Jump to comment:

• Attributing causality to observed associations between psychosocial exposures and health: a response
  John Macleod
  Published on: 27 April 2016
• Confusing confounding and causation
  A Singh-Manoux
  Published on: 27 April 2016
• How does social class get under the skin?
  Hans Bosma
  Published on: 27 April 2016

• Published on: 27 April 2016

  Attributing causality to observed associations between psychosocial exposures and health: a response

  • John Macleod, Clinical Research Fellow
  • Other Contributors:
    • George Davey Smith, Pauline Heslop, Chris Metcalfe, Doug Carroll, Carole Hart

  Dear Editor

  We are glad that our paper has generated interest amongst readers of the Journal of Epidemiology and Community Health.

  We would like to respond first to the points made by Drs Singh-Manoux and Clarke. The first relates to the importance of theory and the need for a conceptual framework within which to consider how social disadvantage comes to be associated with poorer health (given that...

  Show More

  Dear Editor

  We are glad that our paper has generated interest amongst readers of the Journal of Epidemiology and Community Health.

  We would like to respond first to the points made by Drs Singh-Manoux and Clarke. The first relates to the importance of theory and the need for a conceptual framework within which to consider how social disadvantage comes to be associated with poorer health (given that this association seems secure). We completely agree that this issue is important and have discussed it on several occasions. [1,2,3] We have no particular ideological attachment to any explanatory model. We acknowledge that the plausibility of a role for psychosocial factors is well established. [4] Given the importance of the question we feel that it is important that evidence relating to it is appraised critically.

  Singh-Manoux and Clarke’s second point relates to the most appropriate statistical approach to prospective observational data relating psychosocial and other “risk-factors” to health. They suggest structural equation modelling. Interestingly this approach has not so far been used in any of the major papers from their own study (Whitehall II) or indeed in any of the publications we critique in our own study. Had we used structural equation modelling with our own data, direct comparison of our findings with those of others would have been difficult.

  The theoretical advantages offered by structural equation modelling depend largely on data being available on all the key variables, measured or latent, in the hypothesised causal system. Often they aren’t. To proceed with a structural equation modelling approach in this situation may, we believe, generate misleading results. We appreciate that our study hypotheses only address one aspect of a wider model of the relationship between socioeconomic position and health. The dataset used, while invaluable in that it describes a cohort with an unusual social patterning of stress, unfortunately does not contain all the measures that would be necessary for a full evaluation of that wider model. Structural equation modeling may be capable of simultaneously modeling the material, psychosocial, and behavioural pathways that make up the wider model, but not if the data is unavailable.

  It has been proposed that stress provides one of the mechanisms by which socioeconomic health variations arise. This can be represented as “Model A”

  Deprivation -> Psychological stress -> Disease

  Singh-Manoux and Clarke argue that our regression equations ignore the directed relationship between Deprivation and Psychological stress and so the estimated coefficients for these covariates will be biased. They state that this is because the values of one or both of these variables will be correlated with the error term [5 and personal communication]. It is not stated how this correlation comes about, but it may well not have an important effect. Directional relationships between regression covariates are commonly encountered when modeling longitudinal data where the covariates are repeated measurements of a risk factor. To our knowledge, the modeling of their effects within a multiple regression model is not seen as problematic [6].

  In any case, Singh-Manoux and Clarke’s concerns may be misplaced. We propose that the relationship between stress and disease is based on confounding by deprivation. That is, there is no direct relationship between stress and disease (this can be represented as “Model B”)

  Psychological stress <_- deprivation="deprivation" _-="_-"/> Disease.

  In which case the comparison of two regression equations:

  [i] Disease regressed upon stress
  [ii] Disease regressed upon stress and deprivation
  is the standard method of evaluating the role of confounding. Movement of the stress coefficient towards the null value in equation [ii] is observed in our study and is taken as evidence consistent with a confounding role for deprivation. Given the strong suggestion in our data that Model B is the more important (the *positive* effect of stress on disease is likely to be due to confounding by deprivation) we are confident that the standard method of evaluating confounding is appropriate. The chosen statistical methods are the appropriate ones to use in that they address the study hypotheses using the available data without introducing more complexity than can be accommodated.

  Hans Bosma and Martin Bobak criticise what they take as our suggestion that psychosocial factors do not matter, and can effectively be ignored in attempts to explain and ameliorate social inequalities in health. This was not the message we wished to convey in our paper. Rather we suggested - and supported our suggestion with empirical evidence – that much of the evidence purportedly supporting an important causal role for psychosocial factors in this regard may be flawed. Acknowledgement of this issue, and attempts to address it in future work are, we believe, essential if progress is to be made in relation to these important questions.

  We now turn to some of their specific points. They state that “careful examination” of our data reveal no relation between stress and mortality “before or after adjustment for social class”. On this basis they suggest that the issue of confounding is irrelevant in this context. Yet inspection of table 4 in our paper shows that high (compared to low) stress was associated with an age adjusted relative risk of mortality from smoking related cancers of 0.64 (95% CI 0.42 – 0.96) that was attenuated to 0.71 (95% CI 0.47-1.08) after adjustment for occupational class. Age adjusted relative risk for all cause mortality in medium (compared to low) stress participants was 0.85 (95% CI 0.76-0.94), attenuated to 0.93 (95% CI 0.83-1.04) after adjustment for occupational class. Conventionally these data are best described as showing small positive effects of stress, which are attenuated on adjustment for social class. Bosma and Bobak appear to acknowledge this in their first paragraph. As discussed above – conventionally – this suggests the presence of confounding.

  Our point was simple; in this population, psychosocial adversity (as indexed by perceived stress) was associated with general social advantage and thus confounding generated a protective association between stress and health. That this was due to confounding was supported by the implausibility of stress being protective. However in populations where psychosocial adversity is associated with general social disadvantage (such as the populations where Bosma and Bobak have reported negative effects of psychosocial exposure) such confounding will tend to generate a negative association between psychosocial factors and health and its confounded nature may be obscured by the plausibility of the association. Bosma and Bobak do not indicate how they can be confident that this criticism is irrelevant to associations they have described between psychosocial adversity and health.

  Instead, they criticise our use of the Reeder Stress Inventory and speculate that if we had used a more modern and “specific” measure of psychosocial adversity we would have seen effects on mortality. We used the measure of stress that was available to us; a measure that, objectively, appears as valid and reliable as other contemporary psychometric instruments and was widely used during the historical period when this study was undertaken. We accept that the measure is not perfect and that, in common with other instruments measuring self-perceived stress (Cohen’s Perceived Stress Scale for example), it is likely to be influenced by reporting bias. A more objective approach might be to use a life-event scale. However this approach also has problems. First it ignores the dimension of how different individuals cope with the same life event and second, life-event scores can also be confounded by social position (certain negative life events may be associated with social disadvantage). The important issue here is that the points made in our paper are largely independent of arguments concerning the status of the RSI amongst psychometric instruments.

  The Kaunus-Rotterdam Intervention Study (KRIS) also incorporated the RSI and also found no effect of stress on mortality. [7] Interestingly these findings were not published in an accessible form (would they have remained unpublished if they had been positive?), just one indication of how publication bias bedevils this area. [8] The social patterning of stress in KRIS was not reported. “Effects” of one item of the RSI on short -term mortality in KRIS were reported (“at the end of the day I feel completely exhausted physically and mentally”). [9] In this general population study a positive response on the above item predicted death from ischaemic heart disease amongst those apparently free of IHD at baseline up to 40 months of follow-up. In our occupational based study, 30 subjects apparently free of IHD at baseline had died within the first 40 months of follow-up. None of these reported “vital exhaustion”. Prevalence of this symptom was higher in KRIS. 7% of KRIS subjects reported vital exhaustion at baseline compared with 2% in our study Again, in our study, vital exhaustion was associated with social advantage. 3.1% of those in social class I reported vital exhaustion compared to 1.7% in social class V. At 5 years follow up the relative risk of death from IHD associated with vital exhaustion. (adjusted for age and occupational class) in our study was 1.14 (0.15-8.48); at 21 years follow-up it was 1.09 (0.56-2.12).

  Our discrepant (compared to KRIS) finding could have several explanations. We think the most likely is that in our relatively healthy, working cohort the item was not as strongly related to existing disease (including sub-clinical disease) as it would be in a general population sample. This explanation would parallel what is observed in the studies of circulating cholesterol and mortality. In general population studies – where low cholesterol is more strongly associated with existing disease – there is generally an inverse association between cholesterol and non-CHD mortality; in healthy working populations this is seen to a lesser extent. [10]

  We think the response by Bosma and Bobak illustrates another important point in psychosocial epidemiology. For example consider their use of the adjective “specific” in this context. None of the proposed psychoneuroendocrine or behavioural theories linking psychosocial exposure to disease outcome demands a psychosocial stimulus any more specific than “perceived threat”. [4] Specificity in this context has simply been attributed to “vital exhaustion”, post hoc, on the basis of a negative relation with health.

  A similar iterative process appears to have influenced the conception of “job strain” – originally proposed as the toxic interaction between high job demands and low job control. [11] However experience in a number of studies suggests that it is low job control alone that predicts poorer health. [12] The consistency of the relation between job control and other indices of personal autonomy and health could relate to their reliability as measures of social position. Worker autonomy is even incorporated in contemporary social classifications. [13]

  Again our points are simple: responses to psychometric instruments, no matter how elaborate the instrument, are often socially patterned and such patterning must be considered when attempting to attribute effects of psychosocial exposure – furthermore all are likely to be influenced, to a greater or lesser extent, by reporting bias, an issue we elaborate elsewhere. [14] We believe that it is important to address these issues and consider how they might best be overcome.

  Bosma and Bobak suggest that we are saying that in the relevant publications resulting from the Whitehall II and GLOBE studies, no attempt was made to adjust for social position. This is not what we wrote. In both these studies analyses incorporated such measures of social position as were available. Our point was that all measures of social position are inevitably imprecise, and that a considerable degree of residual confounding is possible.

  In their final paragraph Bosma and Bobak seem to disagree with our suggestion that, observational studies of psychosocial factors in populations where these factors are associated with material disadvantage (and thus are potential candidates for explaining social health inequalities) may be difficult to interpret. This was simply an observation and did not reflect any value judgement on their chosen research priorities. Indeed, our underlying motivation for involvement in these investigations is, we assume, the same as their own – a desire to explain health inequalities so that they may be more effectively reduced. In this context we have not dismissed the possibility of an important role for psychosocial factors. We certainly do not suggest that research on these factors be abandoned. If interventions targeting psychosocial exposures are truly to be an effective means to improve health and reduce health inequalities then this should be demonstrable in experimental studies. We strongly support any research team wishing to undertake – and report the findings – of such studies.

  References
  (1) Smith GD, Morris JN, Shaw M The independent inquiry into inequalities in health. BMJ 1998; 317: 1465-1466.
  (2) Lynch JW, Davey Smith G, Kaplan GA, House JS Income inequality and mortality: importance to health of individual income, psychosocial environment, or material conditions BMJ 2000; 320: 1200-1204.
  (3) Heslop P, Davey Smith G, Carroll D, Macleod J, Hart C. Perceived stress and coronary heart disease risk-factors: The contribution of socio- economic position. British Journal of Health Psychology 2001; 6: 167-178.
  (4) Brunner E. Stress and the biology of inequality. BMJ 1997;314:1472 -6.
  (5) Knoke D, Bohrnstedt GW. Statistics for social data analysis. Itasca, Illinois: Peacock, 1993. p286.
  (6) Finkel SE. Causal analysis with panel data. Thousand Oaks, CA: Sage, 1995. p14.
  (7) Bosma H. A cross-cultural comparison of the role of some psychosocial factors in the etiology of coronary heart disease. Follow-up to the Kaunas-Rotterdam Intervention Study (KRIS). (PhD thesis). Maastricht: Datawyse/Universitaire Pers.
  (8) Egger M, Davey Smith G. Bias in location and selection of studies. BMJ 1998;316:61-6.
  (9) Appels A, and Otten F. Exhaustion as precursor of cardiac death. Br J Clin Psychol 1992 ;31:351-356.
  (10) Davey Smith G, Shipley MJ, Marmot MG, Rose G. Plasma cholesterol concentration and mortality. The Whitehall Study. JAMA 1992;267:70-6
  (11) Karasek R, Theorell T. Healthy work: stress, productivity and the reconstruction of working life. New York: Basic Books,1990.
  (12) Schnall PL, Landsbergis PA. Job strain and cardiovascular disease. Annu Rev Public Health 1994;15:381-411.
  (13) Rose D, O'Reilly K. Final report of the ESRC review of government and social classifications. Swindon: Economic and Social Research Council and Office for National Statistics, 1998.
  (14) Macleod J, Davey Smith G, Heslop P, Metcalfe C, Carroll D, Hart C. Limitations of adjustment for reporting tendency in observational studies of stress and self-reported coronary heart disease. J Epidemiol Community Health 2002;56:76-78.

  Show Less

  Conflict of Interest:
  None declared.

  • Back to top
• Published on: 27 April 2016

  Confusing confounding and causation

  • A Singh-Manoux, Research Psychologist
  • Other Contributors:
    • P Clarke

  Dear Editor,

  Macleod et al[1] reported the influence of confounding by socioeconomic factors in observational epidemiology, questioning the role of psychosocial exposures in predicting ill-health. The authors concluded that the association between psychosocial exposures and health reported in the literature might be spurious, a result of the association between socioeconomic position and psychosocial exposures. We read wit...

  Show More

  Dear Editor,

  Macleod et al[1] reported the influence of confounding by socioeconomic factors in observational epidemiology, questioning the role of psychosocial exposures in predicting ill-health. The authors concluded that the association between psychosocial exposures and health reported in the literature might be spurious, a result of the association between socioeconomic position and psychosocial exposures. We read with interest this attempt to explicate the causal pathways linking social position to health but would question the way in which the authors have conceptualised the problem. Analysis of causal influences in social epidemiology necessitates two kinds of questions. First, is socioeconomic position (A) causally related to ill-health (B)? There is now widespread evidence on the existence of a relationship between socioeconomic position and ill-health. Second, is the "how" question; the mechanisms by which A leads to B. The search for satisfactory answers to the "how" question now concern researchers in social epidemiology. Research over recent years suggests that the relationship between socioeconomic position and health is likely to be mediated by three pathways: materialist, health behaviour, and psychosocial[2] [3]. If one is to start out by accepting the existence of multiple pathways, the conclusions drawn by Macleod et al[1] are misleading on two accounts:

  1. The first problem concerns the way in which regression analysis has been used to deal with the effect of occupational class. The authors conceptualise occupational class as a confounder/covariate in the regression equation linking psychosocial exposures and health. However, this approach can produce biased estimates if the covariates in the regression are themselves causally related[4]. (Note that this does not apply to any set of correlated covariates, only to those where one precedes another in the causal sense.) As the authors accept that the social patterning of psychosocial constructs implies that psychosocial exposures are causally linked to occupational class, the estimates of occupational class and perceived stress in the regression may be biased, and cannot be interpreted as the independent effects of these variables on ill-health. In other words occupational class takes precedence in the temporal causal sequence linking social position to ill-health. The causal sequence would be A (socioeconomic position) leading to X (materialist, health behaviour, and psychosocial pathway) which in turn leads to B (ill-health). Multivariate regression analysis used by Macleod et al[1] ignores the temporal and causal pathways. Therefore, adjusting the effect of occupational class when analysing the relationship between perceived stress and ill-health in order to reveal the independent effect of perceived stress on ill-health is erroneous.

  2. The second problem with the conclusions drawn is that the authors favour the materialist or the behavioural pathway without assessing the "independent" effect of these pathways on health outcomes. It isn't difficult to envisage a strong correlation between occupational class and material advantage. Treating occupational grade as a covariate (adjusting for its effects) when analysing the relationship between material disadvantage and ill-health is likely to lead to results similar to the one reported in this paper. We suggest that a way forward in the investigation into "how" social class affects ill-health needs to be theory driven. This would ensure that the temporal and causal sequence of variables analysed isn't lost, and the statistical procedures are chosen with care. Structural equation modelling[5] allows modelling of all three pathways (material, psychosocial, behavioural) concurrently in order to assess the relative importance of each in predicting health.

  (1) Macleod J, Davey Smith G, Heslop P, Metcalfe C, Carroll D and Hart C. Are the effects of psychosocial exposures attributable to confounding? Evidence from a prospective observational study on psychological stress and mortality. J Epidemiol Community Health 2001; 55:878-884.

  (2) Davey Smith G, Hart C, Hole D, MacKinnon P, Gillis C, Watt G, Blane D, and Hawthorne V. Education and occupation social class: which is the more important indicator of mortality risk? J Epidemiol Community Health 1998; 52:153-160.

  (3) Marmot M, and Theorell T. Social class and cardiovascular disease: the contribution of work Int J Health Serv 1988; 18:659-674.

  (4) Knoke D, and Bohrnstedt G W. Statistics for Social Data Analysis. Itasca, Illinois: Peacock, 1993 (pp. 283-285).

  (5) Finkel S. Causal Analysis with Panel Data. London: Sage, 1995.

  Show Less

  Conflict of Interest:
  None declared.

  • Back to top
• Published on: 27 April 2016

  How does social class get under the skin?

  • Hans Bosma, Researchers
  • Other Contributors:
    • Martin Bobak

  Dear editor,

  With great interest we read the paper by Macleod and co-workers on the influence of psychosocial exposures on 21-year mortality. According to the authors, the small positive effects of stress on mortality became attenuated after control for social class. From this the authors concluded that the effects of psychosocial factors on health reported in a large number of studies from a number of populations...

  Show More

  Dear editor,

  With great interest we read the paper by Macleod and co-workers on the influence of psychosocial exposures on 21-year mortality. According to the authors, the small positive effects of stress on mortality became attenuated after control for social class. From this the authors concluded that the effects of psychosocial factors on health reported in a large number of studies from a number of populations were due to (residual) confounding by socioeconomic factors. They suggest that the effect of psychosocial exposures can only be examined when socioeconomic factors are stringently controlled for or in populations where psychosocial exposures are not related to social class. The discussion implies that such a careful control for confounding would entirely eliminate psychosocial exposures as true risk factors for poor health and premature mortality, since it is only socio-economic structure that matters and should accordingly be intervened upon.

  We agree with the authors that studies of psychosocial factors need to carefully control for socioeconomic circumstances, the more so because unfavourable psychosocial factors often result from an unfavourable social and material environment. However, we feel that the overall thrust of the paper – that psychosocial factors have no effect of their own - is not justified by the evidence provided.

  A careful examination of the paper reveals that the argument is not consistent with the authors’ own data. The particular measure of stress used by the authors was not related to mortality, before or after adjustment for social class. How can there be any confounding, residual or otherwise? Rather than to speculate about confounding, it would be more appropriate, in our opinion, to examine the reasons why there was no association between stress and mortality in this study. A likely reason for this is that the Reeder stress scale is an outdated and unspecific measure. In the Kaunas-Rotterdam Intervention Study, the Reeder stress scale did not predict either mortality or nonfatal MI.[1] However, one particular item (“At the end of the day, I am completely exhausted mentally and physically”) strongly predicted short-term mortality from CHD.[1,2] Using measures with higher specificity (in this case on vital exhaustion) would increase the predictive power. There is growing evidence that other, more specific measures, such as low social support, depression, and low perceived control are independently related to poor health.[3]

  The conclusion that psychosocial factors are not really related to poor health is not based on an exhaustive literature review showing that social class was not adequately controlled for in all previous studies. In fact, careful evaluation of the Whitehall II papers on job control and the Dutch GLOBE reports on perceived control (referenced in the paper) shows that the effects of psychosocial factors persisted after controlling for socio-economic factors.

  The recommendation that psychosocial factors should only be examined in populations where psychosocial adversities are unrelated to social class is also problematic. How should we investigate the mechanisms underlying socio-economic differences in health? Should we accept a priori that all social variations are due to materialist explanations? (And which ones?) The problem is that such factors typically explain only a minor part of the socio-economic differences in health. Researchers are therefore increasingly looking for psychosocial mechanisms as the link between low socio-economic status and poor health. These may affect health through psycho-physiological and/or behavioural pathways. In either case, there is no alternative to studying psychosocial factors that are related to (and “caused” by) social disadvantage in order to know how social structures get under the skin.

  (1) Bosma H. A cross-cultural comparison of the role of some psychosocial factors in the etiology of coronary heart disease. Follow-up to the Kaunas-Rotterdam Intervention Study (KRIS). (PhD thesis). Maastricht: Datawyse/Universitaire Pers.

  (2) Appels A, and Otten F. Exhaustion as precursor of cardiac death. Br J Clin Psychol 1992 Sep;31(Pt3):351-356.

  (3) Hemingway H, and Marmot M. Evidence based cardiology: Psychosocial factors in the aetiology and prognosis of coronary heart disease: systematic review of prospective cohort studies. BMJ 1999; 318:1460-1467.

  Show Less

  Conflict of Interest:
  None declared.

  • Back to top