Inclusion Criteria
All studies available from 1978 until June 2017 addressing the relation between attachment and depression from infancy to adolescence were included in the current meta-analysis. Multiple inclusion criteria were formulated to select the studies for the present review. First, the studies had to include a measure of the child’s attachment and of child’s depression. Second, studies with samples with an age range between 0 and 23 years were included. Third, only studies reporting on children’s attachment to their parents and/or primary caregivers were included in the present meta-analysis. Fourth, only studies in English, German, French, Italian, Portuguese, or Dutch were included. Fifth, the studies had to provide sufficient statistical information to calculate an effect size. Finally, only studies reporting on bivariate associations between attachment and depression were included, because in multivariate effect sizes, the set of covariates varies greatly among different studies. Therefore, combining and comparing differently adjusted effect sizes limit the ability to estimate a true overall relation between attachment and depression (Mulder et al.
2018).
We excluded studies on the relation between the child’s attachment relationship and maternal depression or studies reporting on associations between the child’s attachment to other persons (such as peers). Furthermore, studies reporting on internalizing symptoms but not specifically on depression (e.g., using the internalizing, anxious/depressed or social withdraw scales of the CBCL/YSR) and studies on parental bonding, family cohesion, or family conflict were excluded.
Selection of Studies and Limiting Publication Bias
According to the recommendations by Lipsey and Wilson (
2001), the following search strategy was conducted to find qualified studies. First, multiple electronic databases were searched: Ovid (including Medline, PsychINFO, and ERIC), Wiley Online Library, ScienceDirect, Academic Search Premier, EThOS, and ProQuest Dissertations & Theses. The search string comprised four elements: an attachment element, a depression element, a parent element, and an age element. For the attachment element, the following keywords were used: attachment, “parent–child relation*,” “mother–child relation*,” or “father–child relation*.” For the depression element, the following keywords were used: depressi*, dysthym*, “affective disorder*,” or “mood disorder*.” For the parent element, the keywords parent*, mother, father, caregiver, or caretaker were used. For the age element, the following keywords were used: infant, baby, babies, child*, toddler*, youth, adolescen*, “young adult,” or student. If possible, the keywords were entered in specific text fields of the databases (i.e., the title, abstract and/or keywords) to reduce the number of unqualified hits.
In systematic reviews, the aim is to include all eligible studies previously conducted (Lipsey and Wilson
2001). However, a common problem is that studies may not have been published because of non-significant or unfavorable findings, and therefore are difficult to locate, the so-called “publication or file drawer bias” (Rosenthal
1979). The consequence of publication bias is that the selection of studies is not an adequate representation of all previous studies that have been conducted. In order to prevent the problem of publication bias, we screened unpublished studies by searching the ProQuest Dissertations & Theses database and the E-theses Online Service database (EthOS). Most dissertations were publicly accessible. In case we found unpublished dissertations, we emailed the authors for the full text of the study, or ordered the study from the Proquest Dissertation Express. In addition, we emailed several attachment scholars to ask whether they knew of any unpublished articles.
The search for eligible studies was conducted by the first three authors independently. In case of any doubt, the other searchers were consulted. In total, 4892 titles were screened in the electronic databases. Further, we applied a snowball sampling method (i.e., screening the reference lists of relevant articles and the publication lists of attachment scholars) to find additional qualified studies. The initial search strategy yielded 508 studies (including reviews) of which the abstracts and methods sections were briefly read and excluded in case the study did not fit one of the inclusion criteria. Further examination of the full texts of 174 studies led to the inclusion of 124 studies, with 123 independent samples (s), 643 effect sizes (k), and a total of 54,598 participants in the current review. For a flow chart of the search procedure, see Appendix A. Appendix B presents the references of the included studies, and Appendix C presents the characteristics of the included studies.
Coding the Studies
The first author and a research assistant coded the included studies according to the suggestions of Lipsey and Wilson (
2001). The independent variable was attachment security. The dependent variable in this meta-analysis was depression. The potential moderators of the relation between attachment and depression were grouped into the following domains: study characteristics, sample characteristics, attachment, and depression characteristics.
For the study characteristics, we first coded the year of publication as a potential moderator, because we expected that the quality of recent studies was higher than the quality of older studies, as the statistical and methodological knowledge in social research has increased tremendously over the last decades. Second, the impact factor of the journal in which the study was published was coded, because the impact factor could be a first indication of study quality (Saha et al.
2003). Third, in order to assess the possible effect of publication bias on the association between attachment and depression, we coded whether the study was published in a journal or not. Fourth, at this point, it is not known whether the assumption that secure attachment relationships contribute to positive socio-emotional outcomes is true across cultures (Mesman et al.
2016), and if the strength of the association between attachment and depression varies across cultures. Therefore, the country of the research location (Anglo-Saxon and European countries vs. other countries) was coded. Finally, the study design was coded (cross-sectional vs. longitudinal design), as cross-sectional studies measure the relation between attachment and depression at one point in time (i.e., the co-occurrence of attachment insecurity and depression). Longitudinal studies take into account the developmental aspect of the association between attachment security and depression (i.e., attachment insecurity as a risk factor for depression). Also, the time (number of months) between the attachment and depression assessment was coded.
We coded various sample characteristics. First, we coded the mean age at the time of the attachment measure and the mean age of the depression measure, because in longitudinal studies, these can vary. In addition, we created three age categories: childhood sample (age range between 0 and 10 years old), pre-/early adolescence sample (age range between 9 and 15 years old), and adolescence/late adolescence (age range between 15 and 23 years old), because it is expected that the influence of parental attachment is stronger for younger children than for older children and adolescents (DeKlyen and Greenberg
2008). We choose to let the age ranges overlap, in order to increase the number of samples that could be categorized. For example, if the sample included children between 9 and 12 years old, it was categorized as a pre-/early adolescence sample. Studies with broad age ranges (e.g., 9 to 18 years old) were not categorized in this moderator.
Second, in previous meta-analyses on the relation between attachment and psychopathology, significant moderating effects of gender were found (Groh et al.
2017). Therefore, we coded whether the sample was an all-male, all-female, or mixed sample, and coded the proportions of males in the sample (continuous). Third, in line with previous meta-analyses on attachment and internalizing problems (Colonnesi et al.
2011; Madigan et al.
2013), we coded whether the families in the sample were at risk for psychopathology/attachment problems, whether it was a sample from the general population or it was a mixed sample, containing both at risk children or families and children from the general population. A sample was coded as at risk when either the parents or the children had mental health problems, when the child had experienced maltreatment or was in residential youth care, when one of the parents had died, when the children had academic risk factors (such as receiving special education), and when the sample consisted of teenage mothers. Finally, the percentage of children with Caucasian background in the sample was coded, because of possible cultural differences that may influence the association between attachment and depression (Mesman et al.
2016).
Various attachment variables were coded. First, we coded the attachment figure that was measured (parents, mothers, fathers, or general attachment representation of the child), because mothers and fathers could have unique and different influence on the development of children. Second, because the different attachment styles could influence the strength of the association between attachment and depression, we coded the type of attachment that was measured (i.e., secure, insecure-avoidant, insecure-ambivalent, insecure-disorganized, or a broad insecure measure). Third, we coded whether the attachment measure was a continuous or a categorical measure. Fourth, the type of instrument of the attachment measure (questionnaire, interview, or experiment/observation) was coded, because the different instrument may tap different elements of attachment (Bosmans and Kerns
2015). In addition, questionnaires are more sensitive to socially desirable responding than other attachment measures. Fifth, because previous meta-analyses showed moderating effects of the informant of the attachment measure on the relation between attachment and psychopathology (Colonnesi et al.
2011; Madigan et al.
2016), we coded whether the child, parent, or an observer reported on the attachment relationship. In the analyses, only the child and observer categories were included, because in only one study the parent reported on the attachment relationship.
We coded several depression variables. In line with the meta-analysis of Colonnesi et al. (
2011), we coded whether the study measured depressive symptoms or clinical diagnosis of depression, and coded the informant of the depression measure (child, parent, both, or others). Because too little studies assessed depression by both parent and child or others, we only coded the child versus parent informant effect. Lastly, we coded the instrument of the depression measure (questionnaire vs. interview), because interviews are considered as more objective instruments than questionnaires (Uher and McGuffin
2010).
Ten studies that were coded by the research assistant were randomly selected and double coded by the first author. The percentages of agreement for the moderator variables ranged from sufficient for the variables impact factor (86.7%), attachment Fig. (96.7%), attachment measure (93.3%), depression instrument (96.7%), to perfect (100%) for the variables publication status, publication year, country, study design, the age variables, gender, family risk status, percentage of Caucasians in sample, type of attachment, instrument of attachment, attachment informant, depression informant, depression measure, and time between attachment and depression measure. For the calculated effect size and the sample sizes, the percentages of agreement were 96.7% and 90.0%, respectively.
Calculations and Analyses
Effect sizes were reflected in correlation coefficients. We hypothesized that secure attachment relationships would be associated with less depression, and insecure attachment relationships to be associated with more depression. All correlations were keyed into the same direction so that these correlations could be compared to each other. A positive correlation indicated that the effect size was in line with our hypotheses, that is, attachment insecurity was hypothesized to be associated with more depression and attachment security with less depression. Cohen (
1998) formulated criteria that were used for interpreting effect sizes. Effect sizes around
r = .10 were considered as small, effect sizes around
r = .30 as medium, and effect sizes around
r = .50 as large.
If necessary, statistics were converted into correlational scores using the converter of Wilson (
2013), and formulas from Lipsey and Wilson (
2001). If a study only mentioned that an effect was not significant, the effect size was coded as zero (Lipsey and Wilson
2001). The continuous variables (publication year, impact factor, mean age of the sample, mean age at the attachment measure, mean age at the depression sample, months between attachment and depression measures, proportion of males, and proportion of Caucasians) were centered around their mean, and categorical variables were recoded into dummy variables. Extreme values of the effect sizes (> 3.29 SD from the mean; Tabachnik and Fidell
2013) were adjusted by winsorizing these outliers (i.e., replacing the outlier by the highest or lowest acceptable score falling within the normal range). Correlation coefficients
r were recoded into Fisher
z-values (Lipsey and Wilson
2001). In the reports on the overall relation between attachment security and depression and in the intercepts of the moderator analyses, Fisher
z-values were transformed back into correlation coefficients for the purpose of interpretation. Standard errors and sampling variance of the effect sizes were estimated using formulas by Lipsey and Wilson (
2001).
In the majority of the studies, it was possible to calculate more than one effect size, because for instance, the study reported on the correlation between different types of attachment and depression separately, or because multiple instruments and informants were used to assess attachment and depression. It is possible that effect sizes from the same study are more alike than effect sizes from different studies. Therefore, the assumption of independent effect sizes that underlie classical meta-analytic strategies was violated (Hox
2002; Lipsey and Wilson
2001). In line with recently conducted meta-analyses, we applied a multilevel approach to the current meta-analysis in order to deal with the dependency of effect sizes (Houben et al.
2015; Spruit et al.
2016). The multilevel approach accounts for the hierarchical structure of the data, in which effect sizes are nested within studies (Van Den Noortgate and Onghena
2003). Further, a multilevel meta-analysis enables using all effect sizes reported in the primary studies, so that all information is preserved and maximum statistical power is achieved (Assink et al.
2015).
We used a three-level meta-analytic model to calculate the combined effect sizes and to perform the moderator analyses, using instructions of Assink and Wibbelink (
2016). Three sources of variance were modeled, including the sampling variance for the observed effect sizes (level 1), the variance between effect sizes from the same study (level 2), and the variance between the studies (level 3) (Cheung
2014). The sampling variance of observed effect sizes (level 1) was estimated by using the formula of Cheung (
2014). Log-likelihood-ratio tests were performed to compare the deviance of the full model to the deviance of the models excluding one of the variance parameters, making it possible to determine whether significant variance is present at the second and third levels (Assink and Wibbelink
2016). Significant variance at level 2 or 3 indicates a heterogeneous effect size distribution, meaning that the effect sizes cannot be treated as estimates of a common effect size. In that case, we proceeded to moderator analyses, because the differences between the effect sizes may be explained by study, sample, attachment, and/or depression characteristics. Moderator analyses were only performed in case each category of the potential moderator was filled with at least three studies (Spruit et al.
2016). All significant moderators were subsequently entered in a multivariate model to examine the unique contribution in the explanation of the variance in the effect size distribution.
In case of heterogeneous effect size distribution, we are not able to test for publication bias. One of the assumptions made for statistical publication bias tests concerns the homogeneity of the data. If this assumption is not met, the publication bias tests cannot differentiate between heterogeneity and publication bias and might result in false positives or uninterpretable results (Ioannidis
2005). In the search strategy, however, we have made efforts to limit the possible effects of publication bias (see “
Selection of Studies and Limiting Publication Bias”). We additionally tested in a moderator analysis of publication status whether publication bias could affect the strength of the relation between attachment and depression.
The multilevel meta-analysis was conducted in R (version 3.4.4) with the metafor package, using a multilevel random effects model (Assink and Wibbelink
2016; Spruit et al.
2016). The restricted maximum likelihood estimate was used to estimate all model parameters, and the Knapp and Hartung (
2003) was used for testing individual regression coefficients of the meta-analytic models and for calculating the corresponding confidence intervals (see also Assink et al.
2015; Houben et al.
2015; Assink and Wibbelink
2016).