Can we use the pharmacy data to estimate the prevalence of chronic conditions? a comparison of multiple data sources

One of the most important aim of public health is to provide an accurate evaluation of the population health conditions, its need for care and related costs.

Usually, the estimate of prevalence for the most common chronic conditions (CCs) is calculated using direct methods such as prevalence surveys [1] but also indirect methods using health administrative databases that collect this information for other reasons were used [2].

Ideally, prevalence surveys that estimate the prevalence of CCs by a clinical evaluation, and not only by self-reported information from subjects should be performed. However, they are expensive and when performed were limited to elderly and in specific geographical areas [3, 4].

Prevalence surveys based on self-reported information are regularly conducted in several countries to provide estimates for several CCs [1, 5]. Some of these surveys present the advantage to be not particularly expensive but, at the same time, they are criticized because the presence/absence of the disease is self-referred and thus conditioned by potential bias. Furthermore, these surveys refer to a sample of the population and thus are also limited by the sampling uncertainty. In particular, these estimates could be biased because some individuals likely might not be reached by the survey (e.g., very old people living in retirement homes).

As far as the use of health administrative databases to estimate the prevalence of some diseases, hospital discharge registries are those more often used because they collect specific information about diagnoses [6]. However, in some cases the accuracy of diagnostic code can be low [7, 8]; furthermore, for some diseases the probability of being hospitalized, also for a long period, is very low and thus it might underestimate the actual prevalence.

The health administrative database of the general practitioners (GPs) has also been used to estimate prevalence given that for some conditions it is likely that a subject with the studied disease may be in charge of the GPs [9, 10]. However, GPs are not formally requested to collect specific databases with information about diseases and they collect data quite exclusively for facilitating their routine management such as drug prescriptions, doctor's notes, et cetera. This means that the quality about diagnosis may be heterogeneous; furthermore, for some CCs, GP has likely very few contacts with the patients; finally, at least for Italy, the access by public health services to GP's databases is impossible given that there are no statutory compliances for that.

Recently, the use of drug prescription database has been proposed to estimate the prevalence of specific CCs [11, 12]. This can be done when the drug prescriptions are unambiguously used for the treatment of these diseases (e.g., insulin for diabetes mellitus). In Italy drug prescriptions are collected at regional level and the coverage is expected to be extremely high because they are used for reimbursement by the regional health service (RHS).

The objective of this study is to provide estimates of prevalences of people diagnosed with several CCs in Lazio region, Italy, in 2006 using the drug prescription's database and to compare these estimates with those obtained using other health administrative databases. These prevalence estimates were also compared, when possible, with that reported by the survey performed in 2004-2005 by the Italian bureau of census (ISTAT) [1].

From pharmacy data we identified drugs specific for 20 CCs (table 1). In 2006, about 2.5 million people (48% of the entire population) had prescribed reimbursed drugs for one or more of these CCs; these people were about the 72% of the individuals who had prescribed at least one drug reimbursed by the regional health system.

All drugs identifying the 20 CCs represented 61% of the entire volume of packages. Drug expenditure for these 20 CCs was about eight hundred million of euro, corresponding to 57% of the total expenditure (data not shown).

Table 2 shows the number of subjects, the estimated prevalences, the volumes of prescriptions, the mean annual cost per individual for treatment, and the total cost. About 23% of the population received a drug for treating a cardiovascular disease, 9% for treating a rheumatologic conditions and then diminishing for other CCs. Table 3 compares the estimates of prevalences of the 20 CCs identifiable by the pharmacy data with those estimated by the ISTAT survey of 2005 and those obtained using the HIS and the REP, while the Figure 1(A) and 1(B) show the estimates of prevalences by sex using the four different sources.

The estimated prevalences using the pharmacy data were usually higher than that obtained with one of the other sources. For thirteen CCs at least one of the other sources provided higher estimates; using an arbitrary cut-off of +15% in the relative difference between pharmacy data estimates and the maximum of the estimates from the other sources (last column of table 3), for 12 CCs the estimates were more greater than those obtained using the pharmacy data. It is of note that for 11 CCs the relative difference was higher more than 50% of that obtained with the pharmacy data.

Regarding the comparison with the ISTAT survey it is of a good agreement for cardiovascular disease, diabetes and thyroid disorder whereas for seven CCs (i.e., rheumatologic conditions, chronic respiratory illnesses, psychiatric diseases, Paget's disease or other osteoporoses, migraine, Alzheimer's disease, and cirrhosis) the prevalence estimates were 70% or more higher than those obtained by the pharmacy data. Estimates of prevalences were usually lower than those of the pharmacy data when derived by the HIS (but malignancies, Alzheimer's disease, chronic renal diseases, HIV/AIDS and cirrhosis) and by the REP (but malignancies, chronic renal diseases, and chronic hepatitis and selective malignancies). Also stratifying by sex there were similar results.

This study evaluated the possible use of pharmacy data on identifying individuals with several CCs in Lazio region, Italy.

We found that the highest prevalences of people diagnosed with CCs were for cardiovascular diseases and rheumatologic conditions. These results are coherent with previous estimates in analogous studies performed in another Italian region and in US [10, 23]. The approach of measuring the prevalence of CCs using pharmacy data provided reliable estimates for diseases particularly impacting the health and social services such as Parkinson and Alzheimer disease. Prevalence estimates of these diseases using pharmacy data were comparable to those found in other European studies [24‐26].

Data were then compared in terms of prevalence with other health administrative databases and with prevalence estimates obtained by the ISTAT survey. Assuming that all sources correctly identified each specific CCs, we observed that for several of these CCs the pharmacy data was better on identifying cases. This was particularly pronounced with respect to the HIS and to the REP that only in few cases provided higher estimates than the pharmacy data. With respect to the prevalence estimates by the ISTAT survey we highlighted that for some CCs the prevalences estimated by the pharmacy data had a quite good level of agreement (i.e., cardiovascular diseases, diabetes, thyroid disorders, malignancies, cirrhosis). For rheumatologic conditions, chronic respiratory illnesses, psychiatric diseases, osteoporosis, migraine the agreement was very low and the ISTAT survey provided extremely higher estimates of prevalences compared to those obtained with the pharmacy data.

One possible explanation for this discrepancy between pharmacy data and ISTAT survey is that the latter measured the self reported CCs and several studies suggested that the accuracy of self-reporting can be low for some CCs. As an example, it has been shown that subjects over-report rheumatologic conditions in surveys where the diagnosis is self-reported [27, 28]. Furthermore, self-reporting accuracy is likely to be very low for CCs with a vague definition such as migraine. Otherwise, it is likely that those treated (and then identified with pharmacy data) are likely to refer to a more severe case definition.

The HIS is likely less sensitive because it identifies only the more severe cases that need hospitalization. Also the REP is likely less sensitive due to the fact that the free access to health care services is also given to citizens belonging to specific groups of age (e.g., people aged ≥ 65 years old) or of low income and thus there is no practical reason in some cases to require the exempt for a specific CCs.

Some CCs prevalences were comparable with those estimated by an Italian health administrative database of GPs [10]. For this source there were no available data of CCs at regional level, but our findings showed a good agreement with some Italian prevalence of CCs, particularly with diabetes [29] and chronic obstructive pulmonary disease [10].

The prevalence estimates obtained by pharmacy data have several advantages compared to those obtained by other health administrative databases and by cross-sectional surveys.

The prevalence estimates can be easily obtained and they provide estimates not conditioned by sampling problems.

In particular, these estimates can be provided also by very small geographical areas while this is not possible in surveys planned to provide reliable information, in terms of precision, only at a national or regional level. Furthermore, these estimates can be updated frequently. Another advantage is that the ATC coding used to identify CCs is internationally used and this allows immediate comparisons of prevalence estimates in other countries. Finally, it is important to highlight that pharmacy data could also be used to evaluate the incidence for some acute diseases in case a specific treatment would be available.

The present study has several limitations. All the health databases used are affected by selection bias because they likely do identify more severe cases. Statistical techniques had been proposed to correct for the selection bias using external information [30] such as health surveys but in our case this approach was not feasible because the non-availability of survey data. Another approach to correct for selection bias is using capture-recapture techniques [31] but we were not authorized to link the health databases due to privacy reasons.

Using pharmacy data to identify a specific CC implies that those drugs are used exclusively for the treatment of that CC. Furthermore, it is also important that the drug identifying CCs be used in any stage of the disease. We feel that for some CCs the coverage of drug treatment is low and hence a poor proxy for prevalence (i.e., anticholinesterase agents for dementia, interferons for chronic hepatis B, drugs listed for chronic renal disease, for cirrhosis, and for malignancies). It is also important to remind that for some diseases, such as osteoporosis and diabetes, the pharmacological treatment is not given mainly because of under-diagnosis of the conditions.

Another limit regards the potential inclusion of individuals without the specific CC evaluated who used the drug as incidental users or for other CCs not considered in this study. However, this limit has had little impact because we referred to drugs that had restrictive notes for dispensing, restricting the use only to individuals diagnosed with that CCs (see methods section). Furthermore, we included only individuals who had prescribed three or more packages of drug used to identify the CCs, but no sensitivity analysis was performed to determine if increasing the number of packages resulted in substantial changes in the prevalence estimates.

Finally, this study did not consider drugs directly prescribed/administered in hospital setting.

Our study showed that pharmacy data can provide, in several cases, reliable prevalence estimates of CCs in the general population. The estimates obtained could be a quick and priceless alternative to survey data that assess the health population status.

The methodology offers the possibility of international comparison of disease prevalence, prescribing and drug costs in managing CCs.