Intergenerational Stability in Parenting Across Two Generations: A Multi-Level Meta-Analytic Review

Parenting Dimensions

Intergenerational transmission of parenting may differ depending on the type of parenting assessed in both generations. Research on intergenerational transmission of parenting initially focused on harsh and abusive parenting (e.g., Hopps et al., 2003; Huesmann et al., 1984), and later on constructive parenting (e.g., Jeon & Neppl, 2016; Kerr et al., 2009). Apart from distinguishing negative and positive forms of parenting, parenting can also be classified based on content. Two distinguishable dimensions are acceptance and control (Baumrind, 2013; Bornstein, 2015), and sometimes a third negative dimension of conflict and hostility is added (e.g., O ‘Connor et al., 2002).

Parental acceptance refers to the quality of the affective bond between parents and children as well as the behaviors parents use to express their feelings towards their children (Rohner et al., 2005). In contrast, parental control refers to how parents guide, manage, or regulate their children (Ballash et al., 2006). Scholars have noted that this relatively heterogeneous domain of control may be better understood by distinguishing parenting characterized by a lack of pressure, intrusiveness, and dominance (i.e., parental autonomy support) from parenting that involves guiding children to engage in positive behavior and refraining from negative or dangerous behavior (i.e., parental structure; Grolnick & Pomerantz, 2009). This latter dimension of parental structure entails behaviors such as monitoring and setting rules and is expected to be relatively orthogonal to the other parenting dimensions (Soenens & Beyers, 2012). Lastly, we defined negative parenting as parental behavior where there is the intention or risk of hurting children (Backhaus et al., 2022), as well as any parental feelings of enmity, anger, resentment, and conflicts between parents and children (O’Connor, 2002).

Intergenerational studies can examine homotypic (e.g., G1 acceptance predicts G2 acceptance) or heterotypic (e.g., G1 acceptance predicts G2 structure) stability in parenting across generations (Thornberry, 2016). Such predictions can be theoretically grounded; for example, parents who grew up experiencing hostility may try to avoid expressing hostility toward their own children (low homotypic stability)but they may still have trouble setting boundaries (heterotypic stability). An empirical study showed that parent-adolescent relationship quality—defined by trust, communication, and low levels of alienation—predicted next-generation parents’ acceptance (i.e., self-reported and observed warmth and sensitivity) but also indicators of control, albeit less consistently (i.e., observed child management but not self-reported inconsistent discipline; Friesen et al., 2013). As such, it may be expected that measures of G1 and G2 parenting will be more strongly related if both measures fall within the same domain, as opposed to different domains.

G2 Children’s Age

Parenting in different periods of childhood (e.g., toddlerhood vs. adolescence) may have distinctive contributions to the children’s development. First, various developmental theories, including attachment theory, propose that early relationships play a foundational and shaping role in how interpersonal relationships form throughout development (Roisman & Fraley, 2013; Sroufe, 2000). Children’s development has been conceptualized as involving building blocks (Cicchetti & Richters, 1993), whereby the development of age-appropriate skills at each stage (e.g., developing impulse control in early childhood) provides the foundation for development at the next stage (e.g., positive adjustment to school). Thus, parenting at an early age may be highly predictive of children’s future development and their future behavior toward their own children. A second factor is that young children are more dependent upon their parents than older children. Time spent with parents also decreases as children get older (Price, 2008), and parental influences may wane compared with other influences, particularly from peers. Based on these considerations, G1 parenting in early development may be more strongly related to G2 parenting than G1 parenting later in G2 development.

At the same time, some theorists question the notion that early life experiences have a disproportionate influence on later development, suggesting that continuities in environmental circumstances may instead explain why early childhood experiences are relatively strongly associated with later developmental outcomes (e.g., Clarke & Clarke, 2000; Kagan, 1998). Later childhood and adolescent years may provide several opportunities to model and communicate both positive and negative parenting behavior as compared to early childhood. As such, the effects of G1 parenting in early childhood on G2 parenting may not necessarily endure and they may be altered or superseded by later parenting experiences. For example, Thornberry and Henry (2013) reported that G2 childhood maltreatment experiences that did not last into adolescence were not associated with G2 maltreatment of G3. Rather, only maltreatment experiences that lasted into or started in adolescence predicted maltreatment perpetration. Based on these considerations, the association between G1 and G2 parenting may not be dependent on G2 age in childhood—given that in most families there is considerable stability in parenting—or may even be slightly larger when G1 parenting was measured in older G2 children; more recent experiences might potentially suppress earlier experiences in those families where experiences change throughout childhood.

Guided by the notion that parents should adapt their behavior in response to their child’s development, parents may also be most likely to emulate the parenting they received in the same developmental period (Bornstein, 2015; Conger et al., 2009). For example, whether a parent adequately addresses the developmental needs of an adolescent may be more dependent upon their own experiences with their parents as an adolescent than as a toddler. Thus, a third possibility is that the association between G1 and G2 parenting is strongest when parenting of both generations is measured in the same developmental period.

G2 Timing of Parenthood

Although the transition to parenthood can provide opportunities for individuals to adopt adult roles that potentially disrupt intergenerational stability, these opportunities may be less likely to occur when individuals transition at a very young age (Kerr et al., 2011). A peculiarity of three-generation studies is that they tend to focus on younger G2 parents, as they are the first subset of parents that can be analyzed. This characteristic has led some researchers to hypothesize that current estimates of intergenerational stability are overstated (Belsky et al., 2012). Individuals who have more life experiences before they become parents may be less affected by their childhood experiences because they have developed a more certain idea of who they are (i.e., greater self-concept clarity; Diehl & Hay, 2011; Lodi-Smith & Roberts, 2010), more time has passed since being parented, and they have gained more relational experiences outside the family of origin that may have superseded their parenting experiences. In addition, compared to younger parents, older parents are often more independent from their family of origin, and they are generally in a more favorable socioeconomic position (Bray et al., 2006). As a result, they have less need to rely on grandparental involvement for childcare (Fergusson et al., 2008). So far, although theoretically convincing, empirical studies have not found evidence for such moderation within study samples (Belsky et al., 2012; Raby et al., 2015; Shaffer et al., 2009). In this meta-analysis, we examine the extent to which studies of younger G2 parents report larger effect sizes than studies of older G2 parents. Relatedly, we examine if effect sizes are larger when less time has passed between the assessments of G1 and G2 parenting.

Gender

Gender is one of the most fundamental characteristics by which human social lives are organized (Blakemore et al., 2008; Heymann et al., 2019). Importantly, gender-role attitudes and behaviors tend to become more traditional after the transition to parenthood (Baxter et al., 2015; Yavorsky et al., 2015); a trend that continues throughout parenthood (Endendijk et al., 2018). On average, Western mothers spend at least twice as much time on childcare compared to fathers (Organization for Economic Cooperation and Development; OECD, 2016). Mothers are traditionally considered the primary parents; norms for mothers regarding their central and enduring role are stricter than those for fathers (Doherty et al., 1998).

In contrast, fathering is typically a more contextually sensitive process than mothering. For instance, fathers are more likely to lose contact with their children or become significantly less involved in their children’s lives after divorce (Cano & Gracia, 2022; Ivanova & Kalmijn, 2020). Moreover, fathers in high-risk environments (e.g., living in higher-crime neighborhoods, experiencing low-income levels) are less likely to be involved in their child’s life and thereby also less likely to meet the inclusion criteria of intergenerational studies (Capaldi & Kerr, 2023). A study on intergenerational intervention effects reported that at least 95% of the G2 mothers met the study’s eligibility criteria (regular, face-to-face contact with their G3 child), whereas this was only 75% for the G2 fathers in some of the assessment waves (Hill et al., 2020). The lower participation rate of fathers, with potentially the highest-risk fathers not participating in studies, may constrain the ability to detect effects for fathers.

Gender roles and differences have several implications for how G1 gender may moderate the strength of intergenerational stability in parenting. With women more regularly in the role of primary caregiver, and fathers’ risk of noninvolvement and study dropout, G1 mothers’ parenting may be more strongly associated with G2 parenting compared to G1 fathers’ parenting. Studies that have examined associations between mothers versus fathers thus far have found differing effects (Gettler et al., 2019; Macdonald et al., 2018). Nevertheless, most empirical work on gender differences in intergenerational stability in parenting has not examined G1 gender, largely because of sampling restrictions (Kerr & Capaldi, 2019).

In addition to contextual factors that may disrupt father-child contact, other factors may also lead to differences in intergenerational stability of parenting for G2 mothers versus fathers. Compared to mothers, fathers may be more strongly influenced by their partner’s parenting (Capaldi & Kerr, 2023; Capaldi et al., 2008). As mothers are often the primary caregivers, G2 fathers may feel that their partners are more knowledgeable about parenting approaches, therefore defaulting to their partners’ lead. This potentially diminishes the influence of fathers’ families of origin. Some studies indeed find larger G1 parenting effects for G2 mothers compared to G2 fathers (Belsky et al., 2012; Thornberry et al. 2003), although another study found larger G1 parenting effects for G2 fathers compared to G2 mothers (Savelieva et al., 2017).

G2 gender may also moderate the degree to which G1 mothers versus fathers influence G2 parenting practices. Social learning theory predicts a greater tendency for individuals to imitate same-gender models (Perry & Bussey, 1979); thus, G1 fathers’ parenting could be more strongly associated with that of their sons than that of their daughters (and similarly for mothers and their daughters). Evidence supporting gender-specific intergenerational associations includes that parents tend to spend more time with same-gender children (e.g., McHale et al., 1999), that the parenting of children’s same-gender parent is a stronger predictor of children’s developmental outcomes than that of the opposite-gender parent (e.g., externalizing problem behavior; Vanassche et al., 2014), and that parents tend to have a warmer relationship with their same-gender children (Shanahan et al., 2007). Applying these theories and findings to the intergenerational transmission of parenting, it is expected that G2 fathers’ parenting will show a stronger resemblance to G1 father’s parenting, whereas G2 mothers’ parenting will show a stronger resemblance to G1 mothers’ parenting.

Ethnicity

Transmission processes may operate differently across ethnic communities, as many communities hold different beliefs regarding appropriate parenting (Farver et al., 2007). Given that the meta-analysis included studies from many countries, measures of race and ethnicity were not expected to index comparable cultural identity groups across studies. However, families who identify as the ethnic minority within a country may face discrepancies between their own cultural and social capital and that of the dominant culture and institutions (Coll & Pachter, 2002). Parents and children with a minority background may differ in the degree to which they acculturate with dominant societies, with children on average being more engaged with the dominant society than their parents (Smokowski et al., 2008). It is therefore expected that studies of samples with fewer families who identify as the ethnic majority in their national context will report less intergenerational stability in parenting than studies with samples of more families who identify as the ethnic majority.

Methodological Characteristics and Quality as Moderators of Intergenerational Stability

Besides substantive moderators, methodological characteristics and quality may also affect estimates of intergenerational stability in parenting. First, most if not all three-generation studies are extensions of existing studies that had different research objectives (e.g., predicting delinquency or substance use problems). As a result, the samples of these studies vary substantially. A handful of studies are representative of the region, and these studies may produce more trustworthy estimates of intergenerational stability. Various longitudinal three-generation studies also focus specifically on at-risk samples, as these samples suited the original research questions (e.g., Bailey et al., 2009; Capaldi et al., 2003; Farrington et al., 2015). There are reasons to believe that transmission of parenting may be stronger in these at-risk samples. Models such as the Family Stress Model (Conger et al., 2010) highlight that stability in characteristics such as parenting is intertwined with stability in contextual factors. Studies have shown that the stress associated with economic risk factors can constrain parents’ ability to interact effectively with their children (Hardaway & Cornelius, 2014; Neppl et al., 2016; Newland et al., 2013), and these risk factors are found to be relatively stable across generations (Musick & Mare, 2006). In this meta-analysis, we therefore examined whether the type of sample (nationally representative, at-risk, low-risk non-representative) affected the effect size reported in studies. We expected studies with at-risk samples to report larger effect sizes than both nationally representative and non-representative low-risk studies.

Second, how G1 and G2 parenting are measured may be related to the reported degree of intergenerational stability in parenting. Parenting can be assessed in various ways, such as through interviews, observations, and questionnaires. In general, the association between observed and reported parenting is found to be weak (meta-analysis; Hendriks et al., 2018). Questionnaires on parenting often involve both cognitions and behavior, and they do not always distinguish satisfactorily between unidirectional and relational aspects of parenting (Lovejoy et al., 1999; Mastrotheodoros et al., 2018; Verhoeven et al., 2017). In addition, a large part of the variance in questionnaires on family functioning can be accounted for by the subjective experience of the reporter on that questionnaire, and thereby inevitably capture relational schemas or working models of relationships (Branje et al., 2003; Eichelsheim et al., 2009). Observational measures may be better suited to assess parental behavior (Gardner, 2000). In this meta-analysis, we examined if the type of measurement of G1 and G2 parenting explains heterogeneity in effect sizes. We distinguished between questionnaires (G2 or other-reported), observations, and multimethod measures.

Third, studies that rely on different reporters or methods for G1 and G2 parenting are considered to have a stronger design, because this avoids reporter bias (Thornberry et al., 2012). When considering reported parenting, children and parents are routinely found to report differently on a variety of parenting domains (De Los Reyes & Ohannessian, 2016). It has been argued that the child’s perception of parenting may be most influential in predicting developmental outcomes (Rohner & Melendez-Rhodes, 2019). Therefore, we also examined if the association between G1 and G2 parenting is larger when both are reported by G2 (i.e., in the presence of shared-reporter variance).

Fourth, the samples of G2 participants as parents are typically substantially smaller than the original sample, in part because not all participants have children. In addition, the long timespan between assessments of G1 and G2 parenting might induce selective attrition, where high-risk participants have a higher chance of dropping out of the sample (e.g., Assendorpf et al., 2014). We therefore also examined if the association between G1 and G2 parenting differed depending on the degree of study attrition.

Inclusion Criteria

The following inclusion criteria were defined a priori: (a) The study had a predictor variable concerning parenting by G1 parents and an outcome variable concerning parenting by G2 parents. We excluded studies that only measured characteristics of parents that were not directly related to their child (e.g., parental depression). We also excluded studies that only included measures of family environment processes, which can encompass behaviors of family members other than a focal parent or child. (b) The mean age of G2 and G3 children at the time of the G1 and G2 parenting assessments, respectively, was 18 years or younger. (c) The study involved non-retrospective measures of parenting. (d) The study had a multigenerational, longitudinal design. (e) The study was described in a peer-reviewed empirical article and written in English. (f) Information to calculate effect sizes (i.e., correlations) was present or available upon request. (g) The participants did not receive any intervention. For intervention studies, we only used baseline measures or restricted the analyses to the control group. Later in the study process, we also determined that measures of parenting should not include child abuse. Sexual abuse is often part of composite scores representing child abuse but is qualitatively different from any parenting dimension. Additionally, there have been recent meta-analyses on intergeneration transmission of child abuse (Assink et al., 2018; Madigan et al., 2019).

Search Strategy

We searched for longitudinal studies that included two generations of parents using the electronic databases APA PsycInfo, ERIC, Scopus (social sciences), and Web of Science¹ (social sciences). All English-language articles published in peer-reviewed journals before November 2021 were included in the search. The search strings can be found in the supplementary material and included keywords within three categories: (a) parenting² (including general parenting [e.g., parent* style], acceptance [e.g., parent* warmth], control [e.g., parent* monitor* and psychologic* control], negativity [e.g., parent* aggressi*], parent-child relationship quality [e.g., mother adj3 conflict*], and cognitions [e.g., parent* adj3 self-efficacy]); (b) a multigenerational (e.g., intergeneration*); and (c) longitudinal design (e.g., prospective). Studies had to have at least one keyword within each of the three categories in their title, abstract, or keywords to be included. Additionally, studies were located by reviewing the bibliography of a recent review on intergenerational transmission of parenting (Kerr & Capaldi, 2019). Finally, the authors of the large three-generation datasets included in this meta-analysis were contacted to confirm we had included all relevant articles from their dataset.

Selection of Studies

The selection process for this meta-analysis is schematically depicted in Figure 1. After removing duplicates, 5063 unique studies were screened on the inclusion criteria using the information in the title, abstract, and keywords in Rayyan (Ouzzani et al., 2016). In total, 548 (11%) articles were screened by the first and second authors, with an agreement rate of 92% (κ = .67), and any disagreements were discussed and resolved. The remaining articles were screened by one of the two authors. Next, full-text articles (n = 546) were screened based on the inclusion criteria. We took a stepwise approach, starting with screening for (a) human subjects (label: animal study); (b) English peer-reviewed empirical article (label: wrong publication type); (c) quantitative data (label: wrong study design); (d) samples that did not receive any treatment (label: wrong study design); (e) parenting measured in both G1 and G2 participants (labels: wrong outcome and/or wrong predictor); (f) prospective measures of parenting (label: wrong study design); (g) a longitudinal study with one measurement of G2 aged <18 years (label: wrong study design); and (h) two generations of parents (label: wrong population). Whenever a study did not fulfill a criterion, we stopped screening and labeled the article with the first inclusion criteria it did not satisfy. For this step, 59 (11%) articles were screened by the first and second authors, with an agreement rate of 95% (κ = .70), and any disagreements were discussed and resolved. The remaining articles were screened by one of the two authors.

Initially, 73 studies were included in the meta-analysis. We contacted the authors of 41 of the articles for additional information, as one of the following situations occurred: (a) correlations were not (all) reported, (b) the study included both participants with and without children (e.g., when parental attitudes were measured). In this case, we requested correlations and basic information about the parents in the sample. (c) The study included a treatment and control group. In that case, correlations and basic sample information for the control group were requested. And (d), the measures about parenting were not specific enough to fall within the control or acceptance domain. In that case, we requested correlations for specific subscales.

In several cases, multiple studies were based on the same dataset. When the same effect size was reported in two publications (i.e., based on the same measure of G1 and G2 parenting and the same dataset), we included the effect size with the largest sample size. Typically, this effect size was in the most recent publication, as data are continuously collected in ongoing three-generation studies. When publications reported composite scores that were only partially overlapping, we included both composite scores.³

During the coding process, 22 studies were excluded for the following reasons: either G1 or G2 parenting included child abuse (n = 8), all variables had been addressed in other articles on the same dataset (n = 4), the sample size after excluding nonparents was too small to calculate correlations (n = 1), the average age of children was above 18 years (n = 3), no correlations were provided and it was impossible to calculate an effect size (n = 5), the measures were not considered to be parenting after closer inspection (n = 1), or the study was not longitudinal but cross-sectional with adolescent mothers (n = 1). Subsequent screening procedures of titles and abstracts as well as full texts in November 2021 resulted in the inclusion of one additional article (see Figure 1). Eventually, 51 studies on 24 datasets were coded in this meta-analysis. The included datasets are listed in Table 1.

Data Extraction

We coded study characteristics and effect sizes using a detailed coding manual. A total of 10 articles (20%) were coded by the first and second author to determine interrater reliability. Deviating codes were discussed between coders until a consensus was reached. The remaining studies were coded by one of the two authors. If required information could not be retrieved it was coded as missing. Interrater reliabilities (Intraclass Correlation [ICC] or kappa) were good to excellent and ranged between .71 and 1.00.

Meta-Analytic Procedure

All analyses were conducted using the metafor package (Viechtbauer, 2010) in R 4.2.2 (R Core Team, 2022). To assess the overall correlation between G1 and G2 parenting, we applied a random-effects model, which assumes that heterogeneity in effect sizes is caused by both sampling variability as well as variability in true effects (Borenstein et al., 2009). All parameters were estimated with the restricted maximum likelihood estimator, and the Knapp and Hartung method was used to calculate the standard errors of the coefficients in the model (Knapp & Hartung, 2003). To determine the magnitude of the effect sizes, we followed the guidelines of Cohen (1992), with correlations of .10, .30, and .50 indicating small, medium, and large effect sizes, respectively.

Because effect sizes could be from the same dataset, we used a three-level approach to account for their dependency. In these analyses, the first level reflects the sampling variance around the estimated population effect size, the second level represents the variance between effect sizes within the same dataset, and the third level reflects the variance between datasets (Cheung, 2014). The significance of level 2 and level 3 heterogeneity in effect sizes was determined following the guidelines of Assink and Wibbelink (2016). Specifically, two independent one-sided loglikelihood-ratio tests were performed, in which a full model— allowing for level 2 and level 3 variance—was compared with the model without either level 2 or level 3 variance.

If the analyses yielded significant heterogeneity between effect sizes (i.e., level 2 and/or level 3 variance), the hypothesized moderators were added to the model to examine whether they might explain this heterogeneity. Continuous moderators were centered on their mean. All moderators were entered separately in the model. Subsequently, a multiple moderator test was conducted with the moderators that were significant in the univariate tests.

Publication bias was assessed in multiple ways. First, we visually inspected a funnel plot. A funnel plot is a scatter plot in which effect sizes are plotted against their precision, which, in this case, is the inverted standard error. Without any selection bias, it is expected that the effect sizes of studies with a larger sample are closer to the overall meta-analytic effect. In addition, smaller studies would be widely and evenly spread on both sides of the overall effect, resulting in a symmetrical, inverted funnel shape. Asymmetrical funnel plots signal the presence of small-study effects, although they may also arise for other reasons such as between-study heterogeneity (Terrin et al., 2003).

Second, we applied the PET-PEESE method (Stanley & Doucouliagos, 2014). Both PET and PEESE are regression models in which effect sizes are regressed on a measure of precision (i.e., the standard error for the PET model and the variance for the PEESE model). As such, the intercept of the models reflects the estimated effect size for a hypothetical study with an infinite sample size and no error. The models are weighed for the studies’ inverted variances to give more weight to studies with more precision (i.e., studies with a larger sample size). If the intercept in the PET model is not significant, it can be interpreted as the true underlying effect. In contrast, if the intercept of the PET model significantly differs from zero, the PET model is biased and the intercept of the PEESE model provides the most reliable estimate of the effect size.

Third, three-parameter selection models were fitted using the R Weightr package to examine the impact of potential selective publication (3PSM; Iyengar & Greenhouse, 1988) (Coburn & Vevea, 2017). Specifically, we examined whether effect sizes with a one-sided p-value below .10 (i.e., including both significant and near-significant effect sizes) were more likely to be published than effect sizes with a p-value above .10. To do so, we compared an unadjusted model with a bias-adjusted model with default p-value cut points of .05 (equaling a one-sided p-value of .10) and 1, using a likelihood ratio test. Significant values indicated that the probability of being published differed between (near-)significant and non-significant effect sizes. Lastly, we tested whether high-impact journals and older articles reported larger effect sizes by including the journal’s impact factor and year of publication as predictors in the analyses.

Although it is common to ignore dependencies between effect sizes within the same dataset when testing for publication bias, this results in inflated Type 1 error rates (Rodgers & Pustejovsky, 2021). We, therefore, extended the logic of the PET-PEESE model to a multilevel model by running two models in which the effect sizes were regressed on their standard error (PET) and variance (PEESE), weighted for the studies’ inverted variances using the lme4 package (Bates et al., 2015; See also Ujma, 2022). In addition, we conducted multiple 3PSM tests on repeated random samples of a single effect size from each dataset.

Transparency and Openness

This review’s protocol, including the analysis plan, was preregistered at https://www.crd.york.ac.uk/prospero/display_record.php?ID=CRD42020175231. We followed the APA Style Meta-Analytic Reporting Standards reporting guidelines for the final report. Any deviations from the preregistration, and the rationale for these deviations, can be found in the supplementary material. The search string, coding scheme, data, and script for this meta-analysis are available online at https://osf.io/3nts6/. We report how we determined our sample size, all data exclusions, all manipulations, and all measures in the study.

Descriptive Characteristics

The present meta-analysis included 51 publications reporting on 24 unique datasets (k), published between 1984 and 2021 (median year: 2013), with data collection beginning between 1958 and 1998 (median start year: 1989). The datasets were collected in North America (k = 17), Europe (k = 3), Oceania (k = 3), and Asia (k = 1). Datasets were classified as nationally representative samples (k = 5), (predominantly) risk samples (k = 10), or other non-risk samples (k = 10)⁴. Two studies were intervention studies, therefore only the control group was included in this meta-analysis. On average, 87.48% of the samples’ participants identified as the country’s ethnic majority (SD = 18.13, range 5–100%). Characteristics of all coded datasets can be found in Table 1 and Table S1 in the Supplementary material.

In total, 408 effect sizes (j) were extracted, see Figure 2 for the distribution of effect sizes. Sample sizes for these effect sizes ranged from 17 to 2338 (mean n = 373.32). Attrition rates were on average .66 (SD = .16 ; range = .30 to .98), indicating that, on average, effect sizes were based on less than half of the original sample size. The average time between the assessments of G1 and G2 parenting was 14.81 years (SD = 6.52; range = 6.00 to 34.10 years). G2 average reported age was 13.45 years (SD = 3.71, range = 1.25 to 17.5) in childhood and 23.58 (SD = 3.14, range = 18.50 to 32.30) when becoming parents of G3. The average age of G3 children was 4.93 years (SD = 3.72, range = 0.21 to 15.8). On average, G2 and G3 children differed by 9.89 years in age (SD = 5.12, range = 0.00 to 16.59).

Overall Effect Size

Overall, a small positive association between G1 and G2 parenting was found (r = .12, 95% CI .09 - .16, p < .001; Table 2), which indicates significant but modest stability in parenting across generations. Most of the retrieved effect sizes (j = 355; 87%) were in this expected direction. Negative effect sizes were typically small, with a median value of −.03. The prediction interval ranged between −.06 and .30, indicating that future effect sizes of intergenerational stability in parenting will most likely range between these values. Thus, overall, without considering potential moderators and the possibility of publication bias, there is modest stability in parenting across generations.

Moderator Analyses

Likelihood ratio tests indicated significant within (σ² level 2 = .003) and between study (σ² level 3 =.005) variance (Table 2). The variance at level 1 was estimated to comprise 24.82% of the total variance, with 24.11% at level 2 and 51.07% at level 3. This heterogeneity may be explained by moderators.

The results of the moderator analyses are presented in two ways. First, Table 2 presents the results of the meta-regression models, and provides the intercept and regression coefficients for all models with continuous and categorical moderators. For all categorical moderators, Table 2 also presents the average correlation for each specific moderator category (e.g., G1 mothers, G1 fathers), and Table S2 shows which moderator categories differ significantly from each other. For all continuous moderators, Table 2 presents the average correlation at various levels of the moderator.

Second, the results of all moderator analyses are visualized. For all analyses involving categorical moderators, the average correlation and 95% Confidence Interval for each moderator category are visualized in Figure 4. Conceptually, Figure 4 indicates whether associations between G1 and G2 parenting for a specific moderator category are significantly different from zero and from the other moderator categories. As the sample sizes differed between analyses (i.e., due to listwise deletion caused by missing information about moderators), the average correlation slightly differed between analyses. Therefore, the solid blue line in Figure 4 provides the average correlation for each moderator analysis.

For all continuous moderators, Figure 5 displays the regression lines and confidence bands across the full range of moderator values. The solid blue line represents the average correlation within each moderator analysis. The circles represent used data points in these analyses, with their size reflecting the corresponding sample size. Data points from the same sample are shown in similar colors.

Publication Bias and Sensitivity Analyses

Inspection of the funnel plot (Figure 3) indicated asymmetry in the distribution of effect sizes, with most of the smaller studies clustering on the positive side of the plot. The PET-PEESE models estimated the true effect to be r = .12 (p < .001; PET) and r = .11 (p < .001; PEESE). When compared to the unadjusted overall estimate of r = .12, little evidence was found for small-study effects. Moreover, effect sizes were not dependent on journals’ impact factors (F(1,405) = 0.03, p = .853) nor on publication year (F(1,406) = 2.44, p = .119).

Results from the 3PSM models on 50 repeated random samples of single effect sizes per dataset also indicated that the estimated model adjusted for publication bias on average did not fit the data significantly better than the unadjusted model (mean p .233). Most randomly drawn samples did not produce significant p-values for the 3PSM analyses (n = 33; 66%). Overall, few indications for publication bias were found. Results also did not change when transformed Beta’s were excluded from the data (Table S4 in the Supplementary material).

Moderators of Intergenerational Stability in Parenting

The relatively broad prediction interval around the overall association between the parenting that individuals experienced in their youth and the parenting they enacted later with their own children suggests there may be circumstances under which stability is stronger or weaker. Indeed, effect sizes differed within and across studies which could be explained by both theoretical and methodological moderators.

Limitations, Future Directions, and Implications

The current meta-analysis comes with several limitations and directions for future research. First, studies from the United States were heavily overrepresented in this meta-analysis, and with one exception, all studies in this meta-analysis were from Western countries. Although we only included English-language studies, there seem to be few longitudinal studies of multigenerational parenting conducted elsewhere. Additionally, many of the samples were predominantly White (e.g., Conger et al., 2013; Kerr et al., 2009), which further hampers the examination of cultural differences. Intergenerational stability in parenting may be universal to some extent, but cultural differences have also been reported. In a sample of parents from 12 cultural groups and nine countries, Rothenberg et al. (2023) found evidence for a universal association between G2 parental rejection and their recollections of G1 parental rejection, but the magnitude of this association varied across cultures. Stability was generally higher in cultural groups in which parental rejection was more normative. Perhaps, ongoing longitudinal studies that focus on ethnically and regionally underrepresented samples will expand into three-generation studies in the upcoming years, creating the opportunity to examine further cultural differences in the intergenerational transmission of parenting. Moreover, multi-sample collaborations considering ethnic differences may form a promising avenue for future research.

A second limitation is that moderators were examined at the study level. This approach is well suited to examine moderators that do not vary substantially within studies but it also runs the risk of committing the ecological fallacy for moderators that do vary substantially within studies. As an illustration, G2 age in childhood did not vary greatly within studies, as participants are generally sampled based on their age. Hence, the study means were generally a good representation of the study participants’ age. However, the age at which G2 participants became parents varied substantially within studies, and the aggregated age may not be a good representation of the ages at which individual study participants became parents. Future efforts to synthesize individual participant data are better suited to examine individual characteristics such as ethnicity and age when becoming a parent.

The results of this meta-analysis may also have clinical implications. Various parenting programs address how parents look back on their childhood and how they feel this may affect their current functioning (Chamberlain et al., 2019). As prospective and retrospective reports on parenting only partly align (Nivison et al., 2021), the results of this meta-analysis provide limited information for such approaches. Yet, although the focus of parenting programs is mainly on how to improve the immediate family situation, longitudinal follow-up studies indicate that the effects of such programs may also affect G2 (parenting) behavior years later (Mahrer et al., 2014; Rothenberg, Lansford, Godwin, et al., 2023). It has been proposed that extending the focus of parenting programs by also explicitly investing in families with an eye on future generations may extend their public health impact (Chen et al., 2016). The current meta-analysis provides tentative evidence for intervention targets; interventions may be more effective in disrupting negative intergenerational cycles if they focus on improving parental acceptance and avoiding negative parenting, in comparison to behavioral control. Further, present findings indicate that interventions earlier in the child’s life may be more effective in disrupting negative intergenerational cycles than later interventions.