Implementing high‐dimensional propensity score principles to improve confounder adjustment in UK electronic health records

Recent evidence from US claims data suggests use of high‐dimensional propensity score (hd‐PS) methods improve adjustment for confounding in non‐randomised studies of interventions. However, it is unclear how best to apply hd‐PS principles outside their original setting, given important differences between claims data and electronic health records (EHRs). We aimed to implement the hd‐PS in the setting of United Kingdom (UK) EHRs.


| INTRODUCTION
Electronic Health Records (EHRs) are increasingly used for research investigating the effects of medications. 1,2 Adequate adjustment for confounding remains a key issue and incorrect conclusions can be drawn amid concerns of residual or unmeasured confounding. 3,4 Developed in US claims data to improve confounder adjustment, the high-dimensional propensity score (hd-PS) approach treats information stored within healthcare databases as proxies for key underlying confounders. 5 Some proxies may be strongly correlated with variables typically included in a traditional propensity score (PS) analysis; others may represent information about patients that is otherwise unmeasured, for example, frailty. 5 Despite application in various settings (including UK EHRs), [6][7][8][9] detailed guidance on how to apply the hd-PS outside US claims data is lacking. Important differences between data sources mean that careful consideration is needed when implementing hd-PS principles to ensure source-specific characteristics are handled appropriately.
We propose a series of modifications to the hd-PS that aim to characterise key features of UK EHRs whilst adhering to the underlying principles. 5,6 2 | PROPENSITY SCORES The PS is the conditional probability of being treated given a set of observed covariates. [10][11][12] PSs model the treatment allocation process and therefore offer advantages over multivariable analysis in EHRs, since investigators are forced to consider indications for treatment use and can convert large amounts of confounder information into a single number. 4 At a particular value of the PS, the distribution of observed covariates is balanced between treated and untreated individuals, allowing consistent estimation of treatment effects, assuming all confounders are included in the model. 13 3 | DESCRIPTION OF THE hd-PS APPROACH AND UNDERLYING PRINCIPLES

| Preliminary steps
Demographics (d) and clinical factors believed to be important confounders (l) are forced into the PS model. 5 A baseline time-window for assessing patient confounder information is established (often 1 year before study entry date).

| Identification of most relevant covariates
Relevant information in the database is separated into p dimensions. 5 The underlying principle is that each dimension should represent a different aspect of care relevant to the healthcare system under investigation (principle 1). For example, in US claims data, it is typical to separate information pertaining to diagnoses, procedures and prescribing. 5 The hd-PS uses the Bross formula to prioritise covariates across dimensions by their potential to bias the treatment-outcome relationship. 5,15,16 This has three components. Firstly, it takes the confounded apparent relative risk (ARR) for a particular binary covariate as a function of the relative risk (RR) in the absence of confounding by this covariate. Secondly, the imbalance in prevalence amongst the exposed (P C1 ) and unexposed (P C0 ) patients. Thirdly, the independent association between a confounder and the study outcome (RR CD ): Each dimension is sorted in descending order by the magnitude of jlog(bias M )j. This bias term takes a larger value the greater the potential a covariate has to bias the relationship of interest. Therefore, the top k empirical covariates are included in the PS. Typically several hundred covariates are selected.

| Estimation of the hd-PS
The selected empirical covariates are added to the predefined variables before estimating the PS. Traditional PS methods are then used to estimate the treatment effect. 12 The final principle is that after accounting for the top k empirically selected covariates, residual confounding effects are assumed to be negligible (principle 4).

| PROPOSED IMPLEMENTATION OF hd-PS PRINCIPLES TO UK EHRS
In this section, issues surrounding the translation of hd-PS principles to UK EHRs are discussed alongside our proposed modifications (summarised in Figure 1).

| Principle 1: Identification of dimensions
There are important differences between insurance claims and EHR data in terms of data availability, structure and the reasons for data recording. 17 For the prescription dimension the British National Formulary (BNF) coding system is used. We classified prescriptions at the BNF paragraph level which typically groups prescriptions by indication rather than mechanism of action. 22

| Principle 3: Code recurrence
Code frequency is assessed by the hd-PS to provide an indicator of a patient's underlying health. 5 In claims data all relevant information is recorded at each instance a claim is completed which leads to an intrinsic link between disease severity and code frequency.
EHRs exist for clinical record keeping which means that such a link is harder to discern since all relevant information will not necessarily be recorded at each consultation. Frequency of recording is instead likely to be a function of several factors including severity of illness, frequency of consultation and GP preference.
We classified the frequency of codes in a pre-specified baseline time-window, 1 year prior to study entry. Recognising the variability in recording we replaced the "Once" indicator with an "Ever" indicator which captured whether a code had been recorded during a patient's entire history. The remaining frequency indicators were assessed during the baseline time-window.
We hypothesised that the degree to which information is recorded at each consultation was likely to vary by dimension, with more complete recording likely in the prescription and referral dimensions. However, in the clinical dimension relevant information is often not re-recorded at each consultation. For example, a patient receiving prescriptions relating to a diagnosis of T2DM will have this diagnosis recorded but not necessarily at each relevant consultation.
To investigate whether this information was likely to be overlooked when assessing information in a narrow time-window we extended the baseline time-window for the Clinical dimension.
Acknowledging the fact that patients will have varying lengths of baseline information available we classified the frequency of codes by assessing rates in instead of counts. We used three indicators to classify our revised frequency assessment (see Figure 1 for full definition). A recent cohort study using the CPRD linked with the Myocardial Ischaemia National Audit Project (MINAP) investigated the combined use of proton pump inhibitors (PPI) with clopidogrel and aspirin. A possible interaction whereby PPIs may reduce the conversion of clopidogrel to its active metabolite had been suggested, raising concerns that combined use may lead to a reduction in clopidogrel effectiveness and an increased risk of vascular events. The cohort analysis found that combined use was indeed associated with an increased risk of myocardial infarction (MI). 3 The pattern of associations found strongly suggested that residual confounding between patients may have explained the results as they

| Principle 4: Selected number of variables
were not specific to MI and were found for both strong and weak inhibitors of cytochrome P450 3A4 (the mechanism proposed for the drug interaction). Furthermore, a self-controlled case series (SCCS) analysis 25 conducted on the same data found no evidence of increased risk.
The authors concluded that the results from the cohort study reflect confounding in the cohort estimate. In addition, unconfounded studies based on genetic instrumental variable approaches using genetic effects on drug metabolism pathways also suggested no evidence of increased risk. 26 A randomised double-blind trial has The clinical dimension also contains information relating to administrative codes or references to measurements that occurred without results.
subsequently also suggested a lack of clinical effect of PPIs on MI risk, when used in combination with clopidogrel (HR = 0.92; 95% CI: 0.44-1.90). 27

| Design
We summarise the original study design conducted by Douglas et al. 3 Patients had to be present in the CPRD with at least 12 months of prior registration before first prescription for clopidogrel. Study entry was defined as the latest of first recorded clopidogrel prescription in combination with aspirin or 1 January 2003. Patients were censored at the earliest of stopping treatment for aspirin or clopidogrel, death, transferring out of the practice, last data collection date for the practice, 31 July 2009 or an occurrence of MI. Exposure was defined as any prescription for a PPI. We focus on the incident MI outcome which was ascertained using the MINAP database.

| Statistical analysis
The original study analysed the hazard ratio (HR) for the association between PPI treatment and MI using Cox models, adjusting for 14 selected confounders. Missing data for body mass index, smoking and alcohol consumption were handled using missing categories.
These conditions were applied consistently across all analyses.
We reanalysed the original data taking an intent-to-treat approach that classified patients according to original exposure status and incorporated baseline confounder information using PSs. We esti- All HR results are presented with 95% confidence intervals in parentheses. Analyses were conducted using Stata 14. 28

| RESULTS
Demographics and clinical characteristics for the cohort study are summarised in Table 2 For the modified analyses, we mapped the clinical and referral dimensions from Read code to ICD-10. A large number of Read codes represent non-clinical information, for example, codes relating to administrative procedures. Since the aim of the mapping procedure is solely to capture clinically relevant information unmapped Read codes were expected. Upon inspection, the resulting unmapped codes could generally be categorised as either administrative information (eg, a letter), an indicator of a completed test without the result (eg, "blood pressure reading was taken") or coarse information we would typically include more granularly in the pre-defined covariates (eg, broad smoking terms). We include a sample of the most frequently occurring unmapped Read codes in the Supporting Information.
Results for all analyses are presented in Table 3. Using the confounders originally identified by Douglas et al 3 we obtained a HR for the association between PPI use and MI of 1.17 (1.00-1.35).
Applying our modifications reduced the HR for the association between PPI use and MI moving it towards a null result ( Figure 2).
The fully modified hd-PS obtained an HR of 1.00 (0.78 to 1.28).
In sensitivity analyses, extending the baseline time-window for the Clinical dimension lead to point estimates further from the null.
Varying the number of covariates did not meaningfully alter point estimates. However, selecting fewer than 500 variables did improve the precision of effect estimates (Table 3).
We investigated the estimated PS distributions by treatment group obtained from investigator led and hd-PS analyses ( Figure 3).
These distributions compare the characteristics of patients in the populations under investigation. Compared to the investigator led approach, the hd-PS exposed greater variation between the treatment groups and captured extra predictors of prescribing which were also causing confounding bias. The framework we have built could also be extended to include laboratory test results and free text information, the latter of which has been previously explored. 6 Whilst there have been several developments to the hd-PS since its inception, 6 there has been little exploration of how to translate the algorithm beyond claims data. Much of this development work for hd-PS has been focussed on demonstrating it obtains known associations, such as the effect of non-steroidal anti-inflammatory drugs on the risk of gastrointestinal bleed. 5,9,24,29 However, these results have also been obtained through traditional methods of confounder adjustment. In the case study we present, a hd-PS approach has removed a known confounded association discovered using traditional methods.
Future applications of the hd-PS in this context will benefit from updates to the cross-map between Read and ICD-10. In the literature accompanying these cross-maps NHS Digital state that not every concept in one coding system can or should be represented in another. 21 NHS Digital's intention was to map clinically meaningful terms only, and it was reassuring to observe that the majority of unmapped Read codes were clinically uninformative and would typically be discarded in an investigator analysis (see Supporting Information).
When calculating the SEs for treatment effects we have ignored variable selection or estimation of the PS. Theoretically, this is likely to result in narrower confidence intervals, 30 although the practical conse- quences are yet to be fully explored. We obtained a bias-corrected bootstrap 95% CI based on 1000 replications for our final model of 0 Our results highlight the potential benefit of employing hd-PS approaches in EHR studies, especially to overcome intractable confounding. However, the hd-PS is not a panacea and we acknowledge that in studies where the confounding structure is relatively simple, the robustness of results is unlikely to differ between traditional and hd-PS methods. We recognise the need for further exploration of the hd-PS in this setting, via both controlled conditions and case studies. One outstanding issue surrounds the transparency of reporting when using hd-PS approaches and there is a need for tools to better communicate proxies included in the final hd-PS model.
This study has shown that the application of hd-PS methods outside the context of claims data requires careful consideration of how to optimally apply hd-PS principles. By adapting hd-PS principles to the UK EHR setting we have demonstrated the potential for hd-PS to improve confounder adjustment in EHRs.

ETHICS STATEMENT
Scientific approval was obtained to use CPRD data by the Indepen-