Types of studies

Eligible trials were randomised controlled trials (RCTs), including cluster‐randomised controlled trials (C‐RCTs) and cross‐over trials, that:

1. Compared two or more alternative intervention programmes to increase the consumption of fruit or vegetables or both of children aged five years and under;

2. Compared an intervention programme to increase the consumption of fruit or vegetables or both of children aged five years and under with a standard‐care or no‐intervention control group.

We excluded trials which did not include fruit or vegetable intake as a primary trial outcome, to avoid the potential confounding effects of other interventions, and because publication bias and selective outcome reporting are more predominant among secondary trial outcomes (or outcomes that were not otherwise stated). We included trials that did not state a primary trial outcome but did assess an eligible fruit or vegetable intake outcome. We included eligible cross‐over trials in the review, as we deemed them a suitable and common method for assessing the effect of interventions to increase the fruit and vegetable consumption of children.

Types of participants

Participants could include:

1. Children aged five years and under. Trials including children older than five years were included only if the mean age of the study sample at baseline was five years or less;

2. Parents, guardians and families responsible for the care of children aged five years and under;

3. Professionals responsible for the care of children aged five years and under, including childcare staff and health professionals.

Types of interventions

We considered any educational, experiential, health promotion and/or psychological or family or behavioural therapy or counselling or management or structural or policy or legislative reform interventions, designed to increase consumption of fruit or vegetables or both in children aged five years and under (as defined in types of participants). Interventions could be conducted in any setting including the home, childcare/preschool services, health services, or community settings. 

Comparison: Any alternative intervention to encourage fruit and vegetable consumption as described above, or a no‐intervention control, usual care, or attention control or wait‐list control. Attention controls in randomised trials for behavioural interventions are those that include clinical attention and induce the expectation of therapeutic benefit for control for non‐specific effects of the intervention (Freedland 2011). Wait‐list control groups that are also designed to control for non‐specific effects involve participants being allocated to receive an intervention at study conclusion (delayed start) (Whitehead 2004).

Types of outcome measures

We include studies with evaluated outcomes measuring biomedical or dietary indices or both.

Electronic searches

We conducted a search of the following electronic databases:

• Cochrane Central Register of Controlled Trials (CENTRAL) Issue 8, 2016 in the Cochrane Library (searched 30 September 2016);

• Epub Ahead of Print, In‐Process & Other Non‐Indexed Citations, MEDLINE Daily and MEDLINE (Ovid, 1946 to 30 September 2016) (searched 30 September 2016);

• Embase Classic and Embase (Ovid, 1947 to 2016 week 39) (searched 30 September 2016);

• CINAHL (EBSCO, 1937 to 5 July 2016) (searched 5 July 2016);

• PsycINFO (Ovid,1806 to June week 5 2016) (searched 5 July 2016). 

The search strategies are described in Appendix 2. We applied the sensitivity‐maximising version of the Cochrane RCT filter (Lefebvre 2011) to MEDLINE, and adaptations of it to the other databases except for CENTRAL. We imposed no restrictions by date or language of publication.

As a living systematic review we will run the following electronic database searches monthly:

• CENTRAL

• Epub Ahead of Print, In‐Process & Other Non‐Indexed Citations, MEDLINE Daily and MEDLINE

• Embase

We will set up auto‐alerts (where possible) to deliver a monthly search yield by email.

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 searched the reference lists of included articles and handsearched all articles published between 2006 and September 2016 in three relevant international peer‐reviewed journals (Journal of Nutrition Education and Behavior, Public Health Nutrition, and Journal of the Academy of Nutrition and Dietetics (previously titled Journal of the American Dietetic Association)). 

We searched the metaRegister of clinical trials (www.controlled‐trials.com/mrct/) and the WHO International Clinical Trials Registry Platform (www.who.int/ictrp/) in November 2016, and www.clinicaltrials.gov on 14 June 2017. We searched a database of published dissertations, Proquest Dissertations and Theses, in November 2016, to identify eligible studies. We conducted a 'grey literature' search in GoogleScholar on 14 June 2017, of which we screened the first 200 citations for eligibility. We contacted the authors of included studies to try to obtain other eligible trials published in peer‐reviewed journals, as well as ongoing trials. We describe ongoing studies, where available, detailing the primary author, research question(s), methods and outcome measures (Characteristics of ongoing studies).

We are now running monthly trial registry searches of WHO International Clinical Trials Registry Platform and clinicaltrials.gov. We will set up auto‐alerts (where possible) to deliver a monthly search yield by email.

We will conduct selected searches of other resources (articles published in three relevant international peer‐reviewed journals: Journal of Nutrition Education and Behavior, Public Health Nutrition, and Journal of the Academy of Nutrition and Dietetics; database of published dissertations; and 'grey literature' in GoogleScholar) manually every six months.

As additional steps to inform the living systematic review, we will contact corresponding authors of ongoing studies as they are identified and ask them to advise when results are available, or to share early or unpublished data. We will contact the corresponding authors of any newly‐included studies for advice as to other relevant studies. We will conduct citation tracking of included studies in Web of Science Core Collection on an ongoing basis. For that purpose, we have set up citation alerts in Web of Science Core Collection. We will manually screen the reference lists of any newly‐included studies and systematic reviews. Also, we will use the 'Related citation' feature in PubMed to identify additional articles.

We will review search methods and strategies approximately yearly, to ensure they reflect any terminology changes in the topic area, or in the databases.

Selection of studies

Pairs of review authors (from RH, RW, FS, SY, NN) independently screened titles and abstracts of identified papers. Review authors were not blinded to the details of the study author or journal. Review authors applied a standardised screening tool to assess eligibility. Articles were screened against the eligibility criteria of participants (mean age of children more than five years), outcome (primary outcome was not fruit and vegetable intake), comparator (was not a no‐intervention, usual care, attention or wait‐list control), intervention (did not aim to increase child fruit or vegetable intake) and study type (was not RCT, C‐RCT or cross‐over trial with random allocation to group). Based on the title and abstract, we excluded papers which clearly did not meet the eligibility criteria of the review. Pairs of review authors (from FS, RH, NN, RS, SY) then independently examined the full text of all remaining articles. We documented Information regarding the reason for the ineligibility of any paper for which we reviewed the full text, and present it in the table 'Characteristics of excluded studies'. A third review author with expertise in review methodology (LW) resolved any disagreements between review authors on study eligibility. For those papers which did not provide sufficient information to determine eligibility, we contacted the study authors for clarification.

We will immediately screen any new citations retrieved by the monthly searches. As the first step of monthly screening, we will apply the machine learning classifier (RCT model) (Wallace 2017), available in the Cochrane Register of Studies (CRS‐Web) (Cochrane 2017a). The classifier assigns a probability (from 0 to 100) to each citation of being a true RCT. For citations that are assigned a probability score of less than 10, the machine learning classifier currently has a specificity/recall of 99.987% (Wallace 2017). We will screen in duplicate and independently all citations assigned a score from 10 to 100 . Citations that score 9 or less will be screened by Cochrane Crowd (Cochrane 2017b). Any citations that are deemed to be potential RCTs by Cochrane Crowd will be returned to the authors for screening.

Data extraction and management

Pairs of review authors (from EJ, RW, RH, KB, KO, ER, TCM, RS) independently extracted data from each included trial. Review authors were not blinded to the details of the study author or journal. We recorded data on data extraction forms designed and piloted specifically for this review. Consultation with a third review author with expertise in review methodology (LW) resolved discrepancies between review authors about data extraction. We tried to contact authors of included papers in instances where the information required for data extraction was not available from the published report, or was unclear. One review author entered extracted data into the systematic review software Review Manager 5 (RevMan) (RH) and another review author checked it (KO). Where available, we extracted the following information from included trials: 

1. Information on the study, research design and methods, such as the study authors; date of publication; date of study initiation; study duration; setting; number of participants; participants' age, gender, ethnicity, and socioeconomic position. 

2. Information on the experimental conditions of the trial, such as the number of experimental conditions; intervention and comparator components; duration; number of contacts; modalities; interventionist; and integrity.

3. Information on the trial outcomes and results, such as rates of recruitment and attrition; sample size; number of participants per experimental condition; mean and standard deviation of the primary or secondary outcomes described above; any subgroup analyses by gender, population group or intervention characteristics; and analyses (including whether studies appropriately adjusted for clustering).

Assessment of risk of bias in included studies

Two review authors (FS and FT) independently assessed the risks of bias in the included studies. We consulted a third review author (RH) with expertise in review methodology to resolve any disagreements between review authors. Review authors used the tool outlined in the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011) to assess the risks of bias. The tool requires an explicit judgement by the review authors, based on trial information, about the risks of bias attributable to the generation of the random sequence, the allocation concealment, the blinding of participants, personnel and outcome assessors, the completeness of outcome data, selective reporting, and any other potential threats to validity. We also judged recruitment bias, baseline imbalance, loss of clusters and incorrect analysis for C‐RCTs. Judgements on the risks of bias for each trial are recorded in the ‘Risk of bias’ tables accompanying the review.

Measures of treatment effect

Where meta‐analyses were performed, we expressed the intervention effect as a mean difference (MD) where outcomes were reported using a standard metric (such as grams), and as a standardised mean difference (SMD) where outcomes were reported using different methods or metrics of fruit and vegetable intake (such as grams, grams per kilogram of body weight, and serves per day).

Unit of analysis issues

We assessed cluster‐randomised trials in the review for unit‐of‐analysis errors. Where cluster‐randomised studies did not account for clustering, we contacted study authors to provide intra‐class correlation coefficients (ICCs) to allow calculation of design effects and effective sample sizes to enable individual‐level pooling. Where ICCs were not available, we estimated a mean ICC from reported ICCs of included studies, and used it to calculate effective sample sizes.

Dealing with missing data

Where available, we reported outcomes of trials using an intention‐to‐treat analysis. If studies did not report intention‐to‐treat analyses, we reported as‐treated analysis of trial outcomes. We explored the impact of including as‐treated trial outcomes in meta‐analysis for studies with a high rate of attrition (more than 20% for short‐term outcomes) in sensitivity analyses (see below Sensitivity analysis). We contacted study authors to obtain any missing data (e.g. standard deviations).

Assessment of heterogeneity

We assessed statistical heterogeneity by visual inspection of forest plots of the included trials, and calculation of the I² statistic where data from included trials could be pooled. Subgroup analyses by trial characteristics (e.g. type of participants, intervention or outcomes) to identify the source of substantial heterogeneity (defined as I² greater than 50%) was not able to be conducted due to the similarity in trial characteristics.

Assessment of reporting biases

We checked for reporting bias by visual inspection of the funnel plots.

Data synthesis

We assessed trial outcomes using a variety of dietary assessment tools and reported in various metrics, including vitamin C from fruit, fruit or vegetable serves, and grams of fruit and/or vegetable consumption. We calculated standardised mean differences (SMDs; to account for variable outcome measures) for each comparison, using the generic inverse variance method in a fixed‐effect meta‐analysis model (where there was no or low statistical heterogeneity in the primary analysis) or a random‐effects meta‐analysis model (where there was unexplained heterogeneity in the primary analysis), using the RevMan software. We selected post‐intervention values over change‐from‐baseline data for inclusion in meta‐analysis, to reduce the risk of selective reporting and to maximise the number of studies that could be pooled.

We synthesised studies that provided data suitable for pooling in meta‐analyses grouped by intervention type (infant feeding, parent nutrition education, and multicomponent interventions). When studies reported multiple fruit or vegetable outcomes, we selected the stated primary trial outcome for inclusion in our meta‐analyses, or if a primary outcome was not stated we selected the first reported outcome for inclusion. For studies which reported multiple follow‐up points, we extracted data from the longest follow‐up period for inclusion in meta‐analyses.

We selected reported study estimates that adjusted for potential confounding variables for inclusion in meta‐analysis over reported estimates that did not adjust for potential confounding variables. Similarly, for C‐RCTs that reported study estimates that were unadjusted and adjusted for clustering, we preferred estimates that adjusted for clustering for inclusion in meta‐analyses. For C‐RCTs that did not report post‐intervention study estimates (and a relevant measure of variance) that accounted for clustering, we calculated a design effect and effective sample size using study data (number of clusters, number of participants analysed) and a reported ICC from one of the included studies (vegetable intake: ICC 0.014, fruit intake: ICC 0.016; De Bock 2012). For such C‐RCTs (De Coen 2012; Martinez‐Andrade 2014; Namenek Brouwer 2013; Nicklas 2017; O'Connell 2012; Roset‐Salla 2016; Verbestel 2014; Williams 2014), we entered the reported post‐intervention outcome data (e.g. mean and standard deviation) and author‐calculated effective sample sizes into Revman to calculate individual‐level adjusted study estimates to enable inclusion in meta‐analyses. We tried to pool studies separately that compared two or more alternative interventions.

For cross‐over trials, we tried to synthesise results separately from parallel RCTs, by pooling results from paired analyses that adjust for within‐individual comparisons. If such data were not available, we combined results by pooling data from the first cross‐over period (i.e. essentially a parallel RCT) with parallel RCTs.

In all instances where we could not combine data in a meta‐analysis, we have provided a narrative summary of the trial findings according to the review objectives.

Whenever we find new evidence (i.e. studies, data or information) meeting the review inclusion criteria, we will extract the data, assess risks of bias and incorporate it into the synthesis every three months, as appropriate.

We will incorporate any new study data into existing meta‐analyses using the standard approaches outlined in the Data synthesis section.

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 (in press)).

Subgroup analysis and investigation of heterogeneity

Where possible, we conducted subgroup analyses of interventions for the following subgroups, which we had planned a priori:

1. Interventions targeting boys and girls (not conducted);

2. Interventions targeting minority groups including indigenous populations (not conducted, described narratively);

3. Interventions delivered in various settings including health and children’s services (conducted where possible for some comparisons and settings);

4. Interventions of varying intensities, defined in terms of the number and duration of intervention contacts or components (not conducted);

5. Interventions delivered in different modes, such as by telephone, the Internet or face‐to‐face (conducted for some comparisons and modalities, otherwise described narratively).

Sensitivity analysis

Where possible, we conducted sensitivity analyses to explore the impact on the overall assessment of treatment effects:

1. Excluding studies at high risk of bias (defined a priori);

2. Excluding studies not reporting an intention‐to‐treat analysis, with high rates of participant attrition defined as greater than 20% (defined a priori);

3. Excluding studies that did not have a primary outcome of child fruit and vegetable consumption (post hoc).

For the sensitivity analysis excluding studies that did not have a primary outcome of child fruit and vegetable consumption, we considered studies to have a primary outcome of children's fruit and vegetable intake even when this was not explicitly stated if: children's fruit and vegetable intake was the only reported outcome, a sample size calculation for children's fruit and vegetable intake was reported, or children's fruit and vegetable intake was the first reported outcome.