Estimates of cancer incidence, mortality and survival in aboriginal people from NSW, Australia

Observational comparative cross-sectional study of the whole Aboriginal population of NSW with the non-Aboriginal population and with the NSW population overall, 1999-2007.

All (de-identified) records of invasive cancers diagnosed from 1999 to 2008 held on the NSW CCR were used in this study. Cancers included in the analysis were ICD-10 codes C00-C96, D45, D46, D47.1, D47.3 and non-melanoma skin cancer (C44) was excluded. In order to estimate incidence and minimise recording bias of Aboriginal status in NSW cancer data, logistic regression modelling of cancer records with non-missing Aboriginal status was used to inform a multiple imputation (MI) approach to assigning Aboriginal status to those records where Aboriginality is missing. Cancers diagnosed from 1999 onwards were chosen for imputation and analysis since the proportions of records with missing Aboriginal status were lowest from then (Figure 1). Cancer mortality data as recorded by the CCR from 1994 onwards are considered sufficiently reliable to use, but for this paper we have estimated Aboriginal cancer mortality for 1999-2007, to inform the survival analysis of cancers diagnosed over 1999-2007.

Essentially, MI is a means for representing uncertainty in missing data, by producing a distribution of plausible values for a missing variable in a record, given the values of that record’s non-missing covariates. The contribution of the non-missing covariates to the distribution of missing values is based on modelling the non-missing co-variates in complete cases to predict the likely value of the variable in records where it is missing.

The key assumption of most methods for handling missing data is that of Missing Completely at Random (MCAR), which essentially means that the observations that are missing are a simple random sample of all observations. This is clearly not the case for Aboriginal status.

However, MI requires only the more relaxed assumption of Missing at Random (MAR), which means that the missing data is a random sample of all the data after adjusting for all other observed (non-missing) variables [5]. In other words, systematic differences between missing cases and the complete cases are allowed so long as all (or most) of the differences can be explained or modelled by the other observed variables. If the proportion of missingness in the records is low then multiple imputation can be used [6].

Once a model predicting values for a missing variable is constructed, Markov chain Monte Carlo (MCMC) simulation is used to create a small number of independent draws (typically 5-10) of records with a range of imputed values from the predictive distribution. Each draw represents a random sample of the missing values and is then used for multiple-imputation inference such that a number of data sets (equal to the number of draws) of ‘complete’ cases are analysed using standard statistical analyses. The results from the ‘complete’ data sets are then combined to produce inferences that account for multiple records of the same individual from the draws. The overall estimate is the average of the individual estimates, and the overall standard errors include both between and within imputation standard errors. Standard errors for the missing variable reflect the uncertainty due to the missing values, leading to valid statistical inferences.

MI is particularly suitable for estimating Aboriginal cancer rates, as there are numerous covariates available in the NSW CCR (such as area of residence, sex, age, country of birth and clinical variables) which make it quite likely that the MAR assumption would be met. For cancer mortality, missing Aboriginal status has been very low on the NSW CCR since 1999, ranging between 0.2% and 2.8% over 1999-2007 (cf. Figure 1). Accordingly, multiple imputation was applied to cancer incidence data only.

The statistical modelling approach used to develop the imputation model for the MI procedure was logistic regression, since the missing variable is a binary quantity. No covariates other than Aboriginality had missing values, which simplified the estimation procedure considerably. Covariates included in the model were 5-year age group, sex, country of birth, cancer degree of spread at diagnosis, clinical grouping, one-year survival, Area Health Service of residence at diagnosis, and percentage of the local government area population identifying as Aboriginal in the most recent population census. Area Health Service was included to capture geographical remoteness and differences in health resources and access to health services. The logistic regression model was fitted to all cancer data with non-missing Aboriginal status for 1999-2008, treating Aboriginal status as a binary outcome (1 = Aboriginal or Torres Strait Islander, 0 = otherwise).

For the logistic regression modelling and imputation, we used Proc Logistic, Proc MI and Proc MIAnalyze in SAS/STAT^® software, version 9.2. Directly age-standardised cancer incidence and mortality rates were calculated using the 2001 Australian census population as the standard. Indirectly age-standardised incidence and mortality ratios were calculated using cancer incidence and mortality expected in the whole NSW male or female population, as appropriate, for 1999-2007. Cumulated risks of cancer incidence and mortality to age 75 years were calculated according to the method of Day [7]. Cox proportional hazards regression modelling was used to produce cancer-specific hazard ratios by Indigenous status by sex, adjusted for age and year of diagnosis, and cancer degree of spread at diagnosis.

A sensitivity analysis was conducted in which 0-5% random samples of all cancers with missing Aboriginal status were assigned as Aboriginal and overall Indigenous cancer incidence estimated with these added to cancers recorded as Indigenous.

The Cancer Institute NSW is the organisation with delegated authorisation to manage the NSW Central Cancer Registry on behalf of the NSW Ministry of Health. The management of this registry includes assessment, review, maintenance and enhancement of the quality of all data held in the registry. The work for this project was undertaken within this ambit. In addition, no identified individual case information was utilised. Therefore formal ethical approval was not required to undertake this work.

As Figure 1 indicates, Aboriginal status on cancer registration in NSW has improved considerably since the mid-1990s: for cancer deaths, missing Aboriginal status affects 1-2% of cases diagnosed since 1999 (the rise for 2008 is artefactual due to CCR death information yet to be fully updated from the Australian Bureau of Statistics). For still-living cancer cases, the proportion with unknown Aboriginality status declined sharply from 75% in 1996 to 30% in 1999 and steadily thereafter to under 20% by 2008. For all registered cancers, the proportion with missing Aboriginal status varied between 12% and 14% over 1998-2007. While there is still a substantial proportion of missing Aboriginal status in live cancers this is not so great as to preclude multiple imputation. From the perspective of care episodes, proportions of missing Aboriginal status in alive cancer cases are generally around 90% if no inpatient hospital admission occurs, but declines to 10-20% by the first hospital admission for the cancer, then to negligible levels by the fourth admission (Figure 2).

From the logistic regression model applied to cancer records with known Aboriginal status, most of the odds ratio estimates were statistically significant (Table 1). All age groups of cancer diagnosis younger than 75+ years were positively and significantly associated with Aboriginality, as were distant and regional degree of cancer spread, death within 1 year of diagnosis, percentage of Aboriginal population in the local government area of residence (at diagnosis), and non-metropolitan NSW Area Health Services with relatively high Aboriginal population compositions. Significant negative predictors included country of birth outside Australia, and most cancer clinical groupings (using respiratory cancer as the referent). The concordance from the logistic regression model was 74.4% and the percentage discordant was 20.0%, leaving 5.6% tied. The resulting area under the ROC curve of 0.772 suggests that the model correctly predicts Aboriginal status in most cases.

After imputation, 146 additional cancers diagnosed in 1999-2007 were assigned as Aboriginal in males (12.2% more than recorded), and 140 cases to females (12.5% more than recorded). Mean annual Aboriginal cancer incidence was estimated to be 660 per 100,000 in males and 462 per 100,000 in females, equivalent to 9% and 6% higher than all NSW males and females respectively, as measured by standardised incidence ratios (Tables 2 and 3). Head and neck (SIR = 2.05, males and 2.01, females), oesophagus (1.85 and 2.16), liver (1.67 and 2.21) and lung (1.93 and 2.43) were significantly higher in both sexes. Males also had significantly higher incidence rates of cancers of unknown primary origin (SIR = 1.63) and stomach (1.85). Females had a significantly higher incidence of cervical cancer (SIR = 2.43). Male and female Aboriginals had significantly lower incidence of melanoma (SIR = 0.44, males and 0.47 females). Incidence of prostate cancer (SIR = 0.82) was also significantly lower than for all NSW males, and thyroid cancer incidence was significantly lower in Aboriginal females (SIR = 0.54). Prostate cancer is the most commonly diagnosed cancer in Aboriginal males (145 per 100,000), followed by lung cancer (111 per 100,00) and large bowel (75 per 100,000). Breast cancer is the most common cancer diagnosed in Aboriginal females (116 per 100,000), followed by lung (69 per 100,000), and cancer of the large bowel (56 per 100,000). The cumulative risk of cancer over ages 0 to 75 years in Aboriginal males was 1 in 2.6, compared to 1 in 2.8 for non-Aboriginal males; in Aboriginal females the cumulative risk was 1 in 3.5, compared to 1 in 3.8 for non-Aboriginal females.

The age distribution of cancer incidence in Aboriginal people shows most of the excess over non-Aboriginal people to be occurring in the oldest age groups, manifesting by age 70+ years (Figure 3). By age 75+ years cancer incidence in Aboriginal people is approximately 30% and 40% higher in males and females respectively.

Mean annual cancer mortality in NSW Aboriginal people is estimated to be 373 per 100,000 in males (from 657 deaths over 1999-2007), and 240 per 100,000 in females (from 572 deaths; Tables 4 and 5). Cancer mortality in Aboriginal people was 68% and 73% higher than in all NSW males and females respectively, as measured by standardised mortality ratios. The higher excess mortality over incidence indicates considerably lower cancer survival. Outcomes for Aboriginal males with prostate cancer are particularly poor: despite 18% lower incidence of prostate cancer than for NSW overall, Aboriginal males have 86% higher prostate cancer mortality. Mortality from cervical cancer, one of the few easily preventable, was 374% higher in Aboriginal than all NSW women, considerably surpassing the 143% excess incidence of cervical cancer (Table 5). Aboriginal mortality is substantially higher than NSW from several other key cancers including lung (SMR = 1.99, males and 2.73, females), female breast (SMR = 1.54). Even for melanoma, with standardised incidence ratios of 0.44 for males and 0.47 for females, the standardised mortality ratios were not significantly different from 1 (0.67 in males, 1.00 in females).

Lung cancer is the leading cause of cancer death in Aboriginal males (91 per 100,000), followed by prostate cancer (57 per 100,000) and large bowel (39 per 100,000) (Table 4). Lung cancer is also the most common cancer cause of death in Aboriginal females (58 per 100,000), followed by breast (36 per 100,000) and large bowel (24 per 100,000) (Table 5). The cumulative risk of cancer mortality over ages 0 to 75 years in Aboriginal males was 1 in 4.7, compared to 1 in 7.5 for non-Aboriginal males; in Aboriginal females the cumulative risk of cancer mortality was 1 in 6.5, compared to 1 in 11.1 for non-Aboriginal females.

As for incidence, the age distribution of cancer mortality in Aboriginal people shows most of the excess over non-Aboriginal people occurring in the oldest age groups, except the excess mortality manifests earlier at 60+ years (Figure 3). By age 75+ years cancer mortality in Aboriginal people is approximately 80% higher in males and 75% higher in females.

Differences in cancer stage at diagnosis are evident but indicate only some of the sources of differences between Aboriginal and non-Aboriginal cancer mortality. Incidence of cancer with regional or distant degree of spread is significantly higher in Aboriginal males and females than non-Aboriginal counterparts, and mortality correspondingly is significantly higher (Figure 4). However, while incidence by localised cancer in Aboriginal people is somewhat lower (non-significantly), and is non-significantly higher for cancer of unknown degree of spread, Aboriginal mortality from localised and unknown degree-of-spread cancers is significantly higher. This can be seen more clearly by examining mortality to incidence (M:I) ratios. In Aboriginal males the M:I ratio for localised cancer is 0.32 compared to 0.21 in non-Aboriginal males; in Aboriginal females the M:I ratio is 0.23, compared to 0.12 in non-Aboriginal females. For regional degree-of-spread cancers, these ratios are 0.32 versus 0.17 in males, and 0.25 versus 0.14 in females; for metastatic cancer the M:I ratios are 0.49 versus 0.28 in males, and 0.32 versus 0.19 in females; and for cancer of unknown origin the M:I ratios are 0.52 versus 0.32 in males and 0.25 versus 0.16. Thus not all the excess mortality from cancer in Aboriginal people is explained by higher cancer stage at diagnosis. Mortality outcomes for Aboriginal people diagnosed with cancer of the same degree of spread are significantly and substantially worse than in non-Aboriginal people.

Survival from cancer in Aboriginal people is worse: at 5 years, 35% of non-Aboriginal people in NSW have died from their cancer, compared to 47% of Aboriginal people (Figure 5, Table 6). When viewed as hazard ratios, adjusted for age and year of diagnosis and cancer degree of spread, Aboriginal people had an estimated hazard ratio of dying from cancer of 1.62 (p<.0001). In males, 51% of Aboriginal cancer cases have died of the cancer at 5 years compared to 36% of non-Aboriginal cancer cases. The corresponding male hazard ratio was 1.65 (p<.0001). For females the proportions surviving 5 years were 43% versus 33%, and the corresponding female hazard ratio was 1.59 (p<.0001). Cancers showing significantly lower survival in Aboriginal people include head and neck (males), stomach (females), large bowel and lung (males and females), breast and cervix. Of note, the largest Aboriginal/non-Aboriginal difference in 5-year cancer survival is in head and neck cancer among males (41% versus 67%), but in females this is identical (68%). Notable also are differences in survival from lung cancer: a similar proportion (91%) of Aboriginal males and females with lung cancer had died of it after 5 years, compared to 85% of non-Aboriginal male lung cancer cases and 81% of non-Aboriginal female lung cancers.

If Aboriginals comprised 1% of cancer records with missing Aboriginal status missing, then the estimate of cancer incidence is 735 per 100,000 males and 481 per 100,000 females, with corresponding male and female SIRs of 1.16 and 1.07 and cumulated risks to 75 years of 1 in 2.5 and 1 in 3.4 respectively (Table 7). These estimates exceed those imputed here and of course increase with higher proportions of cancers with missing Aboriginality assigned as Aboriginal (cf. Tables 2 and 3).