Causality in Health Inequalities Research: New Solutions to Old Challenges 2? |
|
| Session Organisers |
Mrs Katharina Loter (Martin Luther University Halle-Wittenberg) Professor Oliver Arránz Becker (Martin Luther University Halle-Wittenberg) |
| Time | Wednesday 17th July, 11:00 - 12:30 |
| Room | D31 |
Questions that motivate most studies on the emergence and reproduction of health inequalities over the life course are longitudinal and causal in nature. Fortunately, after several decades of socio-epidemiological research that produced numerous cross-sectional (and thus associational) findings, the need for studies aiming at disentangling causal effects and mechanisms from non-causal associations is becoming increasingly recognized. Especially, attempts to estimate causal effects of social background variables (e.g., marital and family status or occupational status) and/ or life changing events (e.g., pregnancy, separation and divorce, death of a family member, retirement) on health over time – including selection and reverse causality – have to be undertaken using advanced data structures and analytical techniques suited for causal inference. Following recent developments in causal theory including directed acyclic graphs (DAG), new opportunities for innovative health research are being generated by the increasing availability of rich panel data with multiple times of observation and different types of health measures (biomarkers, fitness tracker data, results of objective medical tests, and a variety of subjective indicators). Nonetheless, causal analyses on health inequalities over time remain challenging for many reasons, and one of them is health being a strong predictor of attrition in panel studies (due to nonresponse or death). Several analytical approaches have been proposed as a remedy for the manifold challenges regarding causal inference in health research. The first one deals with health-related self-selection as a missing data problem and uses techniques for correcting causal effect estimates such as weighting or dropout models for nonignorable nonresponse (e.g., Heckman selection models). The second one focuses on reducing unobserved heterogeneity (e.g., via instrumental variables, propensity score matching, fixed-effects regression). The third one places emphasis on testing potential intervening mechanisms (mediator and suppressor effects). The fourth one corresponds with the use of agent-based modeling (e.g., with SIENA) enabling simultaneous estimation of selection and causation effects in a social network framework. Finally, the consideration of measurement error while dealing with latent variables may serve to adjust causal effect estimates for (lack of) reliability. In the light of the above challenges, we invite submissions applying any of the mentioned approaches (or other innovative analysis techniques) with the goal to improve understanding of causality and selection in (preferably longitudinal) research on health inequalities. Welcome are contributions that aim at unraveling the causal complexity between social background variables and/ or life course transitions and events, and any kind of objective or subjective health outcome.
Keywords: health inequalities, life course, causal inference, selection, panel data analysis
Mr Alessandro Procopio (University of Luxembourg) - Presenting Author
Professor Robin Samuel (University of Luxembourg)
In our contribution, we assess the possibilities and limits of Cox regression models in social stratification research in the area of health. We are motivated by the need for a structured analytical strategy through which researchers can deal with health inequality. Previous findings suggest considering health as a relevant resource but also one, which is unequally distributed among the members of a population. Along these lines, we focus on the inequality of risks distribution and the social stratification of (non) access to health as a resource.
Using the substantive example of health inequality, we perform five Monte Carlo simulations in constructed longitudinal data. Each setting simulates a different source of bias. Specifically: a) Measurement error (misspecification of time measurement); b) Linear dependency between class of origin, destination and mobility effects; c) Omitted variables bias; d) Disentangle of timing/probability effects, namely speed/overall occurrence likelihood of an event; and e) Unobserved heterogeneity among groups.
The health-related risks approach in analysing health inequalities has a twofold advantage: a) it splits the health outcome in a true differential and in a stochastic component due to chance and b) it considers only the first – and in most cases more interesting part – as a source of inequality. Moreover, Cox regression models allow for a flexible parameterization conditional to the specific research settings. For instance, addition of frailty parameters to the regression equation can help social scientists to reduce unobserved heterogeneity. This problem is especially encountered in social stratification research when comparing logit transition probabilities.
In summary, this study contributes to the current literature by demonstrating the flexibility of Cox regression models in social stratification research in the area of health. It further provides valuable analytic avenues for theory-driven empirical research in social scientific health research as it uncovers how various sources of bias affect estimates.
Ms Julia Waldhauer (Robert Koch Institute) - Presenting Author
Dr Jens Hoebel (Robert Koch Institute)
Dr Lars Kroll (Robert Koch Institute)
Dr Thomas Lampert (Robert Koch Institute)
Although there is isolated evidence of converging health inequalities from childhood to adolescence, the impact of social origin on health never disappears completely. It remains unclear, however, to what extent different measures (family-based and non-family) of social differentiation are interrelated and influence mental health in the course from childhood to adolescence. We investigated the importance of different aspects of social differentiation as causes of mental health problems in adolescence.
Prospective data of 1,178 boys and 1,286 girls were derived from the German KiGGS cohort. Participants were aged 7–13 at baseline (2009-20011) and 11–17 at follow-up (2014-2017). Mental health problems are assessed using the Strengths and Difficulties Questionnaire (SDQ) in childhood and adolescence. Family socio-economic status (SES) and area deprivation (DEP) is measured in childhood, type of school attended (SCO) and subjective social status (SSS) in adolescence. We used directed acyclic graph (DAG) to construct a structural causal model (SCM) of the complex relationships between these measures of social differentiation and SDQ from childhood to adolescence.
The total effects on SDQ in adolescence are highest for SDQ in childhood (0.55 [0.50; 0.62]) followed by childhood family SES (-0.16 [-0.23; -0.09]). SCO (-0.08 [-0.13; -0.03]) and SSS (-0.09 [-0.14; -0.03]) show similar effects. Gender differences are minor, but boys show a direct effect of SCO (-0.10 [-0.17; -0.04]) on SDQ in adolescence, while girls show an effect of SDQ in childhood on the later assessment of SSS (-0.11[-0.22; -0.01]), which influences SDQ in adolescence as well.
In contrast to conventional regression models, the relation of different measures of social differentiation can be analyzed using SCM. The results indicate simultaneous effects of causation and selection if the underlying assumptions of the DAG are valid.
Mr Manuel Holz (Chemnitz University of Technology) - Presenting Author
The aim of the study is to compare health outcomes of migrants and the native German population, testing if there is a Healthy Immigrant Effect (HIE). The HIE is marked by an observed health advantage for migrants, when compared to the host population, which declines with the years since migration. In order to reduce unobserved heterogeneity propensity score matching and a more adequate native comparison group is used. According to selection perpectives, migrants do not represent a random sample of the country of origins. Therefore, comparing migrants to a random sample of the host population will yield biased estimates. This problem is adressed by using a native comparison group that has more similar characteristics, in this case native internal movers. Physical and mental health scales are computed from the Short Form 12 Item Health survey from the 2016 wave of the German Socio-Economic Panel. The study contributes to ongoing discussions on health selection perspectives by calculating average health differences between recent migrants (≤ 10 years since migration), non-recent migrants (> 10 years since migration), native Germans and native internal movers via propensity score matching.
Dr Erwin Stolz (Medical University of Graz) - Presenting Author
Dr Hannes Mayerl (Medical University of Graz)
Professor Éva Rásky (Medical University of Graz)
Professor Wolfgang Freidl (Medical University of Graz)
Background: Frailty constitutes an important risk factor for adverse outcomes among older adults. In longitudinal studies on frailty, selective sample attrition may threaten the validity of results.
Objective: To assess the impact of sample attrition on frailty index trajectories and gaps related to socio-economic status (education) therein among older adults in Europe. Methods: A total of 64,143 observations from 21,044 respondents (50+) from the Survey of Health, Ageing and Retirement in Europe across 12 years of follow-up (2004–2015) and subject to substantial sample attrition (59%) were analysed. We compared results of a standard linear mixed model assuming missing at random (MAR) sample attrition with a joint model assuming missing not at random sample attrition.
Results: Estimated frailty trajectories of both the mixed and joint models were identical up to an age of 80 years, above which modest underestimation occurred when a standard linear mixed model was used rather than a joint model. The latter effect was larger for men than women. Substantial education-based inequality in frailty continued throughout old age in both the mixed and joint models.
Conclusion: Linear mixed models assuming MAR sample attrition provided good estimates of frailty trajectories up until high age. Thus, the validity of existing studies estimating frailty trajectories based on standard linear mixed models seems not threatened by substantial sample attrition.