The National Cancer Data Repository (NCDR) contains information about every patient diagnosed with cancer in England and allows their treatment pathway to be mapped from diagnosis to cure or death. It consists of linked cancer registry, Hospital Episode Statistics (HES) and National Bowel Cancer Audit Project (NBOCAP) data.
Information was extracted from the NCDR on all individuals who underwent a major resection for primary colorectal cancer (International Classification of Diseases 10th revision C18–C20) diagnosed between 1 January 2006 and 31 December 2008. Information on age, sex, tumour site, date of diagnosis, Index of Multiple Deprivation (IMD) income category (based on postcode at diagnosis) and modified Dukes' stage at diagnosis were extracted from the registry data component of the NCDR. Modified Dukes' stage was used as, over the time period of this study, this was the only staging information captured both by English cancer registries and by the NBOCAP. Information about patient management, including operation type, approach to surgery and hospital of treatment, was derived from HES. If data on modified Dukes' stage at diagnosis or approach to surgery were missing from the HES and cancer registry data in the NCDR, this information was taken from the NBOCAP data set. Standard methods were used to identify whether each patient underwent a major resection for colorectal cancer up to 1 month before or 12 months after the date of diagnosis5,6
Patients undergoing laparoscopic operations were identified as those with Classification of Interventions and Procedures version 4 (OPCS-4) codes indicating minimal access to abdominal cavity (Y75) or other specified approach to abdominal cavity (Y508) recorded on the same date as the major resection. Converted laparoscopic operations were identified as those with an OPCS-4 code indicating failed minimal access approach converted to open (Y714). Information on approach to surgery was also incorporated from the NBOCAP data set.
A Charlson co-morbidity score7
was calculated for each individual based on diagnostic codes (excluding cancer) recorded for any hospital admission in the year before diagnosis of the colorectal tumour, excluding any admission spanning the date of diagnosis. The cancer component of the Charlson index was derived from the cancer registry information in the NCDR. Patients were grouped into Charlson score categories of 0, 1, 2 and at least 3, higher scores indicating greater co-morbidity.
The urgency of surgery is known to have a strong prognostic impact on outcomes, but this information is not recorded routinely in HES. The method of admission is, however, available. Patients who were admitted as an emergency and underwent surgery within 2 days of admission were deemed to have undergone emergency surgery.
The proportion of procedures performed via open, laparoscopic or converted surgery were examined in relation to patient age, sex, year of diagnosis, modified Dukes' stage of disease at diagnosis, tumour location, IMD category, operation type and Charlson co-morbidity score. Factors associated with the use of laparoscopic surgery were also investigated using a hierarchical random-effects binary logistic regression model, fitted using Stata® Statistical Software Release 11 (StataCorp LP, College Station, Texas, USA). The model was built with a hierarchy of patients clustered within hospitals (level 2), so allowing for correlations between patient outcomes. Co-variables (explanatory variables) in the risk-adjusted model included age (per 10-year increase), sex, tumour site, IMD income category, year of diagnosis, stage at diagnosis, Charlson co-morbidity score, operation type (elective or emergency) and operative approach. Approach to surgery was categorized as open or laparoscopic; converted operations were included in the laparoscopic group on an intention-to-treat basis. Some case-mix information (such as stage of disease and socioeconomic deprivation category) was missing from the NCDR as it was not recorded routinely in the database.
Analyses restricted to patients with complete data would not have allowed the overall outcome to be assessed. Missing data were, therefore, imputed deterministically using the ICE command8
, with passive and substitute options and ordered logistic regression for ten imputations and ten cycles of regression switching. It was assumed that the data were ‘missing at random’. The imputation model consisted of 30-day postoperative mortality, survival time, length of hospital stay, age at diagnosis, sex, median annual workload of the hospital, modified Dukes' stage, IMD income category, operation type (elective or emergency), admission type (elective or emergency), year of diagnosis, year of operation, method of access (open, laparoscopic completed, laparoscopic converted), Charlson co-morbidity score, tumour site, hospital and cancer registry. For comparative purposes the models were built using both the imputed data set and a data set restricted to cases with complete data.
To investigate the relationship between laparoscopic treatment and the outcomes postoperative mortality, long-term survival and postoperative length of hospital stay, logistic, Cox and linear regression hierarchical random-effects models were fitted. Thirty-day postoperative mortality was defined as death within 30 days of major resection. Survival time was calculated from the date of major resection to the date of death or when censored (30 June 2010). Length of stay was defined as the number of days from major resection to the end of the associated hospital stay (calculated taking into account transfers between different hospitals). Length of stay was log-transformed before analysis with estimates back-transformed and interpreted as length of stay ratios. Length of stay values of less than 1·00 indicate a shorter stay, values greater than 1·00 indicate a longer stay, and values of 1·00 indicate no change in the duration of hospital stay due to the variable of interest.