# Interpreting mutual adjustment for multiple indicators of socioeconomic position without committing mutual adjustment fallacies

**Part of the following topical collections:**

## Abstract

Research into the effects of Socioeconomic Position (SEP) on health will sometimes compare effects from multiple, different measures of SEP in “mutually adjusted” regression models. Interpreting each effect estimate from such models equivalently as the “independent” effect of each measure may be misleading, a mutual adjustment (or Table 2) fallacy. We use directed acyclic graphs (DAGs) to explain how interpretation of such models rests on assumptions about the causal relationships between those various SEP measures. We use an example DAG whereby education leads to occupation and both determine income, and explain implications for the interpretation of mutually adjusted coefficients for these three SEP indicators. Under this DAG, the mutually adjusted coefficient for education will represent the direct effect of education, not mediated via occupation or income. The coefficient for occupation represents the direct effect of occupation, not mediated via income, or confounded by education. The coefficient for income represents the effect of income, after adjusting for confounding by education and occupation. Direct comparisons of mutually adjusted coefficients are not comparing like with like. A theoretical understanding of how SEP measures relate to each other can influence conclusions as to which measures of SEP are most important. Additionally, in some situations adjustment for confounding from more distal SEP measures (like education and occupation) may be sufficient to block unmeasured socioeconomic confounding, allowing for greater causal confidence in adjusted effect estimates for more proximal measures of SEP (like income).

## Keywords

Socioeconomic position Education Occupation Income Regression DAGs Causal inference## Abbreviations

- CI
Confidence Interval

- DAG
Directed Acyclic Graph

- OR
Odds Ratio

- SEP
Socioeconomic Position

## Background

Socioeconomic position (SEP) can be defined as the “the social and economic factors that influence what position individuals or groups hold within the structure of a society” [1] and is widely recognised as associated with health [2, 3]. Understanding causal effects of SEP on health can be problematic as SEP is a broad and heterogeneous concept, difficult to operationalise [1, 4], and it may act via a multitude of mechanisms or pathways [2]. Various SEP measures are often used interchangeably despite differences in their theoretical grounding and interpretation, and without regard to the level (e.g. individual, household, area) or life-stage at which they are measured [1, 4, 5, 6, 7, 8, 9]. Nevertheless, researchers sometimes try to differentiate the importance of different SEP indicators. While a range of more advanced methods have developed in recent years that could be employed in this context, we have noticed that many researchers still default to mutually-adjusted regression models of health, which require careful interpretation. Regression models may be seen as ‘good enough’, while more advanced methods are seen as offering diminishing returns to effort. With this in mind, the aim of this article is to highlight potential interpretive pitfalls when mutually adjusting for multiple measures of SEP, explaining best practice and the limitations and assumptions that are implicit with such models. Using directed acyclic graphs (DAGs), we emphasise that causal interpretations of such analyses are based on assumptions about causal relationships between SEP measures.

### Multi-dimensionality of SEP

SEP measures are often correlated but could rarely be treated as proxies for each other [7]. Where SEP is viewed as a confounder for effects of other exposures, this multi-dimensionality may be a minor nuisance: adjustment for a single measure of SEP may leave residual confounding, and best practice would include adjustment for multiple (or all available) measures of SEP [7, 16]. At least this is true as long as SEP is really a confounder and not a mediator of the exposure, if any of the SEP measures could be considered mediators of the exposure effects then adjusting for these could induce collider bias via unobserved SEP [17, 18]. However, when SEP is the causal exposure of interest this multi-dimensionality is more critical, and may even be informative.

Consistent associations with health across a range of SEP measures suggest a relationship between the overall ‘core’ construct of SEP and health. Some even investigate this ‘core’ effect of SEP by aggregating information from multiple SEP measures, e.g. as a latent variable [19], or index counting disadvantages experienced [20, 21].

However, investigation of heterogeneity across multiple measures may help identify which characteristics or resources associated with SEP are most important for health, and which factors could best be intervened on in order to alleviate health inequalities [9, 16, 22, 23]. For example, one study found inequalities in a range of health outcomes were stronger for occupational class when stratifying by education than vice versa, concluding adult occupational class was “a better discriminator of socioeconomic differentials” [24]. Authors of a US study showing clearer associations with mortality for income than for education and occupation argued for the systematic collection of administrative income data for the monitoring and analysis of health inequalities [15]. A Finnish study finding independent associations with smoking for a range of socioeconomic variables concluded that alleviation of inequalities in smoking requires “efforts directed at various dimensions of socioeconomic position” [13]. Numerous studies have compared SEP measures across the lifecourse to infer regarding the lifecourse mechanisms leading to health inequalities [16, 22, 24, 25].

Certainly, more advanced methods are becoming more popular for addressing such issues, with examples including: applications of machine-learning methods to identify the most important predictors of health outcomes [26], path analyses or structural equation models to test hypotheses regarding lifecourse pathways [19, 27], and Bayesian approaches where observed data are used to indicate the most likely causal structure linking a set of variables [28, 29]. Nevertheless, many studies have used regression-based models where SEP measures are mutually adjusted [13, 15, 16, 22, 24, 25, 30, 31], and it is not hard to find even very recent examples e.g. see [32, 33, 34, 35]. We focus therefore on explaining best practice for such models, highlighting potential interpretative pitfalls and their implicit assumptions and limitations.

### The “mutual adjustment” fallacy

Interpretation of mutually adjusted regression models requires care and conceptual clarity (as exemplified in many of the citations given above). It is tempting to interpret the mutually adjusted regression coefficients for each SEP measure as the independent effects of the unique aspects of SEP represented by that measure (i.e. the effects of the areas labelled ‘a*’* in Fig. 1). However, if this interpretation is carried across all the included measures of SEP, then aspects of SEP that are common across measures (i.e. the areas labelled ‘b’ in Fig. 1) have been assumed to have no effect. With only three coefficients and each interpreted as the effect of its respective ‘a’ area, the overlapping information in ‘b’ is not represented by a coefficient. Worryingly, if this area of conceptual overlap represents the ‘core’ of the SEP construct, then this assumes ‘core SEP’ has no effect and that all effects are due to the unique characteristics of the different SEP measures employed.

This can be considered a case of the “Table 2” fallacy, where mutually adjusted coefficients are treated as if they all have an equivalent interpretation [36]. The idea was so named because mutually adjusted coefficients have often been presented in the second results table of a paper. We prefer the term “mutual adjustment” fallacy as this is a little more explicit about the error being committed, which could occur in the first, third, fourth or any other results table.

### Causal interpretation of a mutually-adjusted model

*n*= 1248) [38]. Data were coded to indicate low education (left school at age 16 or earlier), manual (compared to non-manual) occupations [39], and low income (lowest tertile of equivalised household income). Self-assessments of health as poor (vs. excellent through fair) were taken as the outcome (5.9% reported poor health). The first column of the table shows unadjusted odds ratios (ORs) and 95% confidence intervals (95% CI) for the association between each measure of SEP and self-assessed health. Other columns show models with adjustment for occupation and education only, and with mutual adjustment for all three measures of SEP. As the outcome is rare we will ignore issues related to the non-collapsibility of ORs [40].

Illustrative example of mutual adjustment

Unadjusted ORs for poor health (95% CI) | Interpretation under Fig. 2a | Partially Adjusted ORs for poor health | Interpretation under Fig. 2a | Mutually Adjusted ORs for poor health | Interpretation under Fig. 2a | |
---|---|---|---|---|---|---|

Low Education | 2.43 (1.32–4.47) | Total effect of education | 1.64 (0.84–3.19) | Direct effect of education, not mediated via occupation | 1.44 (0.72–2.85) | Direct effect of education, not mediated via occupation and income. |

Manual Occupation | 2.79 (1.73–4.48) | Total effect of occupation, confounded by education | 2.33 (1.39–3.93) | Total effect of occupation, not confounded by education | 1.99 (1.15–3.43) | Direct effect of occupation, not mediated via income, or confounded by education. |

Low Income | 2.48 (1.55–3.98) | Total effect of income, confounded by education and occupation | – | 1.68 (0.99–2.85) | Total effect of income, not confounded by occupation and education. |

Under Fig. 2a, the mutually adjusted OR for income represents the total effect of income on health, after adjusting for confounding by occupation and education [36]. However, the mutually adjusted OR for occupation does not have the same interpretation; it is not a total effect of occupation, but rather a direct effect of occupation, i.e. the portion of its total effect on health which is not mediated via income, after adjusting for confounding by education [36]. Similarly, the mutually adjusted OR for education represents a direct effect, not mediated via occupation or income [36]. Thus, a comparison of the mutually adjusted ORs for education and income is not a like-for-like comparison, but a comparison of a total effect (for income) with a direct effect (for education). If the intent was to compare the total effects of education and income then this would be an incorrect comparison: a mutual adjustment fallacy.

Given Fig. 2a, a question about the relative magnitude of the total effects of these SEP measures could be answered by comparing the mutually adjusted OR for income (1.68), with the partially adjusted OR for occupation (2.33), and the unadjusted OR for education (2.43), i.e. adjusting for confounders but not mediators. This highlights quite a different pattern in the magnitude of effects as would be obtained by straight comparison of the mutually adjusted (or the unadjusted) ORs.

### Competing assumptions

Figure 2a is of course not the only set of assumptions that could be made about the relationships between these variables, and analyses may be sensitive to different assumptions [23]. Figure 2b shows an alternative set of (less plausible) causal assumptions to those shown in Fig. 2a. In this case, education is now considered the more proximate and income the more distal cause. Under Fig. 2b, a question about the relative magnitude of the total effects of education and income would require comparison of the mutually adjusted OR for education with the unadjusted OR for income (i.e. 1.44 and 2.48 respectively in Table 1). Thus, different assumptions about the causal relationships between socioeconomic variables can lead to different conclusions about which are more important.

If Fig. 2a and b were both equally plausible, one might want to compare total effect sizes under competing causal assumptions: e.g. the total effect of education with income and occupation viewed as confounders, vs. the total effect of income with education and occupation viewed as confounders. In this case, a comparison of the mutually adjusted ORs for education and income would be appropriate. However, you would only want to do this in situations with multiple plausible competing causal assumptions. In such cases, interpretation could be facilitated by clear description of the competing assumptions, explicitly linking the reported results to those assumptions. We consider ‘mutual adjustment’ to be poor short-hand for such a nuanced interpretation.

Further, we would suggest the causal ordering in Fig. 2a rather than Fig. 2b is more plausible and more consistent with theory regarding the development of SEP over the lifecourse, with education usually preceding and determining occupational level, which in turn generates income. [10, 41] However, while we would favour interpretation under Fig. 2a over interpretation under Fig. 2b, it would be poor practice to assume all other researchers share our assumptions. Reporting the findings in Table 1 as mutually adjusted without explicitly linking them to the assumptions in Fig. 2a could lead to misinterpretation. A naïve reader would have no guidance as to which effects should be interpreted as total and which as direct effects, or a reader who strongly subscribed to Fig. 2b might interpret the results according to those assumptions. Phrases such as “the independent effect of occupation” do not specify whether the “independence” is from confounders or mediators and are thus less informative than they could be with further explication.

### Unmeasured confounding

However, under Fig. 3b the estimate of the total effect of income from the mutually adjusted model in Table 1 remains unbiased. The confounding influence of unmeasured parental SEP on the effect of income is sufficiently blocked by adjustment for own education and occupation, though the direct effects of occupation and education are still subject to bias from the confounding influence of unmeasured parental SEP and from conditioning on a collider [17, 18]. Thus, if Fig. 3b were at all plausible relative to Fig. 3a we might have more confidence in the total effect estimate for income from Table 1 than in the direct effect estimates for occupation and education. This total effect estimate requires less stringent assumptions to be valid (i.e. Figure 3b) than those for the direct effect estimates (i.e. Figure 2 with no unmeasured confounding).

Logistic regression predicting low income from parental occupation, with and without adjustment for own education and occupation

Unadjusted OR for Low Income (95% CI) | Adjusted OR for Low Income (95% CI) | |
---|---|---|

Manual Parental Occupation | 2.80 (2.08–3.77) | 1.28 (0.91–1.80) |

Low Education | – | 4.66 (3.08–7.05) |

Manual Occupation | – | 3.83 (2.89–5.08) |

## Discussion

We focused here on the relatively simple method of mutually adjusted regression, without tackling issues such as interactions between SEP measures, effect heterogeneity, measurement error or statistical mis-specification [23]. As mutually adjusted regression is still commonly used in epidemiology, we hope this discussion will aid best practice, highlight implicit assumptions and limitations, and provide an introductory step towards the ever-developing literature on more advanced causal inference methods, which might include Bayesian analyses [28, 29] or more precisely decomposed definitions of direct and indirect effects that can deal appropriately with interactions between SEP measures [42, 43, 44, 45, 46].

We examined four possible causal diagrams linking three SEP measures, but many more are possible [23], and interpretations of analyses should be altered under differing causal assumptions, as illustrated here. Indeed, the causal structure linking SEP indicators to each other and to health may vary between different social contexts, welfare regimes etc. [28], as may the degree of heterogeneity both within and between socioeconomic categories [9], so researchers may be unsure of what causal model to interpret their data under. Causal direction may also be ambiguous where there are feedback loops (e.g. where a better job leads to a higher income, which then leads to an even better job and so forth). In some cases, such longitudinal ambiguity may be resolved by including repeated measures from different life-stages and interpreting analyses in light of the causal links between measures over time [47], but researchers will not always have the luxury of such detailed lifecourse data. Where there is ambiguity over the causal structure linking socioeconomic variables, Bayesian methods offer a promising avenue for empirically determining the most likely structure given the data [28, 29]. Indeed, such exploratory research is important for developing a good understanding of the causal relationship between socioeconomic variables. Nevertheless, data-driven approaches such as these may be sensitive to idiosyncrasies of the data under study, and/or there may be multiple causal diagrams that would fit the data more or less equally well. Thus, it may still be worthwhile to assess how sensitive analytical conclusions are to alternative theoretically informed causal diagrams.

We highlight the importance of a full causal diagram that includes all relevant common causes of included variables, which might include distal and unmeasured aspects of SEP [23]. Given difficulties associated with gathering inter-generational socioeconomic data, our illustration of residual confounding from parental SEP is a pertinent one. It demonstrates how, in some instances, adjustment for some socioeconomic confounders may be sufficient to block confounding from other more distal and unmeasured aspects of socioeconomic background, and give greater confidence that effect estimates from a proximal measure of SEP are causal. On the other hand, we have shown how estimates of direct effects for more distal measures of SEP can be biased by residual unmeasured confounding [35], and/or collider bias [17]. Where there is the likelihood of residual unmeasured confounding (i.e. almost always) it would seem advisable to test sensitivity of conclusions to different strengths of unmeasured confounding [48].

Estimating “effects” of SEP measures presupposes intention to intervene on and manipulate SEP [23, 46]. Assuming Fig. 2a to be the more plausible set of assumptions regarding the causal relationships between education, occupation and income (which may not hold in all contexts [28]), it follows that we could have greater causal confidence in published estimates of the effects of income that are adjusted for education and occupation (or other more distal socioeconomic variables), than we might have in published estimates of the effects of occupation or education (whether or not they are adjusted for other socioeconomic measures). This is fortuitous, as of the three, it could be argued that income would also be the simplest to intervene on [49], e.g. through welfare or taxation policies. Estimates of effects for education or occupation that are adjusted for income are not without their merit however, and, assuming no unmeasured confounding, could be interpreted for example as the remaining effect of occupation, if we were to somehow intervene and equalise income across occupational strata [46].

## Conclusion

The principles set forth here are not limited to applications involving education, occupation and income but will be applicable whenever researchers are using mutual regression adjustment to compare the effects of different SEP measures, including measures taken at different levels or life-stages. We have explained how effects for multiple socioeconomic measures should not be interpreted equivalently as the “independent” effect of that measure, ignoring the effects of the common information they provide. Causal effects of more distal SEP measures, will be biased with adjustment for proximate SEP measures. A good causal understanding of relationships between socioeconomic variables can improve interpretation and lead to different conclusions about which measure is most important for health than naïve comparisons of “independent” effects. Such understanding is therefore foundational to understanding the causal processes by which SEP influences health over the lifecourse, and to identifying effective points of intervention [23].

## Notes

### Acknowledgements

Not applicable.

### Funding

MG and FP are funded by the Medical Research Council (MC_UU_12017/13) and the Scottish Government Chief Scientist Office (SPHSU13). Funders had no active role in the design of the study, collection, analysis or interpretation of data, or in writing the manuscript.

### Availability of data and materials

The datasets generated and analysed during the current study are available for use by application. Details of how to apply can be found on the study website: http://2007study.sphsu.mrc.ac.uk/Information-on-data-sharing.html

### Authors’ contributions

MG had the initial idea, performed all analyses and drafted the manuscript. FP was a major contributor in writing the manuscript. All authors read and approved the final manuscript.

### Ethics approval and consent to participate

Ethical approval was gained for each wave of the Twenty-07 Study from the National Health Service and/or Glasgow University Ethics Committees. Participants gave informed consent. We had access to the data as researchers at the MRC/CSO Social and Public Health Sciences Unit, where the study is based.

### Consent for publication

Not applicable.

### Competing interest

The authors declare that they have no competing interests.

### Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

## References

- 1.Galobardes B, Shaw M, Lawlor DA, Lynch JW, Davey Smith G. Indicators of socioeconomic position (part 1). J Epidemiol Community Health. 2006;60(1):7–12.CrossRefGoogle Scholar
- 2.Link BG, Phelan J. Social conditions as fundamental causes of disease. J Health Soc Behav. 1995;35:80–94.CrossRefGoogle Scholar
- 3.Department of Health and Social Security. equalities in Health: Report of a Working Group Chaired by Sir Douglas Black. London; 1980.Google Scholar
- 4.Liberatos P, Link BG. The measurement of social class in epidemiology. Epidemiol Rev. 1988;10:87–121.CrossRefGoogle Scholar
- 5.Krieger N, Williams DR, Moss NE. Measuring social class in US public Health Research: concepts, methodologies, and guidelines. Annu Rev Public Health. 1997;18(1):341.CrossRefGoogle Scholar
- 6.Macintyre S, McKay L, Der G, Hiscock R. Socio-economic position and health: what you observe depends on how you measure it. J Public Health Med. 2003;25(4):288–94.CrossRefGoogle Scholar
- 7.Braveman PA, Cubbin C, Egerter S, Chideya S, Marchi KS, Metzler M, Posner S. Socioeconomic status in Health Research: one size does not fit all. JAMA. 2005;294(22):2879–88.CrossRefGoogle Scholar
- 8.Geyer S, Hemström Ö, Peter R, Vågerö D. Education, income, and occupational class cannot be used interchangeably in social epidemiology. Empirical evidence against a common practice. J Epidemiol Community Health. 2006;60(9):804–10.CrossRefGoogle Scholar
- 9.McCartney G, Bartley M, Dundas R, Katikireddi SV, Mitchell R, Popham F, Walsh D, Wami W. Theorising social class and its application to the study of health inequalities. SSM - Population Health. 2018. https://doi.org/10.1016/j.ssmph.2018.10.015.
- 10.Lynch J, Kaplan G. Socioeconomic Position. In: Berkman LF, Kawachi I, editors. Social Epidemiology. Oxford: Oxford University Press; 2000.Google Scholar
- 11.Bartley M. Health inequality: an introduction to theories, concepts and methods. Cambridge: Polity Press; 2004.Google Scholar
- 12.Townsend P. Deprivation. Journal of Social Policy. 1987;16(02):125–46.CrossRefGoogle Scholar
- 13.Laaksonen M, Rahkonen O, Karvonen S, Lahelma E. Socioeconomic status and smoking. The European Journal of Public Health. 2005;15(3):262–9.CrossRefGoogle Scholar
- 14.Mirowsky J, Ross CE. Education, personal control, lifestyle and health. Research on Aging. 1998;20(4):415–49.CrossRefGoogle Scholar
- 15.Duncan GJ, Daly MC, McDonough P, Williams DR. Optimal indicators of socioeconomic status for Health Research. Am J Public Health. 2002;92(7):1151–7.CrossRefGoogle Scholar
- 16.Harper S, Lynch J, Hsu W-L, Everson SA, Hillemeier MM, Raghunathan TE, Salonen JT, Kaplan GA. Life course socioeconomic conditions and adult psychosocial functioning. Int J Epidemiol. 2002;31(2):395–403.CrossRefGoogle Scholar
- 17.Cole SR, Platt RW, Schisterman EF, Chu H, Westreich D, Richardson D, Poole C. Illustrating bias due to conditioning on a collider. Int J Epidemiol. 2010;39(2):417–20.CrossRefGoogle Scholar
- 18.Richiardi L, Bellocco R, Zugna D. Mediation analysis in epidemiology: methods, interpretation and bias. Int J Epidemiol. 2013;42(5):1511–9.CrossRefGoogle Scholar
- 19.Warren JR. Socioeconomic status and health across the life course: a test of the social causation and health selection hypotheses. Social forces; a scientific medium of social study and interpretation. 2009;87(4):2125–53.PubMedPubMedCentralGoogle Scholar
- 20.Degerud E, Ariansen I, Ystrom E, Graff-Iversen S, Høiseth G, Mørland J, Davey Smith G, Næss Ø. Life course socioeconomic position, alcohol drinking patterns in midlife, and cardiovascular mortality: analysis of Norwegian population-based health surveys. PLoS Med. 2018;15(1):e1002476.CrossRefGoogle Scholar
- 21.Whitley E, Benzeval M, Popham F. Associations of successful aging with socioeconomic position across the life-course: the West of Scotland Twenty-07 prospective cohort study. Journal of Aging and Health. 2018;30(1):52–74.CrossRefGoogle Scholar
- 22.Gilman SE, Abrams DB, Buka SL. Socioeconomic status over the life course and stages of cigarette use: initiation, regular use, and cessation. Journal of Epidemiology and Community Health. 2003;57(10):802–8.CrossRefGoogle Scholar
- 23.Kaufman JS, Kaufman S. Assessment of structured socioeconomic effects on health. Epidemiology. 2001;12(2):157–67.CrossRefGoogle Scholar
- 24.Davey Smith G, Hart C, Hole D, MacKinnon P, Gillis C, Watt G, Blane D, Hawthorne V. Education and occupational social class: which is the more important indicator of mortality risk? J Epidemiol Community Health. 1998;52:153–60.CrossRefGoogle Scholar
- 25.Davey Smith G, Hart C, Blane D, Gillis C, Hawthorne V. Lifetime socioeconomic position and mortality: prospective observational study. BMJ. 1997;314:547–52.CrossRefGoogle Scholar
- 26.Koroukian SM, Schiltz N, Warner DF, Sun J, Bakaki PM, Smyth KA, Stange KC, Given CW. Combinations of chronic conditions, functional limitations, and geriatric syndromes that predict health outcomes. J Gen Intern Med. 2016;31(6):630–7.CrossRefGoogle Scholar
- 27.Sweeting H, Green MJ, Benzeval M, West P. The emergence of health inequalities in early adulthood: evidence on timing and mechanisms from a West of Scotland cohort. BMC Public Health. 2016;16:41.CrossRefGoogle Scholar
- 28.Alvarez-Galvez J. Discovering complex interrelationships between socioeconomic status and health in Europe: a case study applying Bayesian networks. Soc Sci Res. 2016;56:133–43.CrossRefGoogle Scholar
- 29.Lewis FI, McCormick BJJ. Revealing the complexity of health determinants in resource-poor settings. Am J Epidemiol. 2012;176(11):1051–9.CrossRefGoogle Scholar
- 30.Katikireddi SV, Niedzwiedz CL, Popham F. Employment status and income as potential mediators of educational inequalities in population mental health. Eur J Pub Health. 2016;26(5):814–6.CrossRefGoogle Scholar
- 31.Ellaway A, Green MJ, Benzeval M, Leyland AH, Macintyre S. Getting "sicker quicker": does living in a more deprived neighbourhood mean your health deteriorates faster? Health & Place. 2012;18:132–7.CrossRefGoogle Scholar
- 32.Cadar D, Lassale C, Davies H, Llewellyn DJ, Batty G, Steptoe A. Individual and area-based socioeconomic factors associated with dementia incidence in England: evidence from a 12-year follow-up in the english longitudinal study of ageing. JAMA Psychiatry. 2018;75(7):723–32.CrossRefGoogle Scholar
- 33.Demakakos P, Biddulph JP, de Oliveira C, Tsakos G, Marmot MG. Subjective social status and mortality: the English longitudinal study of ageing. Eur J Epidemiol. 2018;33(8):729–39.CrossRefGoogle Scholar
- 34.Canney M, Leahy S, Scarlett S, Kenny RA, Little MA, O’Seaghdha CM, McCrory C. Kidney disease in women is associated with disadvantaged childhood socioeconomic position. Am J Nephrol. 2018;47(5):292–9.CrossRefGoogle Scholar
- 35.Kröger H, Hoffmann R, Tarkiainen L, Martikainen P. Comparing observed and unobserved components of childhood: evidence from Finnish register data on midlife mortality from siblings and their parents. Demography. 2018;55(1):295–318.CrossRefGoogle Scholar
- 36.Westreich D, Greenland S. The table 2 fallacy: presenting and interpreting confounder and modifier coefficients. Am J Epidemiol. 2013;177(4):292–8.CrossRefGoogle Scholar
- 37.Shrier I, Platt RW. Reducing bias through directed acyclic graphs. BMC Med Res Methodol. 2008;8:70.CrossRefGoogle Scholar
- 38.Benzeval M, Der G, Ellaway A, Hunt K, Sweeting H, West P, Macintyre S. Cohort profile: West of Scotland Twenty-07 study - health in the community. Int J Epidemiol. 2009;38(5):1215–23.CrossRefGoogle Scholar
- 39.Office of Population Censuses and Surveys. Classification of occupations. London: HMSO; 1980.Google Scholar
- 40.Greenland S, Robins JM, Pearl J. Confounding and collapsibility in causal inference. Stat Sci. 1999;14:29–46.CrossRefGoogle Scholar
- 41.Graham H. Unequal lives: health and socioeconomic inequalities. Maidenhead: Open University Press; 2007.Google Scholar
- 42.VanderWeele TJ. Mediation analysis: a Practitioner's guide. Annu Rev Public Health. 2016;37:17–32.CrossRefGoogle Scholar
- 43.VanderWeele TJ. Marginal structural models for the estimation of direct and indirect effects. Epidemiology. 2009;20(1):18–26.CrossRefGoogle Scholar
- 44.Daniel RM, De Stavola BL, Cousens SN, Vansteelandt S. Causal mediation analysis with multiple mediators. Biometrics. 2015;71(1):1–14.CrossRefGoogle Scholar
- 45.Imai K, Keele L, Tingley D. A general approach to causal mediation analysis. Psychol Methods. 2010;15(4):309–34.CrossRefGoogle Scholar
- 46.VanderWeele TJ, Robinson WR. On causal interpretation of race in regressions adjusting for confounding and mediating variables. Epidemiology. 2014;24(4):473–84.CrossRefGoogle Scholar
- 47.Green MJ, Popham F. Life course models: improving interpretation by consideration of total effects. Int J Epidemiol. 2017:
**46**(3):1057-1062.Google Scholar - 48.Carnegie NB, Harada M, Hill JL. Assessing sensitivity to unmeasured confounding using a simulated potential confounder. Journal of Research on Educational Effectiveness. 2016;9(3):395–420.CrossRefGoogle Scholar
- 49.Duncan GJ, Yeung WJ, Brooks-Gunn J, Smith JR. How much does childhood poverty affect the life chances of children? Am Sociol Rev. 1998;63(3):406–23.CrossRefGoogle Scholar

## Copyright information

**Open Access**This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.