Types of studies

We included randomized controlled trials and quasi‐randomized trials at the level of each twin pair or at the level of each neonatal unit (neonatal intensive care unit [NICU]‐cluster trials). We excluded cross‐over trials, as we considered it highly likely that the major outcomes examined, such as weight gain and neurobehavior, would be significantly "contaminated" by the cross‐over.

Types of participants

We included studies that enrolled pairs of twins of less than 37 weeks' gestation who were considered clinically stable at study entry. We excluded studies that exclusively enrolled infants under intensive care (including those needing any type of respiratory support apart from nasal prong, low‐flow oxygen supplement, and/or any type of circulatory support) or those clinically considered septic at study entry or with congenital anomaly or open skin wounds. For studies with a mixture of stable and less stable infants, we planned to retrieve and analyze only data from stable infants; we planned to contact the primary study authors if necessary.

We placed no limit on weight and did not restrict the lower limit of PMA.

We included both growth‐concordant and growth‐discordant twins in the overall analysis and separated significantly growth‐discordant twins from others in subgroup analyses. We defined significant growth discordance in this review as an intra‐pair difference in birth weight greater than 30%, with the weight of the larger twin taken as the denominator, as weight discordance beyond this level has been consistently associated with worsening of clinically important outcomes (Cheung 1995; Yalcin 1998; Branum 2003; Bagchi 2006).

Participants were eligible regardless of their postnatal age as soon as they were considered clinically stable by the attending physicians.

Types of interventions

Types of outcome measures

We have listed units of analysis for the following outcomes for studies in which randomization occurred at the level of each twin pair. We have provided additional details on how we handled units of analysis under the heading Unit of analysis issues.

Electronic searches

We conducted a comprehensive search of the Cochrane Central Register of Controlled Trials (CENTRAL; 2016, Issue 2); MEDLINE via PubMed (1966 to February 29, 2016); EMBASE (1980 to February 29, 2016); and the Cumulative Index to Nursing and Allied Health Literature (CINAHL) (1982 to February 29, 2016) using the following search terms: (Nurseries OR Neonatal nursing OR NICU) AND (cobed* OR (co bed*) OR (bed shar*) OR (co sleep*), plus database‐specific limiters for randomized controlled trials (RCTs) and neonates (see Appendix 1 for the full search strategy for each database). We applied no language restrictions.

We searched clinical trials registries for ongoing or recently completed trials (clinicaltrials.gov, the World Health Organization International Trials Registry and Platform [www.whoint/ictrp/search/en/], and the International Standard Registered Clinical Trial Number [ISRCTN] Registry [ISRCTN Registry]).

We accepted studies whether published or unpublished, in full article or abstract form, as long as assessment of study quality was possible and other inclusion criteria had been fulfilled.

Searching other resources

We contacted the authors of all selected studies about ongoing trials and unpublished data. We searched the references cited in previous relevant Cochrane reviews and in other relevant papers, review articles, and Internet documents. We assessed the results of handsearching by the Cochrane Neonatal Review Group. We searched the proceedings of relevant pediatric and perinatal conferences (specifically, those of the American Academy of Pediatrics, the European Academy of Pediatrics, the European Society of Paediatric Research (ESPR), the United Kingdom Perinatal Society, and the Perinatal Society of Australia and New Zealand) to identify relevant published abstracts.

Quality of evidence

We assessed the quality of evidence for the main comparison at the outcome level using the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) approach (Guyatt 2011a). This methodological approach considers evidence from randomized controlled trials to be of high quality that may be downgraded on the basis of consideration of any of five areas: design (risk of bias), consistency across studies, directness of evidence, precision of estimates, and presence of publication bias (Guyatt 2011a). The GRADE approach assigns the quality of a body of evidence to one of the following grades: High quality: We are very confident that the true effect lies close to that of the estimate of the effect; Moderate quality: We are moderately confident in the effect estimate: The true effect is likely to be close to the estimate of the effect, but it may be substantially different; Low quality: Our confidence in the effect estimate is limited: The true effect may be substantially different from the estimate of the effect; and Very low quality: We have very little confidence in the effect estimate: The true effect is likely to be substantially different from the estimate of effect (Schünemann 2013).

Review authors independently assessed the quality of the evidence for outcomes identified as critical or important for clinical decision making. These outcomes include rate of weight gain (in gram/kg of baseline weight/d), average from study entry to week three; apnea, bradycardia, or desaturation; episodes in the co‐regulated state (out of 20 observations); infections (any suspected or proven infection); length of hospital stay (in days); parental satisfaction; and infant pain score following a painful procedure (at 30 seconds post heel lance).

In cases in which we considered that the risk of bias resulted from inadequacies in concealment of allocation, randomized assignment, complete follow‐up, or blinded outcome assessment, our confidence in effect estimates was reduced, and we downgraded the quality of evidence accordingly (Guyatt 2011b). We evaluated consistency by noting similarity of point estimates, extent of overlap of confidence intervals, and statistical criteria, such as measurement of heterogeneity (I²). We downgraded the quality of evidence when large and unexplained inconsistency was observed across studies (i.e. some studies suggest important benefit, and others no effect or harm without a clinical explanation) (Guyatt 2011d). We assessed precision on the basis of width of the 95% confidence interval (CI) and the calculated optimal information size (OIS). If the total number of participants included in the pooled effect estimation was less than the number of participants generated by a conventional sample size calculation for a single adequately powered trial, we considered rating the quality of evidence down for imprecision (Guyatt 2011c). When trials were conducted in populations other than the target population, we downgraded the quality of evidence because of indirectness (Guyatt 2011e).

We entered data (i.e. pooled estimates of effects and corresponding 95% CIs) and explicit judgments for each of the above assessed aspects into the Guideline Development Tool ‐ the software used to create "Summary of findings" tables (GRADEpro). We explained all judgments involving assessment of the study characteristics described above in footnotes or comments provided in the "Summary of findings" tables.

Selection of studies

We used standard methods of The Cochrane Collaboration and its Neonatal Review Group. All review authors independently assessed studies and determined their eligibility by using a form with predefined inclusion criteria. We excluded any study that did not fit the criteria and resolved disagreements by discussion.

Data extraction and management

Three review authors (NML, SCF, and WCF) independently entered individual data from each included trial onto a standardized data collection form. Data entered were compared, and differences were resolved by consensus.

Assessment of risk of bias in included studies

Two review authors (NML and SCF) independently assessed risks of bias of included trials, using standard methods of The Cochrane Collaboration as presented in the Cochrane Handbook for Systematic Reviews of Interventions Version 5.1 (Higgins 2011).

Specifically, we assessed trials on the basis of the following: adequacy of sequence generation, allocation concealment, blinding to the intervention, completeness of follow‐up, and blinding to outcome measurement.

Had data been available from cluster‐randomized trials, we planned to assess the following specific additional biases when applicable: recruitment bias, baseline imbalance, loss of clusters or individuals within clusters, incorrect unit of analysis (as elaborated in the corresponding section), and compatibility with individually randomized trials.

Information on risks of bias was included in the Characteristics of included studies table.

In addition, we evaluated the following issues and made a judgment regarding high, unclear, or low risk of bias.

• Sequence generation: Was the allocation sequence adequately generated? 

• Allocation concealment: Was allocation adequately concealed? 

• Blinding of participants, personnel, and outcome assessors: Was knowledge of the allocated intervention adequately prevented during the study? At study entry? At the time of outcome assessment? 

• Incomplete outcome data: Were incomplete outcome data adequately addressed? 

• Selective outcome reporting: Were reports of the study free of the suggestion of selective outcome reporting? 

• Other sources of bias: Was the study apparently free of other problems that could put it at high risk of bias?

Measures of treatment effect

We followed the procedures of the Cochrane Neonatal Review Group. For categorical data, when applicable, we used risk ratios (RRs), risk differences (RDs), and numbers needed to treat for an additional beneficial outcome (NNTBs) along with their respective 95% confidence intervals (CIs). For continuous data, such as length of hospital stay, we used mean differences (MDs) with 95% CIs.

Unit of analysis issues

We identified the unit of analysis for each individual trial to determine whether appropriate adjustments had been made to account for non‐independence between twin pairs via use of appropriate analysis methods, such as generalized estimating equation (GEE) modeling. Had an NICU‐cluster trial been included, we would have looked for evidence of adjustments at the level of the NICU, in addition to adjustments at the level of each twin pair. If the unit of analysis was not stated in the study, we planned to inspect the width of the standard error (SE) or the 95% CI of estimated treatment effects. If we found an inappropriately small SE or a narrow 95% CI, we would have asked the authors of the study to provide information on the unit of analysis.

In this review, we considered each twin pair as representing one unit when we analyzed data from trials in which randomization occurred at the level of each twin pair. For continuous outcomes, we reported the average figure for each twin pair, for example, the average rate of weight gain of the twin pairs. For dichotomous outcomes with low event rates, we counted the event as positive if either of the twins developed the outcome (e.g. in this review, we counted the outcome of sepsis as positive if one or both twins developed sepsis). For dichotomous outcomes with high event rates, had data been available, we would have assigned a score of zero, one, or two, which would correspond to the number of twins in a twin pair with the outcome of interest.

When we found a unit of analysis error, for example, when the included study analyzed outcome data for individual infants instead of for each twin pair, we re‐analysed data derived from the studies and adjusted for the effect of clustering by using different strategies for two main types of studies, as follows.

• For studies in which randomization occurred at the level of each twin pair, as in the included studies, we adjusted individual data, counting each twin dyad as an individual. For continuous data, we recorded the average reading of the twin pair. For dichotomous data, we recorded the number of twin pairs with the outcome (positive if any or both of the twins developed the outcome).

• Had we included studies in which randomization occurred at the level of the neonatal unit (i.e. NICU‐cluster trials), we would have dealt with two levels of clustering: (a) clustering at the level of the twin pair, and (b) clustering at the level of the neonatal unit.

• • To address (a), we would have adjusted the individual data, counting each twin dyad as an individual. For continuous data, we would have recorded the average reading of the twin pair. For dichotomous data, we would have recorded the number of twin pairs with the outcome (positive if any of the twins developed the outcome).

  • To address (b), we would have adjusted the final estimates of each study obtained by analyzing each twin dyad, as stated. We would have done so by multiplying the SEs of the final effect estimates by the square root of the "design effect": (1 + (M ‐ 1) × ICC), where M was the average cluster size (number of twin pairs per cluster) and ICC was the intra‐cluster correlation among twin pairs within each NICU. We would have determined the average cluster size (M) from each trial by dividing the total number of twin pairs by the total number of clusters. We would have sought the best estimate of the ICC from reliable resources such as landmark studies involving twins and NICU clustering.

  • We would have combined adjusted final effect estimates from each trial with their SEs in the meta‐analysis using generic inverse variance methods.

Should we fail to identify a reliable ICC for NICU‐cluster trials for twins after searching through relevant resources, we planned to use partially adjusted estimates (adjusted only at the level of each twin dyad) for our meta‐analysis. We planned to include studies for which determination of the unit of analysis was not possible in the overall meta‐analysis on the basis of effect estimates given by study authors. We would have then performed a sensitivity analysis to assess how the overall results were affected with and without inclusion of these studies.

Dealing with missing data

We obtained a drop‐out rate from each included study. A drop‐out rate equal to or greater than the control group event rate would be considered significant. If a significant drop‐out rate had been found, we would have contacted the authors of the individual studies to ask that they provide additional data when possible.

Had a sufficient quantity of data been available, we would have performed a sensitivity analysis to assess how the overall results were affected with and without inclusion of studies with significant drop‐out rates.

Assessment of heterogeneity

Had we included more than one study that reported the same outcome measure, we would have assessed the treatment effects of individual trials and heterogeneity between trial results by first inspecting the forest plots.

We then would have evaluated the I² statistics to look for an indication of the degree of heterogeneity in study results, with a value greater than 50% indicating moderate or high heterogeneity. If moderate or high heterogeneity was found, we would have explored the causes (e.g. differences in study quality, participants, interventions, or outcome assessments) via post hoc subgroup analyses. We then would have made a decision about whether a meta‐analysis would be appropriate and, if so, whether we should exclude some studies.

Assessment of reporting biases

We constructed a matrix consisting of all outcomes assessed in each included study, as stated in the methods and reported in the results. We compared reported outcomes against outcomes listed in the methods section. In addition, when we found in a study some key outcomes that we considered assessable but were not included, we contacted the study authors to ask for clarification. We sought the study protocol, when possible, from PubMed, a relevant trial registry, or a Web link provided by the study, or we directly contacted study authors. We performed cross‐examination against the conference presentation of the same work, when possible. We planned that if all of the above measures failed to provide further useful information, we would perform a sensitivity analysis by taking an outcome that was reported by all studies and comparing overall results with and without inclusion of studies in which key outcomes were missing.

We planned to screen for publication bias by using a funnel plot if we included a sufficient number of studies (≥ 10) in the analysis. If we suspected publication bias (i.e. significant asymmetry upon visual inspection of the funnel plot), we would have included a statement in our results and a corresponding note of caution in our discussion.

Data synthesis

Had data from more than one study been available for each outcome, we would have used a fixed‐effect model to combine the final effect estimates from each trial with their SEs in the meta‐analysis using generic inverse variance methods in RevMan.

Subgroup analysis and investigation of heterogeneity

Had relevant data been available, we would have performed subgroup analyses for the following.

• Twins with different weight range: VLBW (birth weight of the smaller twin < 1500 g) vs non‐VLBW.

• Twins with and without significant weight discordance at birth.

• Preterm twins (< 34 completed weeks of PMA at commencement of co‐bedding) and twins of borderline maturity (35 and 36 weeks' completed PMA).

• Twins co‐bedded in an incubator and a cot at study entry.

• Studies in which randomization occurred at the level of each twin pair and each neonatal unit.

Sensitivity analysis

We would have performed sensitivity analyses based on the following, if applicable.

• Adequacy of sequence generation and allocation concealment, as detailed under the heading Assessment of risk of bias in included studies.

• Adequacy in handling missing data, as detailed under the heading Dealing with missing data.

• Availability and adequacy of adjustments of units of analysis, as detailed under the heading Unit of analysis issues.