The risk of colorectal cancer with symptoms at different ages and between the sexes: a case-control study

Cases were patients aged 30 years or older with a diagnosis of colorectal cancer between January 2001 and July 2006 and at least 2 years of full electronic records before diagnosis. Up to seven controls per case were selected (the maximum allowed within THIN regulations), using a computerised random number sequence. Controls were matched for practice, sex and age, but had not had colorectal cancer: they also had at least 2 years of records before the index date (the date of diagnosis of their matched case). The full date of birth was unavailable to protect anonymisation; however, the year of birth was available. Where possible, controls were matched to the same year of birth in years as cases; if no controls in this year were available, they were matched to within 1 year, then 2 years and so on up to a maximum of 5 years. THIN staff identified and provided complete records for all cases and controls.

From a review of the literature, 23 candidate variables (features) were identified, either a symptom, or an abnormal primary care investigation, or a predisposing risk marker such as obesity. We also identified codes for irritable bowel syndrome as a potential misdiagnosis. For some symptoms we were also able to take advantage of the availability of data on related prescriptions, for example, prescriptions for antidiarrhoeals and laxatives were obtained as possible surrogates for the relevant symptoms, and similarly antispasmodic drugs for irritable bowel syndrome. Features were designated as new if there were no similar symptoms or prescriptions observed previously in the 2 years before the index date. Haemoglobin values were categorised into 1 g/dl bands, and microcytosis defined as a mean red cell volume ≤ 80 fl. Weight loss was calculated from the change between the last recorded weight and the highest weight in the previous 2 years, separated into two categories: ≥ 10% weight loss or 5% to 10% weight loss. Patients were assigned to their maximum weight loss category. Obesity was defined as a body mass index > 30 kg/m² within 2 years of the index date. Diabetes was considered to be present if it had ever been diagnosed.

The main method of analysis was conditional logistic regression. For the variables that also had prescriptions as surrogates, preliminary analyses compared the odds ratio for a prescription without a record of the feature with the odds ratio for the feature itself. As these were similar for all three of this type of variable, constipation, diarrhoea and irritable bowel syndrome, the relevant pairs of variables were merged.

All variables with a univariable association with cancer significant with a P value < 0.1 were entered into a staged multivariable analysis. The first stage of the multivariable analysis grouped clinically related variables together. These were: intestinal motility features (constipation, diarrhoea, change in bowel habit and flatulence); pain features (irritable bowel syndrome, or prescription of an antispasmodic and abdominal pain); bleeding features (rectal bleeding, anaemia and microcytosis); systemic features (weight loss and thrombo-embolism); obesity features (diabetes and obesity). Only variables with a P value < 0.05 within their group were considered for the final model.

Age was stratified into four bands: 30 to 59 years (there were too few cancers in this band for meaningful sub-division) and 10-year bands thereafter, up to 80+ years. For each age/sex/feature combination we created a 2 × 2 table and calculated the positive likelihood ratio and its 95% confidence intervals. We used Bayes' theorem (posterior odds = prior odds × likelihood ratio) to estimate PPVs [19]. The prior odds were derived from national incidence rates stratified by age and sex for 2003 (the middle year of our cohort), in that an annual incidence of, for example, 5 per 1000, is the equivalent of 199 to 1 odds against having cancer diagnosed in the next year [20]. For males, the annual incidence rates used were: aged 30 to 59 years, 0.026%; 60 to 69 years, 0.19%; 70 to 79 years, 0.35% and 80+ years, 0.43%. Female rates for the same age groups were: 0.02%, 0.11%, 0.21% and 0.27%.

With the 5000 cases available, there was > 99% power to detect 5% having a particular feature, compared with 1% amongst controls. Such a difference was viewed in advance as being potentially clinically useful, and from a previous primary care study was considered possible [15]. Ethical approval was obtained from the London Multicentre Research Ethics Committee.

This is the largest primary care study to examine all colorectal cancer symptoms. Its size has allowed estimates of PPVs for relatively small age bands in either sex, and with reasonably narrow confidence intervals. Furthermore, the THIN database is representative of the UK as a whole, so our results should be generalisable. However, there are weaknesses too. We could not confirm cancer diagnoses, although it is unlikely that such an important diagnosis would have been entered erroneously particularly often. We also could not examine histology or staging, as these are only rarely recorded on primary care systems. The study relied upon good recording of symptoms. This problem is not quite as important as it appears, as under-recording of symptoms occurring proportionately in both cases and controls would not alter the likelihood ratios. Only if under-recording was especially prevalent in one group would a bias arise. It is impossible to know whether cases or controls would be more prone to under-recording. However, the PPVs are very similar to those derived from paper records [15] (these may of course also suffer from differential recoding of symptoms) and to the two previous studies using electronic records [13, 14]. One further limitation was the study design: by matching for age and sex, we could not examine these directly. Even so, the dataset was large enough to allow age-sex stratified analyses and estimation of PPVs.

Only three papers have used primary care data to estimate risks of colorectal cancer across the ages and between the sexes, although none with the size or precision of this one. The first examined GP records directly and showed an approximate tripling of PPV for rectal bleeding for those aged over 70 compared with those aged 40 to 69 years [15]. Two studies using electronic primary care records reported a rise in the PPV for rectal bleeding across 10-year age bands. This was also seen in the study reported here.

There is much less previous literature for other symptoms: one study gives similar figures for change in bowel habit [14], and the other again showed an approximate tripling of PPV for constipation, diarrhoea and abdominal pain between patients aged 40 to 69 years and those aged over 70 [15].

It is important to use primary care data to provide evidence for primary care decisions such as when to refer for investigation. Current referral guidelines essentially ignore age (other than giving a minimum age for referral) and sex (except giving a different threshold of haemoglobin for referral). As the incidence of cancer rises steeply with age, being 16 times higher in the oldest male age group and 13 times higher in the oldest female group when compared with the youngest groups, PPVs could have been expected to differ by as much. They do not, as the likelihood ratios fell with age. This presumably reflects the increasing prevalence of benign causes of colorectal cancer symptoms with age.

Two groups of symptoms emerge from the results. The highest PPVs are for rectal bleeding and change in bowel habit. For men over 60 years, rectal bleeding PPVs ranged from 2.4% to 4.5%; most would agree these figures are high enough to warrant investigation. For women, the figures were lower, but still in the 2% to 3% range. Change in bowel habit is less simple: GPs contributing to THIN must have been using this term very differently from the separate terms of constipation and diarrhoea, in that the PPVs for change in bowel habit were considerably higher. We cannot know what features led GPs to write change in bowel habit in the notes in preference to the specific motility symptoms, however, from our results it is clear that they were identifying a riskier feature, and one that warrants investigation.

In contrast, the PPVs for constipation, diarrhoea, abdominal pain and loss of weight were all below 1.5%, confirming that they are low-risk symptoms, at any age and in either sex. Indeed, this is the first study to show that the risk of colorectal cancer increases with increasing weight loss. This is no surprise (and is a testament to the quality of THIN data). However, these four symptoms are not 'no-risk' symptoms. This creates a problem for design of referral strategies. The high-risk symptoms of rectal bleeding and change in bowel habit were only recorded in 15.6% and 11.2% of cases, respectively. Even allowing for possible under-recording, it is highly likely that only a minority of patients with colorectal cancer have a high-risk symptom. The majority, with only a low-risk symptom, could have their diagnosis expedited in a number of different ways. Firstly, scoring systems such as the CAPER score may refine the risk in such patients [21]. Secondly, it is possible that biomarkers may be identified with adequate performance characteristics for use in such a population. Thirdly, early work suggests that measurement of rectal DNA may allow patients at high risk to be selected for investigation [22]. Without one of these initiatives, or a combination of them, patients with a low-risk symptom will continue to be at risk of delayed diagnosis, and possible emergency presentation [23].