Exploring a causal model in observational cohort data: The role of parents and peers in shaping substance use trajectories
Introduction
Data from observational cohort designs, rather than randomised designs, are often required to track trajectories of substance use behaviour. Thus, research seeking to examine predictors of substance use trajectories are required to work within this methodological framework. However, the observational nature of such data, is closely tied to the pervasive view that causal questions can only be examined within randomised designs (Hernán, 2018). This is due to the issue of confounding within observational studies (Lewis & Kuerbis, 2016). Confounding factors are most simply understood as alternative factors that might be responsible for the observed relationship between an exposure and outcome (VanderWeele, 2019). Whilst randomised designs address confounding through randomisation, observational studies are without such an inherent property. For these reasons, studies examining the predictors of substance use trajectories have yet to interpret estimated associations as causal, and the accompanying use of causal language is scarce.
Advancements in statistical methods are now providing formal approaches and tools for investigating causal effects within observational studies (Hernán, 2004, Pearl, 2009, Rohrer, 2018, VanderWeele, 2019). The counterfactual framework (Lewis, 1973) provides a formal statistical framework for estimating causal effects by accounting for confounding factors. A counterfactual is a “What If?” scenario; for example, what substance use patterns would we observe in those reporting high quality parent-adolescent relationships (i.e., the factual/observed relationship), had they alternatively experienced low-quality parent-adolescent relationships (i.e., the counterfactual)? Counterfactual levels of the outcome exist at all potential exposure levels if variables extend beyond binary (Hernán, 2004).
When using observational data, counterfactuals can be estimated based on the factual (observed) data (Hernan, 2006) when certain assumptions are met. The key assumption underlying this assertion is conditional exchangeability, which implies that conditional on an appropriate set of confounders (i.e., after adjusting for confounders in a regression), the exposure can be considered as effectively randomised and causal effects can be attributed to the exposure (Hernán, 2004, Rubin, 2005, Sauer and VanderWeele, 2013). When assumptions are met, a hypothetical scenario can be estimated in which we know both what happened, and what would have happened, under the counterfactual conditions, permitting the sample to act as its own control for any potential confounding factors.
Determining which variables are necessary to be included as confounders, so that conditional exchangeability holds, remains a theoretical and empirical challenge. The Direct Acyclic Graph (DAG) (Greenland et al., 1999) is one tool used to identify an appropriate set of confounder variables. DAGs are a graphical model developed by the researcher, which defines how exposures, outcomes, and possible confounders are related to one another. Fig. 1 presents an example DAG with some key causal relationships that need to be considered when examining a causal pathway of interest. The specific causal relationships are developed based on substantive understanding; for which DAGs provide the rules to then determine which variables need to be adjusted for in regression analysis to mitigate confounding (Rohrer, 2018, Sauer and VanderWeele, 2013).
Among other considerations, DAGs have helped to illuminate the importance of avoiding adjusting for mediators (a variable that falls on the causal pathway of the exposure to the outcome) and colliders (a variable that is a common causal outcome of both the exposure and outcome), as can be seen in Fig. 1. We encourage readers to review Rohrer (2018) which describes and visualises these common confounder cases in detail using DAGs. However, DAGs require knowledge of the complete underlying causal structure, including the relationships amongst all possible confounders, which in settings with numerous potential confounders is often unavailable (VanderWeele, 2019).
VanderWeele and colleagues (VanderWeele and Shpitser, 2011, VanderWeele, 2019) have proposed the disjunctive cause criterion for confounder selection as an approach to alleviate the burdensome task of determining a complete causal structure. The disjunctive cause criterion suggests that a pre-exposure variable should be considered as a confounder if it causes either the exposure or outcome or both. By selecting pre-exposure confounders, adjusting for both mediators and colliders can be avoided. Known instrumental variables (i.e., variables that are related to the outcome only through the exposure; see Fig. 1) should be excluded, although, substantive knowledge of instrumental variables is often unavailable (VanderWeele, 2019). Additionally, when measurement of confounders of both the exposure and outcome are unavailable, as is potentially the case when using pre-existing data, proxies (e.g., an outcome of a confounder that is not itself related to the outcome; see Fig. 1) can be added. Both instrumental and proxy variables situations have been described in detail by VanderWeele (2019).
Importantly, for every DAG that has a subset of variables appropriate for adjustment, the subset of variables selected by the modified disjunctive cause criterion will also suffice (VanderWeele & Shpitser, 2011). Nevertheless, the ability to attributed identified associations as causal relies heavily on the confounders selected, and there remains no guarantee that a suitable set of confounders are available or selected. This means that unmeasured confounders still have the potential to undermine claims of causality; an issue which remains untestable in observational studies (Hernán, 2012). The use of a causal framework, however, provides the researcher with the opportunity to make explicit their assumptions regarding confounding.
One example relevant to substance use researchers pertains to the examination of how both parent-adolescent relationships and adolescent deviant peer affiliations are related to substance use trajectories. Several studies have examined predictors of substance use trajectories spanning adolescence into young adulthood (Barnes et al., 2000, Barton et al., 2018, Chassin et al., 2002, White et al., 2000), all of which include some level of adjustment, but have not implemented robust causal thinking. Findings from these studies have, thus far, suggested that decreases in parent-adolescent relationship quality and increases in deviant peer affiliations are associated with increases in initial and overall substance use, as well as a faster rate of onset. The application of a causal approach would provide greater insight into the causal impact that parent-adolescent relationship quality and deviant peer affiliations has on the trajectories of substance use.
Thus, the purpose of this study is to explore the process of applying counterfactual thinking to examine the extent to which the quality of the parent-adolescent relationship and early affiliations with deviant peers are causally related to trajectories of alcohol, tobacco, and cannabis use, from adolescence into young adulthood. It is expected that after adjustment for confounding factors, estimated associations will more likely reflect the unbiased causal estimate of the association. From a causal perspective, it is hypothesised that following confounder adjustment: (1) higher quality parent-adolescent relationships in early adolescence will cause lower initial use and delayed patterns of substance use; and, (2) stronger deviant peer affiliations in early adolescence will cause higher initial use and earlier patterns of substance use.
Section snippets
Participants
Participants were from the Australian Temperament Project (ATP), a 16-wave longitudinal study tracking the psychosocial development of young people from infancy to adulthood. The baseline sample consisted of 2443 infants aged between 4 and 8 months, recruited in 1983 from urban and rural areas and representative of the state of Victoria, Australia. Since then, families have been invited to participate via mail surveys approximately every 2 years until 19–20 years and every 4 years thereafter (
Unconditional growth models
Fit indices are presented in the Supplementary Material (Table S2) alongside mean and variance estimates of the growth parameters. The estimated mean growth trajectories are illustrated in Fig. S2. The unconditional LCMs were deemed as having acceptable fit despite some indices being slightly outside the recommended cut-offs.
Conditional growth models
Table 2 presents the estimates from the conditional LCMs for which all growth parameters were regressed onto parent-adolescent relationship quality, adjusting for potential
Discussion
In this multi-wave prospective study, we explored the process of applying counterfactual thinking to examine the extent to which the quality of parent-adolescent relationship and affiliations with deviant peers in early adolescence are causally related to substance use trajectories from adolescence into young adulthood. Attempting to assert causality within an observational design is by no means simple with the untestable possibility of unmeasured confounding remaining a key limitation of
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgements
The ATP study is located at The Royal Children’s Hospital Melbourne and is a collaboration between Deakin University, The University of Melbourne, the Australian Institute of Family Studies, The University of New South Wales, The University of Otago (New Zealand), and the Royal Children’s Hospital (further information available at www.aifs.gov.au/atp). The views expressed in this paper are those of the authors and may not reflect those of their organizational affiliations, nor of other
References (41)
- et al.
Concurrent and prospective associations between substance-specific parenting practices and child cigarette, alcohol, and marijuana use
Journal of Adolescent Health
(2018) - et al.
Premature adolescent autonomy: Parent disengagement and deviant peer process in the amplification of problem behaviour
Journal of Adolescence
(2004) - et al.
A comparison of longitudinal modelling approaches: Alcohol and cannabis use from adolescence to young adulthood
Drug and Alcohol Dependence
(2019) - et al.
Sensitive periods of substance abuse: Early risk for the transition to dependence
Developmental Cognitive Neuroscience
(2017) - et al.
Relationships with parents and adult children’s substance use
Addictive Behaviors
(2017) - et al.
Review of risk and protective factors of substance use and problem use in emerging adulthood
Addictive Behaviors
(2012) - et al.
Family and peer predictors of substance use from early adolescence to early adulthood: An 11-year prospective analysis
Addictive Behaviors
(2012) - et al.
Parental modeling and parenting behavior effects on offspring alcohol and cigarette use: A growth curve analysis
Journal of Substance Abuse
(2000) The determinants of parenting behavior
Journal of Clinical Child Psychology
(1992)- et al.
The inventory of parent and peer attachment: Individual differences and their relationship to psychological well-being in adolescence
Journal of Youth and Adolescence
(1987)
The effects of parenting on the development of adolescent alcohol misuse: A six-wave latent growth model
Journal of Marriage and Family
Trajectory classes of cannabis use and heavy drinking among rural african american adolescents: multilevel predictors of class membership
Addiction (Abingdon, England)
The determinants of parenting: a process model
Child Development
The adolescent brain
Annals of the New York Academy of Sciences
Binge drinking trajectories from adolescence to emerging adulthood in a high-risk sample: Predictors and substance abuse outcomes
Journal of Consulting and Clinical Psychology
The consistency of family and peer influences on the development of substance use in adolescence
Addiction (Abingdon, England)
Prospective childhood predictors of deviant peer affiliations in adolescence
Journal of Child Psychology and Psychiatry and Allied Disciplines
An introduction to instrumental variables for epidemiologists
International Journal of Epidemiology
Causal diagrams for epidemiologic research
Epidemiology (Cambridge, Mass.)
Manual for the social skills rating system
Cited by (3)
Psychosocial predictors of binge-drinking residual harm in adolescence and young adulthood: Findings from the Australian Temperament Project
2021, Drug and Alcohol DependenceCitation Excerpt :Additionally, all young adult associations persisted across young adulthood (3 waves). Binge-drinking remains prevalent throughout Australia, with increases in behaviours observed across adolescence and into young adulthood (Australian Institute of Health and Welfare, 2020; Greenwood et al., 2020). This study used prospective longitudinal data across adolescence and young adulthood to examine predictors of binge-drinking residual harm.