Sociodemographic disparities in survival from colorectal cancer in South Australia: a population-wide data linkage study

A mixture of probabilistic and deterministic linkage methods were used to develop a comprehensive population-wide dataset incorporating demographic, clinical and outcome data for all CRC cases. SACR records pertaining to eligible cases constituted the core dataset. Linkage between the SACR and public hospital separation data was undertaken by SANT DataLink using probabilistic matching based on name, address and date of birth, to retrieve records pertaining to all admissions for each individual, which included an admission code for CRC. Linkage between the SACR and private hospital separation data, radiotherapy services and hospital-based cancer registries was undertaken by the Department of Health through direct matching of hospital codes, patient record numbers, cancer registry accession number and personal identifiers, again to identify all relevant records pertaining to CRC diagnosis or treatment from these data sources. SANT DataLink linkage processes followed the best practice principle of ‘data separation’ to ensure privacy and anonymity [25]. Figure 1 shows the data sources and linkage processes used in the construction of this dataset.

All eligible cases had previously been manually staged by SACR staff for a previous study [5], according to Australian Clinico-Pathological Staging (ACPS) criteria [26], based on data from pathology reports, supplemented by electronic patient records and hospital-based cancer registry (HBCR) data. Grade of the primary tumour at diagnosis was extracted from SACR records and was coded as low, intermediate or high grade corresponding to well-differentiated, moderately-differentiated and poorly-differentiated tumours based on pathological findings. Stage and grade were unknown in 278 (6.0 %) and 362 cases (7.8 %), respectively.

Area of residence was classified into 5 categories:- inner urban (very accessible), outer urban (moderately accessible), rural (accessible), remote and very remote, according to the Accessibility/Remoteness Index of Australia, 2006 [27], based on residential postcode at diagnosis as recorded by the SACR. This measure describes place of residence according to road distance from population service centres. Remote and very remote categories were combined for analyses due to small numbers. Similarly, an area level measure of SES was assigned using the Index of Relative Socioeconomic Advantage and Disadvantage (IRSAD) 2006 [28], based on residential postcode. This index measures area-level rather than individual-level characteristics and represents the average socioeconomic status of specified neighbourhoods (in this case postal areas). It incorporates factors such as average household income, education levels and unemployment rates. SES was categorised into quintiles corresponding to SA's population distribution. No individual level measures of SES were available. Private health insurance status was assigned if any individuals’ hospital admissions flagged the patient as being privately insured. Patient co-morbidity was assessed using the Charlson co-morbidity index (CCI) [29] for the patient’s first admission for surgical treatment or their first admission for CRC if there was no surgical admission. Diagnosis codes for CRC and metastatic disease were excluded when calculating the CCI. CCI scores were further categorised into no co-morbidity (CCI = 0), a single co-morbidity (CCI =1) or multiple co-morbidities (CCI > =2).

Treatments were assessed from multiple sources and coded dichotomously (yes/no) based on any indication in any data source (i.e. any procedure codes for relevant treatment in hospital admissions, HBCR or radiotherapy records). ‘Primary treatment’ was defined as treatment commencing within 12 months of diagnosis, based on the date of the first hospital admission or the treatment start date in the HBCR or radiotherapy datasets if there was no hospital admission record for that treatment.

Information about vital status was derived from the SACR, which routinely links with state and national death notifications to track dates and causes of death occurring in any Australian jurisdiction.

Ordinal logistic regression was undertaken to identify factors independently associated with stage at diagnosis. ACPS stage, categorised as four separate stages (A-D), was the dependent variable. Unknown stage was excluded. Adjusted odds ratios were obtained from a multivariate model adjusting for all covariates simultaneously (age group, sex, cancer site, SES (quintiles), place of residence, co-morbidity, primary surgery, radiotherapy, and chemotherapy, and calendar year of diagnosis (continuous). Cancer-specific survival was examined crudely, in the first instance, using Kaplan-Meier product limit estimates without taking competing risk into account, with differences assessed using log rank tests. We then assessed factors independently associated with risk of CRC death, using competing risk regression analysis according to the method of Fine and Gray [30], where death from other causes was the competing risk. Cases were censored on December 31, 2012 or at their date of death (whichever occurred earlier). Follow-up time was calculated as the time from the date of diagnosis to the censoring date or date of death if death occurred before this date. Multivariable models included cancer site, stage, grade, age, sex, year of diagnosis, comorbidity categories, SES, place of residence, and primary treatments received. Cases with missing stage were excluded from analyses, whereas cases with missing grade were included as a separate ‘unknown’ category. We repeated the competing risk analysis with imputed data for missing stage and grade, derived using multiply imputed chained equations [31], with 20 imputed data sets. Only complete case analyses are reported since results did not differ substantially for multiply imputed data. All analyses were undertaken using Stata v12.0 software [32].

This study received ethical approval from Human Research Ethics Committees of SA Health, the Royal Adelaide Hospital, Flinders University/Medical Centre and the SA Aboriginal Health Council, along with data access approval from SA Health for government managed datasets.

Of the 4949 CRC cases identified from the SACR, 113 were excluded due to ineligible histology, and 195 due to lack of hospital admission records, leaving a total of 4641 for analysis. The average period of follow-up was 4.7 yrs, corresponding to 6.3 yrs among survivors and 2.3 yrs among those who died of CRC or other causes. Demographic and clinical characters of the study cohort are presented in Table 1.

In unadjusted analyses, distribution of stage at diagnosis did not differ by remoteness of residence or SES, but did differ by cancer site, age group, extent of comorbidity and health insurance status (Table 2). Multivariate analysis using ordinal logistic regression indicates that later stage was less likely among rectal cancer cases (0R = 0.81, 95 % CI 0.72-0.91), older patients (0R = 0.77, 95 % CI 0.67-0.89 for 70-79 yrs v 50-59 yrs) and privately insured patients (OR = 0.84, 95 % CI 0.75-0.95) and more likely for those with co-morbidities (OR = 1.50, 95 % CI 1.26-1.79 for CCI > = 2 v 0). Stage at diagnosis was not associated with socioeconomic status (area-level), place of residence, or sex in multivariate analyses.

CRC specific survival was 88 %, 75 % and 69 % at 1, 3 and 5 years respectively (Table 3). In crude analysis, survival was strongly associated with stage of disease, grade of tumour, age at diagnosis and severity of co-morbid conditions. Survival was also associated with receipt of specific treatments (with better survival among those having surgery compared with no surgery and worse survival for those who had radiotherapy or chemotherapy compared with those not receiving these treatments. There were no significant differences by place of residence but survival did differ across SES quintiles and according to whether patients were privately insured. Multivariate competing risk regression (Table 4) revealed similar results, though there were some slight differences to crude survival analyses. Risk of CRC death was higher among patients with later stage disease (e.g. HR = 7.74, 95 % CI 5.75-10.4 for stage C vs stage A), higher grade (HR = 2.25, 95 % CI 1.32-3.84 for high verses low grade), and severe or multiple co-morbidities compared with no comorbidity (HR = 1.22, 95 % CI 1.02-1.44) and lower among rectal compared with colon cancer patients (HR = 0.85, 95 % CI 0.72-0.98). However there were no significant differences by age group or by health insurance status after adjusting for other covariates and accounting for death from other causes. Risk of CRC death was significantly lower among those who underwent surgery (HR = 0.51 95 % CI 0.42-0.62) and higher among those who received radiotherapy (HR = 1.41, 95 % CI 1.18-1.68), but unlike crude analyses, CRC death was lower among those who received chemotherapy (HR = 0.87 95 % CI 0.76-1.00). No differences were found according to place of residence, however risk was significantly lower among the highest compared with lowest SES group (HR = 0.75, 95 % CI 0.62-0.91). Risk of CRC death declined significantly over the study period (HR = 0.95, 95 % CI 0.92-0.98).

Results of stratified analyses, restricted to cases with potentially curable (stages A-C), or metastatic disease (stage D), are shown in Tables 5 and 6. For potentially curable disease, risk of CRC death was higher among residents from remote regions compared with metropolitan residents (HR = 1.35, 95 % CI 1.01-1.80). Regional differences were not evident among patients with stage D CRC at diagnosis. Risk CRC death was lower among those in the highest compared with lowest socioeconomic quintile, both among those with potentially curable CRC (HR = 0.80, 95 % CI 0.63-1.01, non-significant) and for those with metastatic cancer (HR = 0.65, 95 % CI 0.47-0.89). Also private insurance was associated with improved outcomes (HR = 0.80, 95 % CI 0.67-0.96) among patients with metastatic disease at diagnosis but no difference was observed for those with stage A-C CRC. There was no evidence of improvement in survival for metastatic CRC patients over the study period, whereas there was a trend toward improved survival (reduced risk of death) over time for patients with stage A-C CRC (HR = 0.93, 95 % CI 0.90-0.97). Figure 2 shows the cumulative mortality plots for CRC mortality, adjusted for covariates and competing risk of death from other causes by neighbourhood SES, for all cases, and place of residence for stage A-C CRC.

This study demonstrates the utility of record linkage, using administrative and health surveillance databases, to evaluate disparities in outcomes for cancer patients. Record linkage offers considerable advantages over institutionally-based studies when examining disparities according to socioeconomic status or place of residence [39]. The inclusion of all eligible patients, regardless of whether they attended public or private health care services or both, greatly reduces the potential for selection bias to have influenced our results.

Our findings in relation to SES are based on area-level rather than individual level measures of SES, so may not accurately measure the level of socioeconomic disadvantage for individuals. Ideally, individual measures such as income, education level, housing and car ownership would be preferable, however these data are hard to obtain at a population level. Area level measures of neighbourhood deprivation are commonly used to identify socioeconomic disparities at the population-wide level. IRSAD is a rigorously constructed, composite measure of socioeconomic advantage and disadvantage based on census data for small areas which has been widely used to assess SES disparities in Australia [40]. Previous data linkage studies have been hampered by lack of staging data or incomplete information on treatments, particularly those provided through the private health sector. While we made use of staging data from a previous research project, cancer registries across Australia do not routinely collect stage. High priority needs to be given to developing sustainable systems to collect and record staging information within central cancer registries to enhance their value and usefulness in health outcomes research. Cancer Australia initiatives to pilot the collection of staging information in Australian cancer registries are promising developments [41].

Concerns are often raised about the quality and consistency of administrative data, particularly in relation to treatments and comorbidity [42]. Linkage across multiple data sources including hospital admission records, clinical registries and public and private radiotherapy services enhanced our ability to capture broad level treatment information. Our study would have been further enhanced had we had access to national pharmaceutical prescription data, through linkage with the national database, to better ascertainment of chemotherapy administered in outpatient and community settings [43]. The establishment of data integrating authorities, which is underway throughout Australia, will provide greater access to such data which are crucial for evaluating effectiveness of cancer care [44]. While we acknowledge the limitations of using administrative data sources to determine treatment information, undertaking case note reviews at a population level would be impractical and the cost prohibitive. Linkage derived measures are likely to be sufficient for epidemiological studies, though further research is warranted to valid measures and determine the impact of inaccuracies on population-based research.