We will undertake a systematic review of existing observational evidence and will perform meta-analyses where sufficient data are available. The study has been pre-registered with AsPredicted (No. 39505) and registered to the International Prospective Register of Systematic Reviews (PROSPERO CRD42020165345). This protocol follows the Preferred Reporting Items for Systematic Review and Meta-Analyses for Protocols (PRISMA-P) 2015 guidelines (see Additional file 1) [43].
Eligibility criteria
The inclusion/exclusion criteria have been defined according to the Population of interest, Exposure, Comparator, and Outcome (PECO) statements as described below.
Our population of interest will be children and youth aged less than 25 years old, with no geographical limitation. We will include studies carried out in clinical and population settings among individuals aged less than 25 years or whose mean age is under 25. We will exclude studies that do not specify participants’ mean age or age range. No lower age limit will be set in order to inform the scientific community about the youngest ages considered in existing studies and about potential research gaps and limits in very young children.
The term “parent” will refer to the biological or adoptive parent(s), guardian(s), or caregiver(s).
In accordance with the rationale described earlier and with previous meta-analyses examining the relationship of parenting with child outcomes, parenting factors will be classified according to the three broad categories of positive parenting, negative parenting, and parent-child relationship [44,45,46].
Studies will be eligible for inclusion if they assess parenting before the age of 18 or at a mean age lower than 18. We set this age limit knowing that, in most countries and states, reaching 18 years corresponds to legal emancipation of children and is marked by more autonomy, life decisions, and often changes in living arrangement that have a main impact on how parents and children perceive the role of parenting and the parent-child relationship [47].
We will focus on the following two outcomes as defined by the CDC: (i) suicidal ideation, which refers to thinking about, considering, or planning suicide, and (ii) suicide attempt, which refers to a non-fatal, self-directed, potentially injurious behavior with intent to die as a result of the behavior [5].
We will not consider data related to non-suicidal self-injury (NSSI) and non-suicidal self-harm in the present study. Although they are highly comorbid with suicidal behaviors in children and adolescents, these are phenomenologically different [48] and could be influenced by distinct protective and risk factors [49].
Some studies have examined the relationship of parenting with self-injury and self-harm, which refer to any self-injurious behavior, including suicidal (suicide attempt) but also non-suicidal self-injurious behavior (NSSI and non-suicidal self-harm).
In order to identify all relevant data pertaining to suicide attempts, our search strategy is meant to capture these studies by including relevant keywords such as “self-injury,” “self-harm,” or “self-mutilation.” However, we will make a distinction between suicidal and non-suicidal self-injurious behavior based on the presence of an intent to die as a result of the behavior, in accordance with the CDC’s definition of suicide attempt given above and with the standardized nomenclature established based on the Columbia Classification Algorithm of Suicide Assessment [9, 50]. During the article selection process, reviewers will carefully assess the definition of each outcome considered (including in sub-analyses) and will only include studies reporting data on suicide attempts, defined as committed with at least some intent to die as a result of the act. Studies that examine self-harm as a single entity without differentiating suicidal from non-suicidal self-injurious behavior, as well as those in which the intent to die is not ascertained, will be excluded, because we assume that they do include non-suicidal behaviors.
Only observational studies with retrospective, cross-sectional, or longitudinal designs will be eligible for inclusion. We chose to exclude case-control studies for two main reasons. First, they are more prone to selection bias when control subjects are not selected from the same population as the cases [51]. Second, in suicide research, the risk of recall bias might be higher in case-control than in cross-sectional studies. Indeed, in case-control studies, NFSB participants are typically recruited and information collected in the days following suicidal behavior, when parents and their offspring often try to make sense of it [52]. Therefore, they might recall parenting factors in more detail and overreport them compared to controls, which might artificially strengthen the observed associations with NFSB.
To examine the bidirectional association of parenting and NFSB, we will include longitudinal studies that examine either the effect of parenting on subsequent suicidal behavior or the effect of suicidal behavior on subsequent parenting factors. The findings of case reports, case series, therapy/treatment-based intervention studies, discussion articles, exclusively qualitative studies, reviews, or meta-analyses will be excluded. However, the reference lists of literature reviews and meta-analyses will be reviewed to capture possible additional relevant citations.
Studies published (or “in-process”) in a peer-reviewed journal as well as dissertations will be included. The inclusion of dissertations will allow us to consider results published outside of traditional commercial publishing and thus reduce the risk of publication bias [53]. However, we will not include conference posters and presentations for two reasons. First, they may not contain adequate information about the study design, methods, biases, and results, limiting critical appraisal of corresponding studies. Second, the association of parenting and NFSB has already been examined in a large number of studies published as articles and dissertations, and in that case, the inclusion of conference abstracts in meta-analyses has been shown to result in only small differences in the effect estimates [54].
Our research team includes members who are proficient in English and in French, making us able to review research works published in these two languages.
Search strategy
A primary search strategy was developed in APA PsycInfo by a health sciences librarian (TR), and after review and validation by co-authors, the final search strategy was run in APA PsycInfo on November 6, 2019 (Additional file 2). On the same day, it was translated and applied in MEDLINE, CINAHL, Embase, Scopus, and the Cochrane Library databases and was also run in MEDLINE Epub Ahead of Print and In-Process & Other Non-Indexed Citations, to capture the most recent literature. Database-specific subject headings and keywords in natural language were used to capture “parenting dimensions” and “suicidality” concepts, and combined using Boolean logic and operators including proximity searching. These results were then limited to articles where “child” and “adolescent” terms and their synonyms appear in selected fields, and to observational study types. No year limits or language limits were applied.
Data screening
Two principal independent reviewers (FP and XJ) will follow a two-step selection process using the Covidence® software, according to the eligibility criteria described previously. The first decision will be made based on the titles and abstracts. Then, the selected articles will be considered for full-text assessment to determine if they definitely qualify for inclusion. Any disagreement will be discussed by the two reviewers, and any remaining discrepancies will be resolved by a third reviewer (MA).
Data extraction
Data will be extracted separately by the two principal reviewers using a standardized data extraction form and a coding process implemented in the Research Electronic Data Capture System (REDCap®). The following information will be systematically extracted from the included studies:
-
General study characteristics: first author, year of publication, journal, and type of publication (peer-reviewed article or dissertation).
-
Study setting: country where the study was performed and setting in which it took place (mental health care setting, other clinical care settings, or population-based).
-
Study design: type of study (e.g., cross-sectional, longitudinal) and time period for data collection.
-
Sample characteristics: sample size, age of participants (range and/or mean ± standard deviation) or corresponding school grades, gender distribution, and main ethnicity (defined as the ethnicity shared by more than 60% of participants, otherwise ethnicity will be defined as “balanced”),
-
Measurement of parenting: type of parenting, measurement time frame, type of informant (child, parent, other), relationship of caregiver with the child (biological parents only or not), and method for assessment (questionnaire, interview, or observation).
-
Assessment of NFSB: type of outcome (suicidal ideation or attempts), assessment time frame, informant (child, parent, other), and method for assessment (questionnaire, interview, observation).
-
Effect estimates: non-adjusted and/or adjusted effect estimates (along with their standard deviation or 95% confidence intervals) relating to the association of each parenting factor with one or both of our outcomes will be extracted and converted to odds ratios (OR) for dichotomous outcomes and standardized mean differences (Cohen’s d) for continuous outcomes, using conventional conversions.
Any disagreements between the two extraction processes will be resolved by consensus discussion with the third reviewer. In case of unclear or incomplete data, original authors will be contacted.
Risk of bias
The two principal reviewers will independently assess the methodological quality of studies using the Quality Assessment Tool for Observational Cohort and Cross-Sectional Studies developed by the US National Heart, Lung and Blood Institute (NHLBI) [55]. This validated tool includes 14 items for evaluating potential bias induced by study methods or implementation, including patient selection, attrition, confounding, sample size justification, and arguments for causation. Reviewers will select “yes,” “no,” or “cannot determine” in response to each item. Some questions of the tool have been slightly adapted to better capture the strengths and weaknesses of existing studies in the scope of our topic. Reviewers will also rate the overall study quality as “good,” “fair,” or “poor” based on their rating for each item and their own critical appraisal of the risk of bias, as recommended by the guidance document developed by the NHLBI methodology team. If reviewers rate the overall quality of the study as poor, they will state the reasons for the decision. In case of disagreements, consensus will be sought through discussion between raters and, if necessary, with the third reviewer.
Data synthesis
Evidence regarding the association of parenting factors with each of the two outcomes (suicidal ideation and suicide attempt) will be reported according to the Preferred Reporting Items for Systematic Review and Meta-Analysis (PRISMA) criteria [43] and satisfy the Meta-analysis Of Observational Studies in Epidemiology (MOOSE) Checklist for Meta-analyses of Observational Studies [56].
Meta-analyses will be performed using a random effect model when a minimum of three studies with usable data are available. We will calculate the effect sizes as odds ratios (OR) or standardized mean difference (Cohen’s d), with standard errors, and convert information reported in a different metric using conventional conversions.
A narrative summary of the evidence will be provided by outcome, including results from studies that would not be possible to consider in meta-analysis. The results will be presented using forest plots and in a “summary of findings table.” The Grading of Recommendations Assessment, Development, and Evaluation (GRADE) approach will be used to rate the certainty of the evidence [57]. Risk of bias (assessed as previously described), inconsistency of results, indirectness of evidence, imprecision of effect estimates, and risk of publication bias will be considered as reasons to rate down the quality of evidence, whereas a large magnitude of effect and the presence of a dose-response gradient will be considered to rate it up.
Studies often report on multiple effect sizes obtained from the same sample or in the same epidemiological study, for instance, when examining the associations of different parenting factors with NFSB or when considering the measures reported by different informants (mother/father or youth). In that case, we can assume that the corresponding effect sizes are dependent [58], and it is inappropriate to perform a standard meta-analysis because the assumption of conditional independence of effect sizes is violated [59]. A strategy could be to consider only one effect size per study and to perform separate meta-analyses for each type of exposure [60]; however, this implies that some associations are more valid or of greater priority and results in a loss of information. Moreover, it becomes impossible to examine the moderation effects between several exposures of interest (in our case, between parenting factors). The use of three-level models has thus been recommended to model dependent effect sizes without losing available information, especially in studies examining the role of different parenting factors that could influence each other [46]. Our three-level meta-analyses will allow us to consider (1) the effect size level, (2) the sample level, and (3) the study level.
The study design is known to influence the strength of the observed associations, especially since parenting is likely to influence NFSB and NFSB can also have an impact on parenting [61]. As results from cross-sectional studies do not allow these bidirectional associations to be disentangled, we will investigate their results separately from those obtained in longitudinal studies, while distinguishing longitudinal studies that examine the effects of parenting on NFSB from those studying the consequences of NFSB on subsequent parenting.
Heterogeneity will be assessed by visual inspection of forest plots, Cochrane’s Q, and Higgins’ test (I2). The I2 values, corresponding to the observed heterogeneity that would not be expected by chance, will be classified as low (< 30%), moderate (30–50%), and severe (> 50%) [62].
In case of significant heterogeneity, we will conduct moderator analyses considering the participants’ characteristics and methodological aspects of studies. Associations between parenting and NFSB were previously reported to differ according to age [63], child assigned sex [64, 65], and ethnicity [66]. Moreover, parenting takes place in a broader cultural and socio-political context, which differs widely according to participants’ countries of residence. Countries, as well as their income level defined by the World Bank as low- and middle-income countries (LMIC) and high-income countries (HIC), have been shown to influence the risk of suicide behaviors in children and youth [67, 68]. Among methodological aspects, the methods used for the assessment of parenting and NFSB (e.g., using questionnaire or observation data) and whether the informant is the child or the parent could also explain some differences observed in previous findings [28, 29]. We will thus consider participants’ age, sex, and ethnicity; countries and their income level; methods for assessment; and informants as potential moderators in our meta-analysis.
Publication bias will be evaluated through visual inspection of funnel plots and by using Egger’s test. The “trim and fill” method will be applied to correct for publication bias [69].
Sub-group and sensitivity analyses
If possible, sub-group analyses will be conducted according to different study settings (mental health care setting, other clinical setting, or population-based).
To examine whether the inclusion of studies with the highest risk of bias might affect our results, and in accordance with the Cochrane Handbook, sensitivity analyses will be performed by restricting the primary analysis to studies at low risk of bias, after exclusion of those whose quality was rated as “poor” on the Quality Assessment Tool for Observational Cohort and Cross-Sectional Studies. When possible, sensitivity analysis comparing the results between meta-analyses of adjusted and unadjusted data will be conducted to inform about the presence of confounding.
We will identify the effect size outliers, defined as effect sizes falling more than 2.2 standard deviations away from the pooled result, as well as small sample size outliers (n < 100) [70]. These outliers will be considered in a leave-one-out sensitivity analysis, which consists of performing separate meta-analyses on each subset of the studies obtained by iteratively leaving out one outlier [71].
Analyses will be performed using comprehensive meta-analysis and R.