By applying two different study designs in comparable target populations we have demonstrated that simple data structure and longer study period used in the SES improved recruitment and the likelihood of capture of medicines consumption. The probability of capture for a given API was further enhanced by higher frequency of use and longer duration of prescription. This meant that the SES gave a more comprehensive list of trade names, APIs and manufacturers than the PPS. The list of most frequently used APIs was overlapping in 90% with high correlation of use frequencies between the two studies.
The strength of both methods described here is the simplicity and uniformity of data collection. A two-step approach, as used in the ESNEE project may allow simultaneous coverage of both unit-level and individual-level data in neonatal excipient exposure studies, where hardly any multi-country data are available. We have recruited so far one of the largest international cohorts of neonatal units from 21 European countries revealing a surprisingly wide list of medicines. The number of trade names applied in NICU setting was also high. The response rate of 71% on country level as well as the number of prescriptions per neonate [9, 29] are similar to those reported previously [18–21]. Against the background of rising safety concerns of neonatal medicine/excipient exposure, the wide trade name list from the SES allows identification of substitution possibilities, while individual exposure data from the PPS provide a more precise quantification of the problem.
Different study designs can be applied in pharmacoepidemiology [30, 31]. Cohort studies have the advantage of allowing data collection over prolonged time periods which is more likely to capture rare exposures. However, the expense and duration make it hard to implement cohort designs in multinational studies . Case–control studies allow rich data collection when a limited number of medicines/excipients are targeted and data collection is limited to a few centers. The PPS and SES, in contrast, are easy to perform in multinational settings. They are time and resource saving and to some extent cover each other`s shortcomings like under- or overestimated exposure rates of rarely used medicines or need for individual data . Alternately, nested study designs with longitudinal data collection allow linkage between different datasets and may prove the best option. However, to the best of our knowledge, it has hardly been used in paediatric pharmacoepidemiological studies.
When planning medicine/excipient exposure studies the research question to be answered needs to be balanced against the implications of study design. We recognize that a 3-day survey is actually very short in duration. However, even this small longitudinal component resulted in a more comprehensive list of medicines due to the better recruitment when simple data structure was used. On the other hand, when individual exposures of frequently used medicines are targeted, the increasing data volume (both structure and duration) needs to be weighed against evolving decrease in compliance that may lead to underestimation of prevalence . It has been shown that the longer the recall period, the greater the imprecision (especially underestimation) of prevalence estimates [33, 34], although this finding may be more relevant in non-medical informants . Furthermore, not only optimal recruitment but also sample size maintenance over the course of a study becomes an issue .
Additionally, economic evaluation plays a role in the choice of a specific study design. There is considerable overlap between acceptable methodologies and those sanctioned by health economists . Although pharmacoeconomic analysis remains beyond the scope of this study, a few comments can be made. There were no major cost-differences in conducting these studies with the exception of those associated with the development of the PPS database. Web-based collection with automated control of data completeness was applied to improve compliance and quality of reporting. We believe that wherever sufficient to answer the research question, simple data structure is favourable in terms of resource allocation. However, when a more complex format is required, reasonable additional expenditure on data collection tools may result in an improved cost-benefit ratio.
We have also shown a high correlation between APIs exposure on unit and individual level. This should allow an indirect individual exposure assessment from less demanding approach like SES under the condition of limited resources. We realize, that the estimates in the frequency of medicine use in the PPS are systematically lower compared to the SES. However, knowing the relationship between the two an estimate for one can be extrapolated from the other. Nevertheless, such calculations should be applied with caution and to estimate the exposure only for the studied population as a whole. In neonatal studies stratification by GA age allows a more meaningful risk assessment due to the strong dependence of the maturity of metabolic pathways on the postmenstrual age of the newborn [37, 38]. It has been shown previously and also by us that medicines utilization pattern in extreme prematurity is different from that near or at term [9, 29, 39, 40]. Therefore, the PPS approach should be preferred to obtain individual exposure estimates in neonates of different GA.
Some limitations of the present study need to be noted. First, the quality of data collection was dependent on the personnel entering the data. In both studies data insertion was done by medical specialists according to standardized protocols. As for the quality of reported data, in the SES most queries concerned demographic data; only a few queries were related to the information about medicines. Second, neonatal units were given the option of selecting the most appropriate day(s) for data collection. This could lead to underestimation of medicine use because less busy day(s) probably were more likely to be chosen. However, we believe that eventually this approach captured information on a larger number of neonates and medicines through improved compliance and a larger number of participating units. Third, the slightly different representation by country required exclusion of two APIs from final analyses due to differences in regional handling/approach. Fourth, the SES and the PPS were done at different times. Nested design with detailed individual data collection (PPS) performed as part of the SES was not feasible, as an interim between the two studies was needed to prepopulate the PPS database from the SES list of medicines. This could potentially reduce the overlap between the studies. In that situation our data about overlap represent a “worst case scenario”. The true concordance in API reporting may be higher than described by us. Finally, about two thirds of the 31 invited countries agreed to participate with relative inhomogeneity of the regional distribution of participating units. Although the inability to randomly select hospitals probably caused overrepresentation of teaching hospitals (61% and 77% in the SES and PPS, respectively, compared to less than 10%, reported in national hospital statistics ), data on the distribution of the levels of neonatal care will allow appropriate adjustments. Both SES and PPS have achieved representative sample size on patient level, the margin of error being 1.73% and 2.63%, respectively (CI 95%; response distribution 50%; population size 5.4 million live births). However, it would be valid under the condition of randomisation that remained unfeasible in our studies. Still, the current analysis gives a notion of good coverage of the target population in the SES and PPS.