Predictors of variation in colorectal cancer care and outcomes in New South Wales: a population-based health data linkage study

The study dataset included people who had an admission for an identified surgical resection procedure11 in one of 105 hospitals in NSW. Additionally, data were linked to six Area Health Service Clinical Cancer Registries (ClinCRs). The ClinCR contains items about cancer diagnosis, staging, and specialist referrals and consultations that provide additional depth to the data collected in the CCR and other datasets. ClinCR data are based on a minimum dataset collected voluntarily for notifiable cancers in public hospitals in six of the eight former Area Health Services in NSW since 2006, and have been shown to include 85% of people who underwent colorectal cancer surgery in a public hospital in 2008.11

We analysed six outcome and five care process measures in this study. Measures were chosen based on current clinical practice guidelines and literature review (with input from a study advisory group comprising clinicians, data managers and clinical coders), as well as an assessment of the feasibility of calculating them from available data.11 Measures were calculated for colon and rectal cancers separately. Rectosigmoid junction cancers were grouped with colon cancers as they have more similar rates of survival and recurrence.12 Outcome measures were calculated for the whole cohort; however, care process measures required data from the ClinCR, which are collected only in public hospitals (43 hospitals; 61.5% of patients). The available data for each measure and further information about their definition and calculation are reported elsewhere.13

Outcome rates by hospital were calculated using multilevel logistic regression models. Rates were adjusted for age, sex, Charlson comorbidity score,14 extent of disease, emergency admission, socioeconomic status (Socio-Economic Indexes for Areas score),15 and private health insurance status. Each hospital’s expected number of outcomes was obtained by summing their patients’ predicted probabilities of experiencing the outcome, given their covariate values. The observed number (O) of people who experienced the outcome was divided by the expected number (E) to create a ratio (O/E) that was multiplied by the overall NSW outcome rate to obtain the risk-adjusted rate for each hospital.16 Hospital rates for care process measures were unadjusted, as each measure should apply to all people defined in the denominator of the measure.

Funnel plots and multilevel regression models were used to quantify hospital variation. In the funnel plots, the outcome and care process rates for each hospital were plotted against hospital caseload. For each measure, 95% and 99.8% “control limits” were calculated based on the mean NSW rate.8 Hospitals outside of these limits were considered outliers.

Multilevel regression models take into account the clustering of patients within hospitals and allow the partitioning of variance between levels. The following two-level logistic regression models were fitted for each measure:

• a null model (hospital random effects but no explanatory variables);

• a model with fixed patient effects (as reported in Box 1) and hospital random effects; and

• a model with fixed patient and hospital effects (ie, caseload, private or public hospital, metropolitan or rural location) and hospital random effects.

Models were fitted using a Markov chain Monte Carlo approach in MLwiN (Centre for Multilevel Modeling). Intraclass correlation coefficients were calculated to compare the proportion of total variability attributable to hospitals after accounting for known patient and hospital factors.

Adjusted odds ratios (AORs) for patient and hospital factors were obtained from a maximum likelihood estimation approach based on adaptive quadrature in Stata (StataCorp).18 ORs for hospital factors are also presented as 80% interval ORs (IOR-80%), as hospital-level AORs can be problematic to interpret in multilevel models.19 An IOR-80% containing the value one indicates that the hospital characteristic does not add meaningfully to explaining hospital variation in the outcome.

Box 2 shows the number of outlying hospitals identified from funnel plots. Example funnel plots are shown in Box 3. More hospitals performed outside of control limits for care process measures than for outcome measures. The outcome measure with the most outlying hospitals was 90-day venous thromboembolism for both colon cancer (5% above 95% limits) and rectal cancer (8% above 95% limits). Recorded pathological stage (documented tumour–node–metastasis stage based on histological examination) was the care process measure with the greatest proportion of outlying hospitals (colon cancer, 41%; rectal cancer, 23%) (Box 2).

Hospital variance was much greater for care process measures than for outcome measures in the multilevel models, both before and after adjusting for patient and hospital factors (Box 2; and Appendix 1). Known patient factors accounted for most of the hospital variation in outcome measures. For care process measures, known patient and hospital factors generally did not help to explain much of the hospital variation. Recorded pathological stage showed the greatest hospital variation; estimates of total variability attributable to hospitals, after accounting for known patient and hospital factors, were 87% in the colon cancer model and 91% in the rectal cancer model.

Box 4 and Box 5 summarise patient and hospital-level predictors of outcome and care process measures. Complete results are available in Appendix 2. Greater extent of disease, comorbidity, resection after emergency admission and increasing patient age consistently showed strong evidence of an effect on outcomes. Having no current partner increased the odds of prolonged length of stay for colon (AOR, 1.8; 95% CI, 1.5–2.2) and rectal cancer patients (AOR, 1.7; 95% CI, 1.2–2.4).

There was less evidence of an effect of hospital-level factors on outcomes. Wound infection within 30 days was more likely in a rural or regional hospital compared with a metropolitan hospital (colon cancer: AOR, 1.9; 95% CI, 1.3–2.7; rectal cancer: AOR, 2.4; 95% CI, 1.4–4.0).

Care process measures had fewer consistent predictors than outcome measures. Younger people were more likely to be discussed at a multidisciplinary team (MDT) meeting and receive treatment within suggested time frames from the initial referral and the decision to treat. Greater extent of disease increased the odds of appropriate care for nearly all care process measures. Interestingly, males with rectal cancer were more likely than females to be treated within 62 days of initial referral (AOR, 1.9; 95% CI, 1.4–2.8).

Care process measures could only be compared between public hospitals, as these measures are based on data from the ClinCR. A patient with colon cancer treated at a metropolitan hospital or a medium caseload hospital was more likely to be discussed at an MDT meeting. However, the IOR-80% for these factors contains one, meaning that known hospital factors accounted for only a small component of the unexplained hospital variation in care processes.