Overall, differences in prevalence estimates among the different sources in a region were lower than the differences between regions, and differences observed among regions were similar across data sources. The fact that independent sources of data showed consistent values across different regions supports the claim that they correspond to actual population measures. In case systematic differences were observed, they could be interpreted as being due to differences in data collection and associated to demographic and disease characteristics. This provides evidence that administrative data actually measure a population phenomenon that can be interpreted and supports the use of administrative data for surveillance of geographical trends of the diseases in study, with the possible exception of COPD.
In the case of diabetes mellitus, the observed concordance between estimates from GP data and from survey data confirms previous reports [24, 25]. Estimates from GP data were systematically higher than estimates from administrative data. According to reports from other countries , the difference is likely to be due to the proportion of patients who, although being diagnosed with diabetes to the knowledge of their GPs, have mild or well-controlled disease and thus have never had either a hospital admission or a prescription for antidiabetic drugs, and have not received an exemption from copayment of diabetes-related healthcare, therefore escaping the algorithm for administrative databases. Indeed, when the subset of patients undergoing therapy with antidiabetics in the previous two years were extracted from both administrative and GP data sources, the pairs of prevalence estimates almost coincided in all of the regions, with one exception. Estimates adjusted for completeness of ascertainment, on the other hand, provided slightly higher estimates, a finding consistent with a previous study with similar data in another Italian area .
Ischaemic heart disease being congruently estimated by administrative and GP data in all of the regions is an unexpected finding. Angina, a less severe form of the disease, does not lead per se to a hospital admission, and few cases (less than 5%) are detected by the registry of exemptions from copayment. As around 30% of cases are detected only by dispensings of nitrates (data not shown), we observe that nitrates therapy is probably specific in detecting cohorts bearing this condition, as otherwise data would have been less consistent across regions in matching the diagnosis-based figures from GP clinical databases.
Heart failure was underestimated by GP data, although non significantly in the majority of the regions. Underestimation was highest in the oldest age band available in both data sources (85-95), where the prevalence is highest. This is consistent with the hypothesis that GPs belonging to HSD might occasionally perform less accurate data collection when visiting patients at home  or in residential care , or consider heart failure as a complication of other underlying conditions, such as ischaemic heart disease, rather than as a disease of its own. This would imply that the population detected by administrative data had a more severe form of the disease and was more often affected by disability. Another possibility is that administrative database overestimate prevalence because of lack of specificity of the case ascertainment algorithm. Indeed, according to a recent review of validated algorithms for case ascertainment of heart failure , algorithms using secondary discharge diagnosis showed lower positive predictive value (PPV) in several countries.
For COPD administrative data failed to detect the differences between regions that the other two sources consistently measured. Ascertainment of COPD from administrative sources has been shown to be challenging in other studies [29, 30]. In this case, the algorithm detected a particular pattern of drug prescriptions, combining duration, intensity and ATC class, that had been identified through a consensus process in a group of experts that was reported in Anecchino et al. . Although the pattern was specifically meant to avoid misclassification (e.g., with respect to asthma), it is possible that the conclusions of the study were in fact specific for the geographic area where the experts worked.
In light of the limitations of the sampling design of our study, the overall good agreement with other data sources supports a fortiori validity of chronic disease surveillance using administrative data in the regions that were involved in the study. However, support for external validity of our results needs to be discussed. Althought we only collected data from few geographical areas, the same administrative data are available for the whole national population. We are in fact not claiming that administrative data from few geographically sparse areas can be used to estimate national prevalence of chronic diseases, but rather that administrative data seem to be consistently able to detect prevalence of some chronic diseases around the area they were extracted from. Our positive findings (treated diabetes, ischaemic heart disease, heart failure) are indeed probably due to the fact that typical health consumption patterns of such chronic patients are similar across regions. On the assumption that regions of the same macroarea of the country (North, Center, South) are similar the one to the other to this respect, our data support the claim that estimates relying on the same algorithms should prove to be similarly effective. However, in some specific critical areas of the country where incomplete administrative data collection is suspected, a local evaluation is recommended.
Cohorts can be selected from administrative databases to perform population-based studies on patients with chronic diseases through further record-linkage with the same databases. This study cannot provide analytical tools to assess the limitations of the findings of such studies. However, no evidence emerges for major bias, except in the case of COPD, where regional differences with the other data sources are likely to be due to differences in the characteristics of the corresponding local cohorts.
The first limitation of studies that make secondary use of existing healthcare data sources is that only prevalence of diagnosed cases is taken into account, and underestimation of actual population prevalence cannot be estimated .
An implicit assumption of both crude and adjusted rate estimation from administrative databases performed in this study was that PPV of the case detection was 100%, an assumption that we could not verify and that is not to be taken for granted, when, for instance, secondary discharge diagnosis or drug utilisation with no indication is used as a source of case ascertainment. Ecological validation studies cannot directly resolve this issue, as consistent ecological estimates between a data source and a reference gold standard might as well be due to coincidental inclusion of false positive and exclusion of false negative cases. Only validation studies perfomed using individual-level comparison with a gold standard could assess PPV and sensitivity.