Study eligibility criteria

To select the relevant works, we have set a number of procedural criteria of the studies’ design and method, type of analysed behaviour, and participants:

1. The study should conduct an analysis of longitudinal data using SAOM. The number of data collection waves is not a limitation provided that there are at least two waves.

2. The study should explicitly measure social ties between adolescents (friendship ties, romantic ties, support ties and so on).

3. The behaviour modelled in SAOMs should be alcohol consumption. Specifically, the outcomes may be any of the following: drinking quantity, drinking frequency, binge drinking, drunkenness, onset to first alcohol use, lifetime alcohol use.

4. The target population should be children and adolescents from 10 to 19 years old.

5. The study manuscript should be available by September 2019 when the updated database searches will be conducted.

Information sources and search strategy

The current research collects and analyses studies published in academic journals, dissertations/theses, reports, conference materials. The articles will be searched in the following electronic databases: Web of Science, Scopus, PubMed, The Cochrane Library (Cochrane Database of Systematic Reviews, Cochrane Central Register of Controlled Trials (CENTRAL), EBSCOhost (MEDLINE, SocINDEX, Academic Source, ERIC), ProQuest (ProQuest Dissertations and Theses Global), PsycINFO (PsycNET), Excerpta Medica database (Embase) and Cumulative Index to Nursing and Allied Health Literature (CINAHL). The search will include titles and abstracts. We will use Google Scholar for control, but not as the main electronic database; rather, we will use it to check for unpublished studies and grey literature. This procedure will allow us to correct our search query if needed.

Search query will consist of three points that match the research question: drinking behaviour, SAOMs and adolescents, and will be formulated with the inclusion of the maximum number of synonyms corresponding to these three points. Previously conducted systematic reviews on similar topics will help to formulate accurate queries.

Point 1 ‘Drinking behaviour’: drink* OR alcohol* OR booze OR tipple and other synonyms, as well as types of beverages—liquor OR wine OR beer, etc. We will also include ‘Substance Abuse’ to collect a wider selection of studies dedicated to drinking among adolescents.

Point 2 ‘Analytical approach’: SAOM* OR ‘stochastic actor-oriented’ OR The algorithm Simulation Investigation for Empirical Network Analysis in R (RSIENA) OR ‘stochastic actor-based’ OR ‘simulation investigation for empirical network’ and other possible variations of the method name.

Point 3 ‘Adolescents’: adolesc* OR youth OR ‘young person’ OR teen* OR student* OR ‘peer group’ OR child* OR pupil and other synonyms.

After the search query has been tested on Web of Science and Scopus, it will be adapted for other search engines. If the request is modified, another search will be conducted through all the electronic databases. All of the changes will be discussed by the research team.

After forming our database of articles, according to the inclusion criteria, the references will be cross-checked for possible relevant studies not found by the above search queries.

An example of a full electronic search strategy for the Web of Science is given in online supplementary file 2.

Study selection process

The selection process will be conducted by three members of the research team independent of each other.

First, after deleting duplicate articles from the lists retrieved from all electronic databases, three reviewers will make independent decisions on whether the article should be included based on its brief description. Each article will be classified as ‘relevant’, ‘irrelevant’ or ‘unclear’. Articles classified as ‘unclear’ by at least one team member will be discussed by the research group. Articles classified as ‘irrelevant’ by two team members will be deleted.

In the second step, three reviewers will carefully examine the full texts of the remaining relevant articles, make independent decisions on the methodological quality of the studies and conclude whether the studies should be included in the meta-analysis. Disagreements about disputed cases will be resolved through general discussion.

Data collection process

The data collection form will be created and piloted through several iterations in the course of reviewing a sample of primary studies. When the data collection form is finalised, two reviewers will independently extract data from each eligible study, duplicating each other’s work. Duplication is needed to reduce errors in data extraction. The third reviewer will compare the duplicate extractions, and all discrepancies will be resolved through general discussion. If multiple reports of the same study will be identified, only one extraction will be kept. Such identification can be done by comparing the records of these studies and finding identical project names, datasets, characteristics of the study design, samples or model specifications. In case of missing or incomplete information (ie, participants demographics or missing data treatment), authors of primary studies will be contacted (up to three email attempts per study).

Data items

The following data will be extracted from the selected articles:

1. Bibliometrics: authors, name, source, year of publication, language and country. Bibliometric data will be generated automatically and checked for consistency by the members of the research team.

2. Characteristics of the study: country(s) and year when the study was conducted, project, dataset, type of sample (general, high risk, other), sample size (number of participants and number of networks), gender composition of the sample, age of participants and socioeconomic status of participants.

3. Study design: number and frequency of survey waves, data collection techniques, name generator questions.

4. Description of ‘alcohol drinking behaviour’: drinking quantity, drinking frequency, binge drinking, drunkenness, onset to first alcohol use, lifetime alcohol use; scale assessing the frequency/intensity of alcohol consumption, the type of alcoholic beverage, etc.

5. Descriptions of covariates included in the model: gender, age, family socioeconomic status and so on.

6. Description of participant dropout from wave to wave, description of missing data, treatment of missing data.

7. Model specifications.

8. Effect size data: parameter coefficients and SEs for selection effects, influence effects, behaviour dynamics effects, network structural effects.

Primary outcome

Estimation of the effect size of peer influence on alcohol consumption behaviour among adolescents of different age groups.

Quality assessment and risk of bias

Careful reviewing of available risk of bias assessment tools and quality assessment checklists revealed that none of them are applicable to non-experimental, longitudinal panel network studies.35 36 We do not expect high heterogeneity between studies as every eligible study should employ the same methodology; nevertheless, the important characteristics of study design and implementation may differ considerably. We devised a custom checklist of study characteristics to evaluate the quality of each study: the number of networks, the number of participants, response rate, attrition rate, the percentage of missing data, treatment of missing data, study limitations. The checklist will be pilot-tested with a few articles and modified before quality evaluation. When the checklist is finalised, two reviewers will independently evaluate each eligible study. If there is insufficient detail reported in the study, we will contact the primary study authors for more information (up to three email attempts per study). Disagreements will be resolved by general discussion.

While the sample size is important for statistical power, it is not a source of bias. Missing data, on the other hand, if not treated properly, would have a significant impact on research findings. One of the problems that emerge while working with longitudinal network data is the loss of a network agent due to attrition. Other reasons for missing data are item non-response and tie non-response. However, up to 10% of missing data does not lead to biased estimates as RSIENA package includes a data recovery procedure that replaces missing data using information from the previous or next wave or the global average.27 Though imputation reduces the risk of bias in evaluating effects, we will closely monitor the percentage of missing data reported in primary studies.

Data synthesis

If a sufficient number of studies using SAOM are found, we plan to extract model results from individual articles and to conduct a meta-analysis. A meta-analysis of SAOM results is straightforward because parameters coefficients are presented as log odds ratios and require no further conversion. However, in SAOM there are different effects for testing peer influence (average similarity, total similarity, average alter) that are not directly comparable. Hence, we will need reasonably similar models, which further narrows our base for meta-analysis.

Since for some age groups there might not be enough research available, we will need to merge some age categories of adolescents. We may need to consult child psychologists so as to make an informed decision on merging some of the age categories.

For assessing heterogeneity of individual studies, we will use tau squared calculated in metafor package.37 38 For meta-regression, we will use either robust variance estimation in robumeta package39 or multilevel model in metafor package.29 38

Post-hoc sensitivity analyses will be conducted to explore whether covariates (age, gender composition) or choice of outcome measure (specific drinking behaviour) could explain observed heterogeneity in results.

Patient and public involvement

Patients and/or the public are not involved in the current study.

Strengths and limitations

The fundamental strength of the proposed systematic review/meta-analysis is the fact that it is based on articles using rigorous methodology allowing the separation of peer influence from peer selection in the range of adolescent alcohol consumption behaviour. ‘Social contagion’ with harmful habits has been a popular research topic, and many researchers have attempted to evaluate peer influence in health risk behaviour. Recently, several review articles summarise the findings in this area.24–26 However, the papers reviewed have used a variety of methods, none of which allows separation of the effects of peer influence from the confounding effects of peer selection.

Traditional tests of peer influence suffer from ‘projection bias’, that is, the tendency for people to overestimate the similarity between their behaviours and those of their peers and from the inherent entanglement of two processes: social influence and social selection. The only method (SAOM) able to isolate selection and influence was designed explicitly for this purpose less than 20 years ago.28

The SAOM method requires network data and longitudinal surveys, which are quite difficult to implement. Nevertheless, more and more papers employing this methodology are appearing. Papers using longitudinal network data and SAOM analysis will be collected and reviewed.

A possible limitation is that not enough articles for meta-analysis will be found. Our inclusion criteria imply specific age of adolescents (10–19 years old). A second limitation is that the models for meta-analysis should be sufficiently similar. SAOM allows modelling of different peer effects (average similarity, total similarity and average alter) that may not yield similar interpretations. Moreover, structural effects included in the model may also differ; therefore, we will extract data for meta-analysis from studies that use comparable model specifications.