In this study, we estimated the age-specific joint prevalence of all pairs of four of the most prevalent chronic diseases in the Dutch population. Co-occurrence of all disease pairs studied was seen to increase with age. The joint prevalence was highest for diabetes and AMI, while cancer and stroke co-occurred the least frequently. For all pairs not including cancer, co-occurrence was more frequent for all age groups than expected when the individual diseases occur independently. Thus, observed minus expected proportions increased with age, while the corresponding observed/expected ratios became smaller. This implies that although at lower ages co-occurrence is less prevalent, at lower ages chronic diseases tend to cluster more within individuals. Diabetes co-occurred frequently with stroke and AMI, which is in line with what is known about the increased risk for these diseases in diabetics. On a relative scale, as measured by the observed/expected ratio, stroke and AMI co-occurred most frequently. This is not surprising, as both events are related in their etiology and share multiple underlying risk factors such as high blood pressure, cholesterol, smoking, and obesity. Cancer, however, seemed to display a somewhat different behavior: within an age range of approximately 60 to 75 years, it co-occurred less frequently than expected with the other three diseases. This pattern was somewhat unexpected and not easy to interpret. On the face of it, this seems to imply that diabetes and cardiovascular diseases "protect" against cancer and/or vice versa. However, to the best of our knowledge, there is no known patho-physiological mechanism that could explain such a relation. Alternatively, it could be a "survivor" effect: those prone to develop both diseases die of cardiovascular disease before reaching the age at which cancer would become symptomatic. Yet another explanation might be that people adapt their lifestyles after being diagnosed with cancer. It needs to be stressed that the uncertainty in these estimates is large, and that the findings regarding cancer can be due to chance. Although interesting, at this point we cannot attach much significance to this observation.
A limitation of our study was that the institutionalized population was not included in the survey. As the prevalence of chronic diseases is probably higher among those institutionalized [22], this exclusion is likely to have led to some degree of underestimation of the prevalence of co-occurrence of chronic diseases. Furthermore, the response rate in this survey was not much more than 60%, which is a potential source of bias. Also, the self-reported nature of the data may have induced some bias in different ways. First, people might not accurately report their disease status. However, previous studies showed that self-reports of chronic conditions were fairly accurate, suggesting that this form of bias probably remained limited [23, 24]. Second, if nonresponse was related to disease status, bias would result. To investigate whether this was the case, we compared our estimates to other national representative estimates of diagnosed disease prevalence [1, 25, 26]. Although estimates of cancer and diabetes prevalence were very similar, our estimates for AMI and stroke appear to be high. This could possibly be explained by a less-stringent case definition. Third, even if people report accurately and there is no selective nonresponse, undiagnosed cases will be missed. In case of diabetes, it has been argued that for every diagnosed diabetes case, there may be around 0.5 to one undiagnosed case. Thus, the true prevalences include, depending on the type of disease and other factors, variable proportions of people with no current morbidity or disability. Although there has been an upward trend in the ratio of undiagnosed/diagnosed cases of diabetes in the Netherlands [27], there are no recent observational studies in the Netherlands that have presented estimates of disease co-occurrence among diabetics. If there still is substantial underdiagnosis, we hypothesize that having a diabetes diagnosis is more likely in people with comorbidity. This would imply that our estimates of observed/expected ratios could be too high. Other limitations of our analyses are that cancer was treated as a single entity, whereas it is heterogeneous condition, and that no distinction was made between diabetes Types 1 and 2.
A few remarks are necessary regarding the method we used in modeling the joint presence of two diseases in the same individuals. Most importantly, we aimed at expressing prevalence as a function of age. With two diseases there are four possibilities. Hence, the outcome variable has a multinomial distribution, which we related to age using P-splines. The advantage of P-splines compared to polynomial regression is that model fit at the lower ages is not influenced by that at higher ages, and vice versa. That is, P-splines can be seen as a form of "local" regression. Furthermore, with P-splines it is not necessary to choose a more or less arbitrary number of knots, which is often seen as a drawback of other types of splines, such as B-splines. The choice of the smoothing parameter(s) for P-splines is data driven. In our analyses, we used the Akaike information criterion (AIC) criterion, which was also used by Eilers and Marx [21]. In Additional file 1, results are shown when the Bayesian information criterion (BIC) is used to find the optimal smoothing parameter. In general, the results are similar, but a bit smoother and less wiggly when the BIC criterion is used compared to the AIC criterion. For the absolute co-occurrence prevalence, the estimates do not differ much between the AIC and BIC. However, for the observed/expected ratios, there is a clear influence at lower ages, in which the prevalences are generally low. The observed/expected ratios at those ages are much higher if the BIC is used. Finally, it should be noted that in order to increase power, we combined both sexes and pooled all years. This means that the estimates are time- and gender-averaged. Stratifying the analyses by sex and analyzing time trends would therefore be a next step.
Although the "clustering" of chronic diseases is not a surprise, quantitative data on multimorbidity are scarce. Especially at older ages, the co-occurrence of chronic conditions starts to become so common that individuals with more than one disease can no longer be considered the exception. This not only has consequences for disease management programs, but also guidelines should more explicitly address the issue of comorbidity than has hitherto been done. The fact that the care for this category of patients poses specific difficulties requiring a distinctive approach is still insufficiently recognized. A better appreciation of the epidemiology of multimorbidity is a first step to bring the magnitude of the problem into focus. A noteworthy point of our study is that we have presented estimates combining cancer with noncancerous diseases. Although cancer incidence and prevalence are usually well-monitored by cancer registrations, these are not often linked to noncancerous diseases.
In conclusion, in this study we quantified age-specific co-occurrence patterns. It is clear that with increasing age, multimorbidity becomes common. More importantly, the prevalence of multimorbidity most of the time is much greater than would be the case if diseases occur independently from each other. Thus, the practice in epidemiological and public health research to monitor individual diseases tells only part of the story. With an aging population, it is important to quantify the problem of multimorbidity. Those involved in the management of care, the drafters of guidelines, and the doctors treating patients with more than one disease should develop strategies to improve the care for this category of patients that is becoming more numerous as the population ages.