Assumptions made when preparing drug exposure data for analysis have an impact on results: An unreported step in pharmacoepidemiology studies

Abstract

Purpose: Real-world data for observational research commonly require formatting and cleaning prior to analysis. Data preparation steps are rarely reported adequately and are likely to vary between research groups. Variation in methodology could potentially affect study outcomes. This study aimed to develop a framework to define and document drug data preparation and to examine the impact of different assumptions on results.

Methods: An algorithm for processing prescription data was developed and tested using data from the Clinical Practice Research Datalink (CPRD). The impact of varying assumptions was examined by estimating the association between 2 exemplar medications (oral hypoglycaemic drugs and glucocorticoids) and cardiovascular events after preparing multiple datasets derived from the same source prescription data. Each dataset was analysed using Cox proportional hazards modelling.

Results: The algorithm included 10 decision nodes and 54 possible unique assumptions. Over 11 000 possible pathways through the algorithm were identified. In both exemplar studies, similar hazard ratios and standard errors were found for the majority of pathways; however, certain assumptions had a greater influence on results. For example, in the hypoglycaemic analysis, choosing a different variable to define prescription end date altered the hazard ratios (95% confidence intervals) from 1.77 (1.56-2.00) to 2.83 (1.59-5.04).

Conclusions: The framework offers a transparent and efficient way to perform and report drug data preparation steps. Assumptions made during data preparation can impact the results of analyses. Improving transparency regarding drug data preparation would increase the repeatability, reproducibility, and comparability of published results.

Keywords: data preparation; pharmacoepidemiology; reproducibility; transparency.

Figure 1

The drug exposure preparation algorithm. qty = total quantity entered by GP for the prescribed product; ndd = derived numeric daily dose; numdays = number of treatment days; dose_duration = derived duration of prescription. The highlighted pathway is the “primary preparation pathway” we defined in the second phase of each analysis; this pathway was used to generate one dataset, then further datasets were generated by varying a single assumption with respect to this primary pathway. All options that produce a missing value stay coded as missing unless otherwise stated. *For options 6d: If only one stop available, use it; if 2 available and equal, use that date; if 2 available and unequal (but within x days), use mean; if 3 available and unequal, use mean of closest 2 if within x days. **Records with missing stop dates after step 7 are dropped

Figure 2

Influence of drug exposure data preparation assumptions on association between oral hypoglycaemic drug class (sulfonylureas compared with biguanides as referent) and CVD events: Distribution of hazard ratios and standard errors from 50 random data preparation pathways

Figure 3

Influence of drug exposure data preparation assumptions on association between oral hypoglycaemic drug class (sulfonylureas compared with biguanides as referent) and CVD events: Effect of changing one data preparation option from primary pathway

Figure 4

Influence of drug exposure data preparation assumptions on association between oral glucocorticoid use (on vs off) and CVD events: Distribution of hazard ratios and standard errors from 50 random data preparation pathways; 3 years of follow‐up

Figure 5

Influence of drug exposure data preparation assumptions on association between oral glucocorticoid use (on vs off) and CVD events: Effect of changing one data preparation option from primary pathway; 3 years of follow‐up

Figure 6

Influence of drug exposure data preparation assumptions on association between oral glucocorticoid use (on vs off) and CVD events: Distribution of hazard ratios and standard errors from 50 random data preparation pathways; 20 years of follow‐up

Figure 7

Influence of drug exposure data preparation assumptions on association between oral glucocorticoid use (on vs off) and CVD events: Effect of changing one data preparation option from primary pathway; 20 years of follow‐up