We describe the observed comorbidity of the 20 jointly most common episodes of care in primary care populations from the Netherlands, Malta and Serbia. Out of 820 possible comorbidities, 573 (69·9%) showed at least one significant association from at least one of the three databases. Consistently significant sets of ORs were observed in only 76 (9·3%) cases. Only 32 (3·9%) sets contained at least one strong association. Significant inverse associations were rare (only 6 (0·7%) sets), and in only 2 (0·2%) cases was at least one of these inverse associations strong. We did not observe any set with contrasting significant ORs. Observed comorbidity seems very common, but inconsistent across the country databases, not in direction (negative versus positive) but rather in both strength and significance.
The high frequency of observed comorbidity with low consistency in both clinical and statistical significance between populations indicates more casual than causal associations. A causal relationship would be expected to be manifest more consistently across populations. Even in the minority of cases where the associations were consistent between countries and numerically larger, the associations were observed to weaken with increasing patient age.
Method and analysis
We defined comorbidity as all other episodes of care co-existing with an episode of care in a defined time period, that time period being one calendar year data-frame . Our definition is consistent with most literature .
ICPC is a standard instrument to measure the content of primary care, and is an accepted tool to measure comorbidity. In fact, the use of a classification allows more precise measurement of the relationships between unique distinct concepts [1, 4–7]. The granularity of ICPC is appropriate for the study of comorbidity in primary care populations, since the precision of incidence and prevalence estimates is improved with fewer classes (as compared to ICD-10), thus allowing narrower confidence intervals of estimates .
The episode of care data model and the practice of coding the symptom diagnosis, when appropriate, keeps disease classes clean, avoids over-estimation of the prevalence of chronic diseases and corrects for the effect of multiple consultations [3–7].
By applying strict limits for clinical and statistical significance one avoids describing spurious associations. The clinical and statistical significance thresholds we used represent a standard presentation which has been extensively validated [5–7, 9] and fits well with Bradford Hill’s requirement for strength of an association being an indicator of causality .
We considered Bradford Hill’s criteria for causality  and analysed them mathematically. As such our analysis of causality was based on the strength of an association, consistency across different populations), and a temporal relationship. We corrected our ORs for the prior probability, and demonstrated significant departures from the prior probability as recommended by other authors . We then analysed ORs in different age groups as a surrogate for biological gradient or dose-response to re-test our findings. We did not individually assess such associations for specificity, plausibility, coherence, experimental evidence or analogy, partly due to the large number of such associations, and also because our study was not one of individual associations, but rather of general trends.
Validation through extended comparison
We selected triplets of ORs which were consistently significant across three databases, with two of them strong, for further analysis. In fact, many of such comorbid pairs (K86 and T90; T93 and T90; T93 and K74; T93 and K87; T90 and K77; K74 and K77; K77 and K87) fit established medical literature on important comorbid conditions. We then proceeded to test these selected associations in different age groups.
The Bradford Hill criteria include a dose-response or biological gradient effect . One would expect ORs for a causal association to increase as patients are exposed to the disease/s of interest over time. We used age as a surrogate for the passage of time. We would thus expect the ORs in older age groups to increase for a causal association. We actually observed the opposite, with ORs falling in older age groups.
A lower OR in an (older) age group might alternatively be explained by an increased prior probability of either disease in that age group, and, consequently, also of the comorbidity. However, the OR would not decrease should the posterior probability also increase at the same, or at a larger, rate. This latter case would be expected with a causal relationship exhibiting a dose-effect. As such, this is an unlikely explanation for our findings.
A higher OR for comorbidity may manifest in younger age groups as a consequence of the early presentation of a serious disease which might trigger, in turn, more frequent visits to the FD , or increased surveillance or decreased diagnostic thresholds for other related diseases (such as for diagnosing hypertension or hyperlipidaemia in a diabetic). However we have corrected for the effect of multiple visits with the episode of care model, and we have also corrected for the prior probability of such an occurrence.
As such we consider the lack of a dose-response or age gradient as a strong indicator of the lack of a causal relationship. We conclude that the null hypothesis for causality is not rejected.
Arguably, the major reason for the large number of significant associations between the commonest EoCs in three populations is illness diversity: the increasing number of possible diagnoses in medicine and primary care due to new diagnostic entities. This has the effect of increasing the probability of such interactions. We conclude that most observed associations do not reflect actual causal relationships. Evidently, the utilisation of highly granular coding systems in primary care, especially those which do not separate unique concepts or allow multiple terming of individual concepts, runs the risk of worsening this artefactual comorbid landscape.
Our approach to studying comorbidity is consistent with prior definitions of comorbidity and with methods to assess both the strength and the clinical and statistical significance of an association [1–3, 9]. Our finding that most comorbidity is likely to be casual may be more controversial. However, such is supported by publications which have sought to confirm the reported causality of specific comorbidities and instead concluded that prior research had failed to adequately correct for chance and/or for the effect of multiple consultations over time . Studies which analysed comorbidity across a range of common health problems and/or different populations and/or captured episodes of care were rare. The Transition Project databases thus provide a unique perspective on comorbidity in primary care.
FDs are often selected to participate in research using EMRs, and may collect data at a higher level of detail and accuracy than their colleagues. Thus, the analysis of such data sacrifices some generalisability for increased depth, whilst accepting inherent features of family practice which cannot be adjusted for mathematically without introducing new systematic errors and biases. However, we have demonstrated elsewhere that such studies of EMR data are complementary to epidemiological surveys, and are not necessarily less valid or less generalizable [6, 7, 12].
The databases were collected for research purposes from selected practices, with the exception of Serbia. A comparison of data from more practices and more countries, had they been available, would have allowed a more powerful study. A key message is that more data are needed for such comparisons, and this research should be extended to other countries.
The error-trapping and coding support tools in the EMR TransHis, and the advantage of the classification used and the EoC data model have been previously described [4–7]. These qualities are a substantial strength, adding support for the study conclusions.
This is a study of comorbidity which does not focus on a small selection of diseases, but rather analyses data on many common diseases. As such it has significant advantages over studies which either focus on one index disease or collect data from secondary care.
Impact and future research
This study informs clinicians on the landscape of common comorbidities and allows more rational interpretations, discarding assumed causal relationships and helping to avoid over-treatment. Future research could focus on more sophisticated longitudinal analyses to specifically measure the change of comorbidity ORs over time to quantify the dose-response effect, if present.