized and cross over trials will also be excluded.
Trial identification
As is usual with IPD-meta-analyses, initial literature searches and eligibility screening have been carried out as part of protocol development. This is to ensure that a draft protocol can be sent to trial investigators along with their invitation to partner in the project, and is central to the collaborative approach. Inclusion criteria were established at the outset of the project.
Bibliographic searches of MEDLINE, Embase, CINAHL, and the Maternity and Infant Care databases as well as the Cochrane Pregnancy and Childbirth Review Group's specialized register were carried out during the development phase of the project (this is usual for IPD meta-analyses) and will be re-run at the end. An example MEDLINE search strategy is provided in Appendix A. Trial registers (ClinicalTrials.gov, ISCTRN, and the WHO ICTRP portal) were also searched to identify any unpublished and/or important ongoing trials. Unpublished trials that completed data collection before July 2016 were considered for inclusion. This cut-off was designed to allow trial investigators 1 year after recruitment to complete their own research and analyses. Authors of included trials have also been/will be asked to identify any unpublished trials of which they are aware.
Two researchers independently screened all titles and abstracts retrieved from electronic database and other searches. Full paper publications were then obtained for potentially relevant trials. Where no full paper existed and/or trial eligibility was uncertain, study authors were contacted and asked to provide further information.
Two researchers independently assessed the relevance of each trial using the fullest available information. Any discrepancies in screening decisions were resolved by consensus and discussion with a senior team member or clinician, as required.
‘Near miss’ studies that do not meet all of the inclusion criteria and have therefore been excluded from EPPPIC will be tabulated, and their bibliographic details listed with reasons for exclusion in the final EPPPIC report and PRISMA diagram.
Data provision and coding
Trial investigators will be invited to supply data in a standardized format using standardized coding developed for EPPPIC. However, data will be accepted in any reasonable format and re-coded as necessary by the research team. Data will be requested for all women randomized, including any who were excluded from original trial analyses. Trial protocols and forms will also be collected. A list of data items to be collected is given in Appendix B.
Data supplied will have all names and identifying numbers removed. Individuals will either be labeled with numbers known only to the original trial team or numbered sequentially and trial investigators will be asked to keep a record of these numbers. This will enable any data queries to be traced back to the appropriate individual.
Data storage and confidentiality
All IPD will be received via secure online transfer, or any other secure method such as secure FTP transfer or encrypted email. All data will be anonymous and held in a password-protected area of the Centre for Reviews and Dissemination's (CRD) server. No attempt will be made to re-identify participants and in the unlikely case of re-identification, confidentiality will be maintained. Access will be limited to staff working directly on the project. Copying data to laptop computers or memory sticks will be prohibited.
Critical appraisal, data checking, and quality assurance
Critical appraisal and assessment of data quality will be based on trial protocols and publications and on IPD checking. Risk of bias will be assessed using the Cochrane risk-of-bias tool (RoB) [18]. Assessment will be undertaken by one researcher and if discrepant from RoB assessments reported in previous Cochrane [7] or other systematic reviews, will be checked independently by a second. Two researchers will independently assess trials not previously assessed and reported in a Cochrane or other good quality systematic review. Any disagreements will be discussed with a senior member of the EPPPIC IPD-MA research team.
All IPD will be checked on receipt. Data will be checked for internal consistency, baseline imbalance, and integrity of randomization. Patterns of missing data will be examined. Baseline data will be tabulated and compared with the trial publication and any inconsistencies noted. One researcher will run data checks, which will be independently checked by a second researcher. Findings of all data checking will be discussed with senior members of the research team.
Each individual trial will be analyzed (main outcomes only) and compared with corresponding published analyses (bearing in mind that there may be reasonable discrepancies, if, for example, previously excluded participants have been reinstated in the analyses, or additional follow up data have been provided). Any problems, uncertainties, or queries will be passed back to the responsible trial investigator for explanation and discussion.
Results of data checking may up- or down-weight implications of RoB assessments, for example, data checks may show that there is no evidence that risk of bias arising from the method of randomization has been realized. Any datasets that are judged to be of insufficient quality or completeness will be excluded from the analyses. This may be for the trial as a whole or for particular outcomes or analyses, depending on the nature of the problem.
Data description
Descriptive tables and a narrative summary outlining the key design features and demographic characteristics of each included trial will be produced. Excluded trials will be listed with reasons for exclusion.
Planned analyses
A detailed statistical analysis plan will be developed when the extent of available data is known, but before starting meta-analysis.
Analyses will evaluate overall effectiveness of main and additional outcomes on an intention to treat basis, that is, participants will be analyzed according to allocated treatment, irrespective of whether the treatment was received.
Three parallel but separate analyses focusing on (i) singleton and (ii) twin and (iii) triplet pregnancies will be undertaken. If there is no evidence of differences in effectiveness of progestogen in these three types of pregnancy, further exploratory analyses will incorporate all pregnancy types in a multi-level model. Twin and triplet pregnancies may be combined if there are insufficient triplet data for a viable analysis, and there is no clear evidence of differences between twin and triplet pregnancies. Separate analyses will be carried out for each type of progestogen. Further exploratory analyses may combine types of progestogen if there is no evidence of differences in effectiveness between them.
Maternal, pregnancy, neonatal, and longer-term developmental outcomes will be considered. A separate but complementary focus group study on views of women who have experienced preterm birth will identify outcomes that are especially important to pregnant women and their families, and which therefore lend themselves to shared decision-making.
Outcomes align broadly with the Core Outcomes in Women’s and Newborn health (CROWN) list [19]. Consideration of additional outcomes reflects the opportunity that IPD-MA provides to leverage maximum information from preexisting data. As many outcomes are being explored, findings will be interpreted with the knowledge that there is a high likelihood that statistically significant results will arise by chance. Findings will therefore be interpreted cautiously and in the round, with biological plausibility and consistency across related outcomes and across trials lending credence; and inconsistency suggesting that results may be spurious and should be interpreted conservatively.
A final list of outcomes to be addressed will depend on what variables trials have collected and are available from the included trials. Outcomes with low numbers of events may be combined into more general categories for formal analysis. Appendix B lists the data items that will be requested, including baseline and outcome variables.
Main outcomes
-
Preterm birth or fetal death (< =37, <= 34, <= 28 weeks)
-
Serious neonatal complications or fetal/infant death
(broncho-pulmonary dysplasia, severe intraventricular hemorrhage, nectrotizing enterocolitis, confirmed sepsis, patent ductus arteriosus, retinopathy of prematurity)
-
Neurosensory disability (measured at 18 months or later) or infant/child death
(cerebral palsy, visual impairment, hearing impairment, epilepsy, intellectual impairment, developmental delay)
-
Important maternal morbidity or maternal death
(gestational diabetes, gestational hypertension, preeclampsia, maternal infection)
-
Fetal/infant death
(fetal death occurring at any point after trial entry, stillbirth, or death of live born infant before hospital discharge following birth, or to 28 days, whichever is longer)
With component outcomes defined as follows
Additional outcomes
-
Gestational age at birth
-
Prolongation of pregnancy (interval between randomization and delivery)
-
Prelabor spontaneous rupture of membranes (≤28, ≤34, ≤37 weeks and >37 weeks)
-
Non-spontaneous onset of labor (medically induced or caesarian before labor)
-
Mode of birth (vaginal/caesarian birth)
-
Gestational diabetes
-
Preeclampsia
-
Gestational hypertension
-
Maternal infection
(chorioamnionitis during labor; intrapartum fever, postpartum fever requiring antibiotics)
-
Adverse/side effects of treatment
-
Maternal death
-
Fetal death/stillbirth (death of fetus after randomization)
-
Death of newborn (death of live born infant to hospital discharge following birth)
-
Birth weight (adjusted for gestational age and sex)
-
Admission to neonatal intensive or special care unit (NICU/SCU)
-
Length of stay in NICU/SCU
-
Respiratory distress syndrome (as defined in the trial)
-
Use of respiratory support (mechanical, CPAP, high flow nasal cannula)
-
Bronchopulmonary dysplasia (as defined in trial)
-
Periventricular leukomalacia
-
Necrotizing enterocolitis (grade II or III)
-
Neonatal infection (antenatal, early, late; as defined in trial)
-
Confirmed neonatal sepsis
-
Patent ductus arteriosus (treated for)
-
Severe intraventricular hemorrhage (grade III or IV)
-
Retinopathy of prematurity (stage 3 or worse)
-
Major congenital anomalies including cardiac malformation
-
Death and cause of death after discharge (from hospital following birth)
-
Cerebral palsy
-
Visual impairment (visual acuity worse than 6/60, 20/200 in better eye)
-
Hearing impairment (requiring amplification or worse)
-
Epilepsy
-
Intellectual impairment (as defined in trial)
-
Developmental delay (mild, moderate, or severe on Bailey scale or equivalent)
-
Growth outcomes (including height, weight, head circumference, if possible with reference to appropriate growth chart centiles)
Neurosensory disabilities will be collected individually, but it is anticipated that events will be few and that they will need to be combined for analysis.
We will also analyze an exploratory composite outcome of birth after 37 weeks of gestation of a surviving baby with no serious neonatal complication or neurosensory disability and with a surviving mother with no long-term adverse events.
Analysis of potential effect modifiers
Potential effect modifiers will be investigated to explore whether any particular therapeutic approaches are more effective than others, and/or whether there are particular types of women who derive greater benefit (or harm) from intervention.
Trial-level intervention characteristics
Participant-level maternal and pregnancy characteristics at trial entry
-
Previous spontaneous preterm birth
-
Gestational age at previous preterm birth
-
Cervical length (continuous variable and using thresholds ≤15 mm, ≤20 mm, ≤25 mm, ≤30)
-
Positive fetal fibronectin test
(continuous variable and using thresholds ≥ 10 ng/ml, ≥ 50 ng/ml, ≥ 200 ng/ml)
-
Singleton or multiple gestation pregnancy
(singleton/multiple and singleton/twin/triplet gestation pregnancy)
-
Chorionicity and amnionicity
-
Gestational age at randomization/initiation of treatment
-
Maternal age
-
Ethnicity (within comparable geographical locations)
-
Assisted conception
-
Chronic hypertension
-
Prepregnancy diabetes
-
Smoking during pregnancy
-
Maternal body mass index (categorized using WHO definitions)
Sensitivity and supplementary analyses
Sensitivity analyses will assess the impact of trial design features and alternative approaches to synthesis. In addition to those listed, further analyses may be undertaken where principle analyses suggest that further investigation may be informative. All analyses will be described according to whether they were principle or sensitivity analyses and whether they were preplanned or ad hoc.
-
Supplemented (where possible) with aggregate data from published reports of trials for which IPD is not available
-
Restricted to trials at low risk of bias with respect to design features such as blinding
-
Restricted to trials that have been prospectively registered
-
Analysis of cervical length restricted to trials that used ultrasound measurement
-
Separate analyses of trials measuring developmental outcomes directly and those that used questionnaires
-
Separate analyses of trials measuring intellectual impairment based on health records and those that used parent reports
-
Multivariable analysis including singleton and multiple gestation pregnancies
-
Analysis of ethnicity across all trial locations
-
Analyses of covariate treatment interaction for additional outcomes where analyses of main outcomes suggests that further analysis may be informative
Statistical methods
Outcome measures
Dichotomous outcomes will be analyzed by calculating the risk ratio for the effect of progestogen compared to the control treatment (placebo or usual care). Odds ratios may be used where risk ratios cannot be computed. For continuous outcomes mean differences between treatment arms will be reported. Hazard ratios will be calculated for time-to-event outcomes.
Unit of analysis
Maternal and birth outcomes will use the pregnancy as the unit of analysis. Infant outcomes will use the baby as the unit of analysis. Because infants within a multiple pregnancy are more similar to each other than to babies in other pregnancies (statistically, they are not independent), analyses will be adjusted for this clustering (provided that the additional complexity does not make one-stage models computationally intractable) [20].
One- and two-stage models
The IPD will be synthesized across trials using meta-analysis. Both ‘two-stage’ models (where effect estimates are calculated for each trial, and subsequently pooled in a meta-analysis) and ‘one-stage’ models (where all IPD from all trials are analyzed in one step, while accounting for the clustering of participants within trials) will be used.
Two-stage models
Two stage models will be used as an initial analysis for all the main outcomes listed previously. Effect estimates (relative risk (RR), mean difference (MD), or hazard ratio (HR)) will be estimated for each trial, and then combined using random effects meta-analysis. This will generate forest plots enabling results across trials to be compared visually, heterogeneity investigated and differences across subgroups visualized. All of these aspects will be essential in gaining a full understanding of the underlying dataset, and will motivate the choice of more complex one-stage models [21]. Heterogeneity will be quantified using the I
2 statistic.
One-stage models
One-stage models will be fitted for all outcomes. One-stage analyses will pool IPD from all trials using a generalized linear mixed model framework, which accounts for potential heterogeneity across trials. For continuous outcomes, linear mixed-effect models will be fitted. For dichotomous outcomes, logistic mixed-effect models will be used to calculate relative risks or odds ratios where relative risks are computationally intractable.
Impact of trial and participant characteristics
The impact of the trial and patient-level characteristics on treatment effect (that is, treatment–covariate interactions) will be examined. For trial-level covariates, the trials will be divided into groups according to the characteristic and meta-analyses performed within each subgroup. The more formal analysis of interactions will use one-stage models, where treatment covariate interactions will be added to existing one-stage models for treatment effect. This will enable us to take account of multiple participant characteristics when comparing progestogen with placebo or usual care (stratified by trial), and will also enable exploration of potential treatment interactions in a multivariable way. Models will be compared in terms of goodness of fit and parsimony using the Akaike information Criterion (AIC).
Time to event analyses
For the analyses of whether progestogen prolongs pregnancy, the hazard ratio for the effect of progesterone will be calculated within each trial by fitting a Cox proportional hazards model. Tests for proportional hazards will be performed. Hazard ratios will be pooled across trials using random effects meta-analysis. If the proportional hazards assumption is reasonable, one-stage random effects Cox models will be fitted. Treatment–covariate interactions will be included in these models as required.
Relative and absolute differences
Absolute differences will be calculated by applying the resulting risk ratios or hazard ratios to appropriate baseline incidences (calculated from suitable meta-analyses across the trial control arms). Numbers needed to treat and numbers needed to harm will similarly be calculated for a range of plausible baseline measures.
Unavailable trials and missing data
Every effort will be made to minimize the amount of missing data, including requesting information for any randomized participants that were excluded from the original trial analyses. Where IPD cannot be obtained for a trial, and where possible, aggregate data will be extracted from publications and combined with the IPD-MA results in a sensitivity analysis. Where covariate data are missing for some participants, a complete case analysis will be used in the first instance (i.e., excluding patients with missing data). If there are substantial missing data (around 10% for any outcome or covariate), multiple imputation within each trial will be used to impute missing covariates, where this is computationally feasible. Trials that have not recorded particular outcomes or particular covariates will not contribute to those analyses. Sensitivity analyses based on best and worst case scenarios will be used to assess the impact of missing outcome data.
Network meta-analysis
If sufficient suitable data are available, a network meta-analysis will compare all types of progesterone and routes of administration. This will incorporate direct evidence from head to head trials and indirect evidence from trials comparing each type of progestogen with no intervention. Analyses will be conducted for the main outcomes listed earlier.
Two statistical models will be used: first, the Bayesian models of Lu and Ades [22], which are the most commonly used methods for network meta-analysis. The one-stage meta-analysis models described above will also be extended to include multiple treatment arms. The results of the two approaches will be compared. Both approaches will use random effects to account for heterogeneity. Potential network inconsistency will be investigated by comparing results to results from direct pairwise meta-analyses. If there is evidence of differences, node-splitting models will be used to investigate inconsistency further.
Software
All analyses will be performed at CRD using the R software package [23]. Two-stage analyses will additionally use the meta and metafor libraries [24, 25]. One-stage models will be fitted via the lme4 library, and one-stage Cox models will use the coxme library. Forest plots will be produced using in-house R code. For the network meta-analysis, WinBugs and the GeMTC package in R will be used. https://www.mrc-bsu.cam.ac.uk/software/bugs/the-bugs-project-winbugs/
Reporting
Results will be presented and discussed at a dedicated meeting of the EPPPIC group. Discussion will inform the interpretation of results and development of the final report. An audited list of comments received, and actions taken will be maintained. The results of the IPD-MA will be reported in accordance with PRISMA-IPD [26]. Authorship of the final journal publication will be by the EPPPIC group. Plain language summaries of findings will be produced.
Data repository
Our aspiration is that at the conclusion of our analyses, data will be archived in a repository that will be developed and curated under the aegis of the Cochrane Pregnancy and Childbirth Group. This would establish the nucleus of a data repository that can grow over time and underpin future IPD syntheses and other research projects in pregnancy and childbirth topics. Trial investigators will be given the opportunity to elect to share their data supplied to EPPPIC with the repository under a range of options that the repository will offer. No further work would be required to share data in this way. Participation in the repository is optional. Trialists who participate in the EPPPIC group and IPD-MA are not required to contribute their data to the repository.