Factors Explaining Socio-Economic Inequalities in Survival from Colon Cancer: A Causal Mediation Analysis

In high-income countries, people of lower socio-economic position (SEP) have higher excess mortality after diagnosis with colon cancer than people with higher SEP (1, 2). In Australia, bowel cancer is the second most common cancer diagnosed in both men and women, and the second most common cause of cancer-related deaths (3). Previous Australian studies observed lower survival for disadvantaged people with colorectal cancer and reported that disease stage and other tumour characteristics did not explain the observed gaps in survival (4–7). In contrast, studies conducted in the United States, Europe, and New Zealand suggested that cases from more socio-economically disadvantaged backgrounds present with more advanced stage (8–12). Few studies have assessed the potential role of health-related behaviours, comorbidities, and treatment in inequalities in cancer survival (13, 14). Interpreting the findings of studies that have attempted to identify the underlying reasons for socio-economic inequalities in colorectal-cancer survival is difficult because most compared the effect of SEP on survival with and without controlling for mediators, which can give biased results, particularly in the presence of multiple mediators influencing or interacting with each other (15, 16). We aimed to assess 1- and 5-year net survival for patients with colon cancer living in the most and least disadvantaged areas, and investigate the potential role of substage, comorbidities, surgery, chemotherapy, emergency presentation prior diagnosis, surgical hospital type, surgical volume of a hospital, and number of nodes taken in socio-economic inequalities in survival from stage III colon cancer by applying causal mediation analysis to linked population-based health data.

We identified colon cancer cases using the population-based Victorian Cancer Registry (VCR). All hospitals and pathology laboratories in Victoria are required to notify the registry of new malignant invasive neoplasms. The VCR collects demographic and tumour details (17) and derives tumor–node–metastasis (TNM) summary stage at diagnosis from information on pathology reports and from hospital notifications (18). Information about comorbidities and the treatment received was obtained via record linkage between the VCR and the Victorian Admitted Episodes Dataset (VAED; ref. 19). The VAED comprises demographic, clinical, and administrative data for every admission to Victorian public and private hospitals, including rehabilitation centres, extended care facilities, and day procedure centres (19).

Eligible cases were diagnosed with first primary colon cancer [International Classification of Diseases (ICD)-10; C18.0-C18.9] at age 15 to 79 years between 2008 and 2011. Cases 80 years or older at diagnosis were not included as they are more likely to receive oral chemotherapy, for which we did not have information. We excluded cases diagnosed with any previous cancer, diagnosed with colon and rectal cancer on the same day, and cases diagnosed by autopsy or death certificate only. We also excluded the few Indigenous cases as they are more likely to be disadvantaged than the general population. We conducted a complete-case analysis excluding cases with missing data on socio-economic disadvantage, stage, or comorbidities. Follow-up was completed on December 31, 2016.

The Index of Relative Socio-Economic Disadvantage (IRSD), one of the socio-economic indexes for areas (SEIFA) constructed by the Australian Bureau of Statistics using census data, was used to define an aggregated composite measure of SEP (20). For each case, this was based on the smallest geographical area of their usual residential address at the census closest to their year of diagnosis, i.e. 2006 census for cases diagnosed in 2008, and 2011 census for those diagnosed between 2009 and 2011. For individuals diagnosed between 2009 and 2011, addresses were geocoded and mapped to precise latitude and longitude, and then mapped to Australian Bureau of Statistics Statistical Area level 1, which have a mean population size of approximately 400 persons (21); for individuals diagnosed in 2008, addresses were geocoded and mapped to census collection district, which have a mean of 225 dwellings (22). SEIFA quintiles were defined based on the assigned IRSD scores for each geographic area. A lower score represents more socio-economically disadvantaged areas with households with lower income and people in less skilled occupations or with lower education (20). The most and least socio-economically disadvantaged areas were defined as the fifth and first quintiles of SEIFA, respectively. Hereafter, we use the word ‘disadvantage’ for this variable.

Deaths and cause of death were identified by linkage to the Victorian Registry of Births, Deaths and Marriages and the National Death Index (NDI). Causes of Victorian deaths were coded by the VCR staff using information from several medical records such as death certificate, each person's history of cancer diagnosis/diagnoses, recent hospital admission for recurrent or metastatic disease. Causes of death for people who died interstate were obtained from the NDI, which uses coding by the Australian Bureau of Statistics.

We defined comorbidities using the Charlson Comorbidity Index (23) dichotomised into absence or presence of any comorbidity based on admissions to hospital a year prior to diagnosis or 30 days after diagnosis to capture surgical admission which for many cases would be the first opportunity to have their comorbidities reported. For stage III disease, substage was categorized as 3A/3B and 3C; surgery and intravenous adjuvant chemotherapy were defined dichotomously (yes/no) according to the Victorian optimal care pathway recommendations (24); surgery had to occur within 4 weeks of diagnosis and chemotherapy within 8 weeks following surgery to be coded as such.

For mediation analysis restricted to cases who received surgery within 6 months of diagnosis, other potential mediators considered were emergency presentation 0 to 7 days prior to diagnosis (no/yes), type of hospital in which surgery was performed (private or public), annual median number of colon cancer surgeries performed at the hospital, i.e. hospital caseload (≥32 or <32 surgeries, 32 being the median across hospitals in Victoria; ref. 25), as a proxy measure for hospital quality, and number of nodes taken during surgery (≥12 or <12) as a proxy measure for quality of surgery.

Analyses were restricted to the most and least disadvantaged categories to assist with the interpretation of the estimated mediating effects. Including the middle categories in the mediation analysis makes the assumption that the effects of the mediators on the outcome, and the effects of other confounders on the outcome are constant across all quintiles. The number of deaths was too small to consider fitting the interactions to assess these assumptions. One- and five-year net survival was estimated by stage, sex, and socio-economic disadvantage using the Pohar–Perme method (26), based on Victorian population life tables stratified by year, age, sex, and SEIFA quintile. For mediation analyses, the outcome was death from colon cancer; cases who died of other causes were retained in the analysis and treated as without the outcome of interest (censoring was not possible in the mediation analyses). We also assessed the total effect of disadvantage on death from colon cancer applying logistic regression models and Cox regression in which cases who died of other causes were censored at date of death.

Mediation analyses were performed for stage III disease only for the following reasons: (i) death from colon cancer was common for cases with stage III disease but death from other causes was uncommon (Table 1); (ii) few cases with stage I disease died from any cause; (iii) more cases with stage II disease died from causes other than colon cancer, which would obscure any effect of disadvantage on cause-specific survival; and (iv) for cases with stage IV disease, the probability of dying within the first 3 months after diagnosis was substantially higher for those living in the most disadvantaged areas than for their counterparts from the least disadvantaged regions, suggesting that disadvantaged cases were more likely to die before they could receive treatment.

For the causal mediation analyses, we used a multiple mediator interventional approach to decompose the total effect (TE) of socio-economic disadvantage on survival into an interventional direct effect (IDE), not through any of the mediators, and an indirect effect (IIE; ref. 27), estimated as risk differences. This method requires the averaging of effects across Monte-Carlo draws of the mediators; we used 600,000 draws. Confidence intervals were obtained from 1,000 bootstrap samples. We also fitted logistic regression models to assess associations between disadvantage and the outcome, between disadvantage and potential mediators, and between the mediators and the outcome (death from colon cancer).

Our hypothesised causal structure for the effects of socio-economic disadvantage on colon cancer survival were based on existing literature (Supplementary Fig. S1; refs. 1, 14, 28). For stage III disease, potential mediators were comorbidities, substage, receiving surgery within 4 weeks of diagnosis, and chemotherapy within 8 weeks following surgery (Supplementary Fig. S2A). As a secondary analysis, we extended the time to surgery to 6 months to allow more time for events to occur that might be considered ‘mediators’. Thus, for the subset of cases who received surgery within 6 months of diagnosis, mediators were comorbidities, substage, receiving chemotherapy within 8 weeks following surgery, emergency presentation prior to diagnosis, surgical-hospital type, surgical volume of a hospital, and number of nodes taken (Supplementary Fig. S2B).

The models to estimate IIEs for men and women were (i) logistic regression of comorbidities conditional on disadvantage and age; (ii) logistic regression of substage conditional on disadvantage and age; (iii) logistic regression of surgery conditional on disadvantage and age; (iv) logistic regression of chemotherapy conditional on disadvantage and age; (v) logistic regression of the outcome (death from colon cancer) conditional on disadvantage, all mediators and age. A similar approach was used to define models estimating the mediating effects of emergency presentation, hospital type, surgical volume, and number of nodes taken. We repeated mediation analysis that allowed for potential exposure–mediator interactions, but only included interaction term between exposure and mediators that did not create combinations with sparse data (i.e., sufficient number of at least 30 cases in each category).

People residing in nonmetropolitan areas are generally more disadvantaged and have lower cancer survival (3). Because area-level socio-economic disadvantage and remoteness of residence were defined based on the same geographical areas, rather than adjusting for rurality, we conducted a sensitivity analysis restricted to cases residing in major cities to remove any potential confounding effects of factors associated with rurality.

All analyses were performed using Stata/MP version 14.2 (Stata Corporation LP). Approval to conduct the analyses was granted by the data custodians at the Victorian Department of Health and Human Services and the Cancer Council Victoria Human Research Ethics Committee. All procedures performed in this study involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. For this type of study formal consent was not required.

We identified 2,203 eligible colon cancer cases after excluding 4,615 cases; most cases excluded were in the middle SEP groups (n = 3,171; Fig. 1). The characteristics of cases by stage at diagnosis are summarised in Table 1. Eight hundred and seven cases died within 5 years of diagnosis (684 cases died of colon cancer and 123 died of other causes). Cases residing in the most disadvantaged areas were diagnosed at older ages (67 relative to 65 years). The least and most disadvantaged areas had almost identical stage distributions (Table 2; P = 0.4).

For stage I disease, 5-year net survival was close to 100%; while for stage II, it was mostly greater than 90%, but lower for the most disadvantaged cases, relative to the least disadvantaged (Table 3). For stage III disease, 1-year survival did not differ between those residing in the least and most disadvantaged areas; however, 5-year survival was 13% lower for the most disadvantaged areas. The gap in survival was greater for men than women. Of cases with stage IV disease, those residing in the most disadvantaged areas had lower 1- and 5-year survival than their counterparts from the least disadvantaged areas, although the difference in 5-year survival was restricted to men (Table 3).

With respect to the association between socio-economic disadvantage and death from stage III colon cancer, the results of both logistic regression and Cox regression analyses showed that disadvantaged men had lower 5-year survival than advantaged cases although the effect of disadvantage was slightly weaker in the Cox model (Supplementary Table S1).

Of men with stage III colon cancer, those residing in the most disadvantaged areas were more likely to have comorbid conditions than those living in the least disadvantaged regions; we observed no association for women (Supplementary Table S2). Residents of the most disadvantaged areas were more likely not to have received surgery within 4 weeks of diagnosis and chemotherapy within 8 weeks after surgery, and were more likely to have had an emergency presentation prior to diagnosis and to have received surgery at public hospitals than the least disadvantaged cases (Supplementary Table S2).

Cases with comorbidities or those having emergency presentation prior to diagnosis were more likely to die from colon cancer (Supplementary Table S3). Not receiving chemotherapy within 8 weeks after surgery was associated with lower survival. The odds of dying within 5 years after diagnosis were greater for cases with late–stage III cancer compared with earlier stages (IIIA and B). Men with stage III disease who received surgery in public hospitals were more likely to die within 5 years of diagnosis than their counterparts having surgery in private hospitals. There was no evidence of association between survival and surgical volume of the hospital for stage III disease. Results for receiving surgery within 4 weeks of diagnosis and the number of nodes taken in surgery for cases with stage III disease were too imprecise to be informative (Supplementary Table S3).

Men in the most disadvantaged areas had substantially higher excess deaths from stage III colon cancer than their counterparts from the least disadvantaged areas, while for women there was little effect of disadvantage (Fig. 2). For every 1,000 men with stage III disease, we estimated that there were 161 [95% confidence interval (CI), 67–256] additional deaths from colon cancer in the 5 years following diagnosis in the most disadvantaged areas relative to the least disadvantaged (Fig. 2). Almost all this excess was the direct effect (i.e., not through the mediators), with comorbidities and chemotherapy explaining very little of the excess mortality. Surgery and substage did not contribute to the observed additional deaths for men residing in the most disadvantaged areas (Fig. 2).

For men with stage III cancer, we further investigated whether emergency admission prior to diagnosis, the type of hospital in which surgery was done (public or private), surgical volume of the hospital, or the number of nodes taken during surgery (as a measure of quality) explained the excess deaths for men in the most disadvantaged areas. Of the 142 (95% CI, 48–235) additional deaths per 1,000 for men receiving surgery within 6 months of diagnosis and residing in the most disadvantaged areas relative to their counterparts in the least disadvantaged regions, 31 (95% CI, −15–77) were explained through receiving surgery at a public hospital. Emergency presentation prior to diagnosis explained very little of the observed differences in survival (Fig. 3). There were no mediating effects through surgical volume or number of nodes taken.

Adding interactions between the exposure and the mediators did not markedly change any results (Supplementary Figs. S3 and S4). Excluding cases residing outside major cities did not materially alter the results for men, although the mediating effect of comorbidities and chemotherapy was greater in major cities (Supplementary Figs. S5 and S6). For women, the gap in survival and the mediating effect of substage and chemotherapy was greater in major cities (Supplementary Fig. S5).

We observed socio-economic inequalities in colon cancer survival that were greater for men than women. The stage distribution could not explain these differences in survival because it did not differ by socio-economic disadvantage. Lower 5-year survival from stage III colon cancer for cases residing in the most disadvantaged areas, relative to the least disadvantaged regions, was not explained by differential surgical treatment. Comorbidities and chemotherapy contributed very little to the observed differences; however, lower survival for men living in the most disadvantaged areas was perhaps partly mediated through receiving surgery at public hospitals (the CI was wide and included the null). Emergency presentation prior to diagnosis, while more common for patients from the most disadvantaged areas, explained very little of the gap in survival. Surgical volume of the hospital and number of nodes taken, as markers of hospital and surgical quality, did not explain the observed socio-economic inequalities in survival from stage III colon cancer.

None of the previous studies that investigated the underlying reasons for inequalities in colorectal cancer-specific survival by educational level or SEP used causal mediation analyses (4, 8, 11, 29, 30). Interventional mediation analysis, which does not require strong assumptions about the sequence in which mediators act, enabled us to attempt to decompose the total effect and quantify indirect effects through multiple interrelated mediators (27). Other advantages of this method are the ability to identify pathways through multiple mediators even when they share unmeasured common causes, and to include exposure–mediator interactions in the models (27). Another strength of our study is the population-based design and the linkage to inpatient hospital data, which allowed us to assess the role of treatment and comorbidities in socio-economic inequalities in survival from stage III colon cancer. The main limitations are the sample size and the way we measured socio-economic disadvantage. We used an area-based measure of SEP rather than an individual-based measure, which could have resulted in misclassification and underestimation of the total effect of SEP on colon cancer survival. We did not have information on health-related lifestyle behaviours or oral chemotherapy, which is an issue for older cases, who are more likely to receive oral chemotherapy. Also, comorbidities were underreported, as they were only recorded when related to the hospital admission. Moreover, we defined surgery and chemotherapy as binary variables, which could have diluted their indirect effects and increased the direct effects.

Findings from previous Australian studies are generally consistent with our results. One Australian study showed that adjustment for stage at diagnosis did not attenuate the observed socio-economic inequalities in survival from colorectal cancer (4). The other Australian study found that the association between socio-economic status and survival persisted after adjusting for stage at diagnosis, comorbidities, treatment, patient characteristics, or tumour characteristics (30). A Swedish study reported that variation in surgery type (elective or emergency) and hospital location (urban or rural) did not explain the higher excess mortality from colon cancer among cases with low educational level (8). In contrast, studies from New Zealand and England found that tumour stage and treatment partly contributed to socio-economic inequalities in survival from colorectal cancer (11, 29). The inconsistency between countries in the estimated effects of these mediators might be due to differences in health care systems, limitations of the data, the statistical methods applied, or a combination of these factors.

A recent review by the International Agency for Research on Cancer (IARC) reported higher prevalence of unhealthy diet, smoking, heavy alcohol intake, physical inactivity in individuals with low levels of income and education as potential contributing factors to inequalities in cancer survival (28). Differences in screening uptake, stage at diagnosis, and access to diagnostic and treatment services across socio-economic groups have also been highlighted as other explanatory factors (28), but our study and other Australian studies suggest that in Australia, stage does not contribute to the observed gap in survival.

We have previously shown that socio-economic inequalities in survival from colorectal cancer were greater for men than women (31). In this study, we observed a similar effect for colon cancer, for which we have no explanation. Further, in our previous study, socio-economic differences were similar for men and women for most cancer sites, suggesting that differential measurement error by sex is not an explanation (31). It is unclear why lower survival of disadvantaged men with stage III disease appeared to be partly explained by having surgery in public rather than private hospitals. This observation might be due to other unmeasured socio-economic characteristics of men from disadvantaged areas receiving surgery in public hospitals, rather than the public or private hospitals themselves, as we found that the observed gap in survival was not explained by differences in surgical volume of the hospital or the number of nodes taken, i.e. quality of the hospital and surgical resection. Cases with more serious conditions might be more likely to go to public hospitals. Moreover, emergency presentation is more likely to be to a public hospital as few private hospitals in Victoria have emergency departments.

In summary, despite Australia's universal health care system, we observed lower survival from stage III colon cancer for cases residing in the most disadvantaged areas, which was barely explained by differences in comorbidities, chemotherapy, emergency presentation, or surgical-hospital type. Given that the mediators we used in our analyses relate to care in the first year after diagnosis, it is essential that future studies assess the potential role of later factors such as compliance with surveillance for recurrence and supportive care services.

N. Afshar reports this work was supported by the Australian National Health and Medical Research Council [grant number 1150012]; in addition, N. Afshar was the recipient of an Australian Government Research Training Program Scholarship. G.G. Giles reports grants from National Health and Medical Research Council (Australia) during the conduct of the study. R.L. Milne reports grants from NHMRC during the conduct of the study. D.R. English reports grants from Australian National Health and Medical Research Council during the conduct of the study. No disclosures were reported by the other authors.

N. Afshar: Conceptualization, formal analysis, validation, investigation, methodology, writing–original draft, writing–review and editing. S.G. Dashti: Methodology, writing–review and editing. L. te Marvelde: Data curation, writing–review and editing. T. Blakely: Methodology, writing–review and editing. A. Haydon: Writing–review and editing, provide clinical advice. V.M. White: Writing–review and editing. J.D. Emery: Writing–review and editing. R.J. Bergin: Writing–review and editing. K. Whitfield: Data curation. R.J.S. Thomas: Writing–review and editing. G.G. Giles: Writing–review and editing. R.L. Milne: Conceptualization, resources, supervision, funding acquisition, validation, investigation, writing–review and editing. D.R. English: Conceptualization, formal analysis, supervision, validation, investigation, methodology, writing–review and editing.

This work was supported by the Australian National Health and Medical Research Council (grant number 1150012). N. Afshar was the recipient of an Australian Government Research Training Program Scholarship.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.