How measurements affected by medication use are reported and handled in observational research: A literature review

Abstract Purpose In epidemiological research, measurements affected by medication, for example, blood pressure lowered by antihypertensives, are common. Different ways of handling medication are required depending on the research questions and whether the affected measurement is the exposure, the outcome, or a confounder. This study aimed to review handling of medication use in observational research. Methods PubMed was searched for etiological studies published between 2015 and 2019 in 15 high‐ranked journals from cardiology, diabetes, and epidemiology. We selected studies that analyzed blood pressure, glucose, or lipid measurements (whether exposure, outcome or confounder) by linear or logistic regression. Two reviewers independently recorded how medication use was handled and assessed whether the methods used were in accordance with the research aim. We reported the methods used per variable category (exposure, outcome, confounder). Results A total of 127 articles were included. Most studies did not perform any method to account for medication use (exposure 58%, outcome 53%, and confounder 45%). Restriction (exposure 22%, outcome 23%, and confounders 10%), or adjusting for medication use using a binary indicator were also used frequently (exposure: 18%, outcome: 19%, confounder: 45%). No advanced methods were applied. In 60% of studies, the methods' validity could not be judged due to ambiguous reporting of the research aim. Invalid approaches were used in 28% of the studies, mostly when the affected variable was the outcome (36%). Conclusion Many studies ambiguously stated the research aim and used invalid methods to handle medication use. Researchers should consider a valid methodological approach based on their research question.


Key Points
• Methodological studies stressed the importance of adequately handling variables affected by medication use and showed that using invalid methods may lead to substantial bias. However, we found that many clinical studies did not consider this issue.
• A large proportion of the studies did not provide information on whether their interest was in the observed or the untreated underlying values. Without clear reporting on research aims, interpretation of the results will be ambiguous.
• Methods which have been shown invalid, such as restricting a study population to nonmedication users when the outcome variable was affected by medication use, are still often used.
• Justification on methods used for handling medication use was seldom given.

Plain Language Summary
In epidemiological research, measurements affected by medication, for example, blood pressure lowered by antihypertensives, are common. Different ways of handling medication are required depending on the research questions and the function of the measurement in the analysis (e.g., whether the blood pressure is the exposure or the outcome). Incorrect handling of medication use in the analysis may introduce bias or lead to a wrong interpretation of the result. This study aimed to review how medication use are being handled in observational research. We review a total of 127 articles. Most studies did not perform any method to account for medication use and no advanced methods were applied. In 60% of studies, the validity of the method used could not be judged due to ambiguous reporting of the research aim. Invalid approaches were used in 28% of the studies. Many studies ambiguously stated the research aim and used invalid methods to handle medication use. We urge researchers to consider a valid methodological approach based on their research question.

| INTRODUCTION
Measurements affected by medication use are a commonly encountered feature in epidemiological research. For example, blood pressure is lowered by antihypertensive drugs or glucose levels by glucoselowering drugs. Several methods for handling medication use have been proposed and compared. [1][2][3][4][5][6][7][8][9] Studies have shown that different methods may lead to substantially different effect estimates, [2][3][4][5][8][9][10] and the optimal method depends on (a) the research aim and (b) whether the medication effect is on the exposure, outcome or a confounder. 10 If the method used for handling medication effect does not match the research question, substantial bias can be introduced and the interpretation of results will be unclear. 11 Thus, it is essential to carefully think about the research question when some individuals in a study population use medication that affects the variables in the dataset. In some situations, the research interest could be in the observed measurements, regardless of whether some individuals' measurements are lowered due to antihypertensive medication use; for instance, when the effect of current blood pressure on the course of the disease for patients infected with Covid-19 is considered. In other cases, blood pressure values that would have been observed if the medication was not administered (sometimes referred to as underlying values 2,12 ) could be the primary interest, for example, if the effect of genetic factors on blood pressure are examined. In this instance, a method to correct for the medication effect should be used.
Handling of medication use in epidemiological research has received attention, although this was mainly in methodological papers. [1][2][3][4][5][6][7][8][9]13,14 There are studies that adopted some of the methods suggested. 15,16 However, a majority seems to overlook the potential bias due to inadequately handling medication use. 4 To our knowledge, there has been no systematic review on how medication use is being handled in research practice. Therefore, In this literature review, we aim to investigate which methods are used in observational studies to handle measurements affected by medication, assess how often methods used correspond to the research aims stated in these studies and evaluate the validity of the methods used. Select a maximum of 50 arƟcles for each journal field for review. If a journal contained less than ten arƟcles meeƟng the inclusion criteria, all arƟcles from that journal were selected to be reviewed. Rest of the studies were randomly selected unƟl the sample size per journal field meets 50 or no more arƟcles were leŌ to be selected. The full-text of the identified papers was screened, and papers that met following inclusion criteria were selected for review: (a) observational studies in adults, (b) sample size larger than 100, (c) aimed to answer etiological questions, (d) performed linear or logistic regression (including linear mixed modeling), and (e) inclusion of any of the following variables: blood pressure-related measurements (e.g., systolic or diastolic blood pressure, pulse wave velocity), glucose-related measurements (e.g., glucose level, insulin level, HbA1c, HOMA index) and lipid levels (e.g., cholesterol measures, triglycerides).
For studies on type 1 diabetes patients, glucose measurements were not considered because there is no variation in glucose medication use in these patients as insulin treatment is mandatory and unavoidable. If blood pressure related measurements or lipid measurements were used, these studies could be included.
Among the studies that met the inclusion criteria, we selected a maximum of 50 articles to be reviewed from each field. If a specific journal (five per field) contained less than 10 articles meeting the inclusion criteria, all articles from that journal were selected to be reviewed. The rest of the studies were randomly selected until the sample size per journal field met 50 or no more articles were left to be selected. If two or more studies used the same study population within a field, the latest publication was considered ( Figure 1).

| Data extraction
Data extraction for all 127 papers was independently performed by two reviewers, JC (a PhD candidate in clinical epidemiology) and SlC (a senior statistician and epidemiologist). Disagreements between the two reviewers were resolved during a consensus meeting involving the third reviewer, OMD (a senior epidemiologist and endocrinologist). For each paper, the following general information was extracted: 1. Authors, journal, name of the study/cohort/database.

Study population and sample size.
3. Research question with exposure(s) and outcome of interest.

Whether linear, logistic regression or both were performed.
For information related to medication use, we extracted the following: 5. Measurements that may have been affected by medication use (blood pressure, glucose, and/or lipid). "Medication use" was defined as the use of drugs that aim to lower blood pressure, glucose, or lipid level.
6. Whether the measurement potentially affected by medication was an exposure, an outcome or a confounder. We used the following rules: a. When the measurement was mentioned as an "independent variable" and the effect of the variable on the outcome was specifically discussed in the paper, it was coded as an exposure. 7. Percentage of individuals using medication. 8. Whether details on medication information were given (e.g., type and dose of medication, duration of use).
9. Methods used for handling medication use for each affected variable.
a. If different variables had the same role and were handled by the same method, the method was recorded once (e.g., if a study had blood pressure and glucose level as confounding variables and medication use for the both variables are handled by a restriction method, the method was recorded once).
b. When multiple models were used to evaluate the same relationship, the most complex model was considered (e.g., when both unadjusted and adjusted analyses were performed to estimate the relationship between the same variables, the adjusted analysis was considered).
c. Justification for the chosen method.
d. Sensitivity analyses for handling medication use.

| Assessment of research aims and the validity of the methods used
We evaluated the validity of methods used for handling medication use based on the research aims of the study and which variable was affected by medication use. Figure 2 displays our evaluation process.
In detail, the following steps were taken.

| Step I
For each variable affected by medication use, we first evaluated the research aim as stated by the authors, which was categorized as follows: 1. The interest is in the observed values as they are.
2. The interest is in the values that would be observed if no medication was administered (we refer to this as "values if untreated" or "untreated values" in the further text).
3. The interest is ambiguously reported.

| Step II
The validity of the method used for each variable was evaluated in relationship to the research aim and whether the affected variable is an exposure, an outcome, or a confounder. The assessment on whether the methods used are in general valid or invalid was based on recommendations from previous methodological studies. [2][3][4][5][6]10,11,[17][18][19][20][21][22][23][24][25][26] For example, restricting the study population to non-medication users was considered valid when the exposure or a confounder is affected by medication use regardless of whether the research aim is in the values as observed or if untreated. This is because the restriction on a proxy variable of the

| RESULTS
Our search strategy in PubMed retrieved 258 articles in cardiology journals, 331 articles in diabetes journals and 688 articles in epidemiology journals (see Table 1 for the number of papers and Figure 1 for flow-chart). After the screening process, 49 articles in the cardiology field, 73 articles in the diabetes field, and 28 articles in the epidemiology field remained. For the diabetes field, a subset of 50 articles was selected as described in the methods section. We included 49 articles from cardiology journals, 50 articles from diabetes journals, and 28 articles from epidemiology journals, for a total of 127 studies.

| Summaries of reviewed articles
Supplementary material S3 displays the complete list of the reviewed articles and extracted information of each article. Table 2 T A B L E 3 Frequency of methods used for handling medication use in main analyses, n (%) Note: The sum of the subtotals exceed the total number of articles included for each journals field (49 for cardiology journals, 50 for diabetes journals, and 28 for epidemiology journals) because more than one variable was affected by medication use in some studies.
provides a summary of the included studies. Overall, the measurement affected by medication use was most often a confounding variable (In 56% of the studies), followed by an outcome (42%) and/or an exposure (35%). In the epidemiology journals, affected outcomes were more often present (64%). Sample sizes varied largely between the reviewed articles and were generally larger in the epidemiology journals. Included studies performed linear regression analysis (59%), logistic regression analysis (40%), and/or linear mixed modeling (9%).
Overall, a majority of the studies did not report the percentage of medication users (47%) or only reported medication use for part of the variables affected (14%). Among the studies which fully or partially provided information on the percentage of medication users, the median percentage of medication users was 32%. The percentage of medication users ranged from 0 to 100, because some studies restricted their study population to medication users or non-users.
Details of medication use, such as dose or prescription frequency, were seldom given (7%). A binary covariate in a regression model was the next most used method for exposures (18%) and outcomes (19%). For confounders, it was one of the most used methods (45%). The binary variable used for the adjustment was often "using medication (yes/no)." However, Note: The percentages add up to 100 per "affected variable." one study adjusted for "using medication or having high value (yes/no)" (e.g., hypertension vs. no hypertension, while defining hypertension as taking antihypertensive drugs or having blood pressure above a certain level).

| Methods used for handling medication use
Adding an estimate of the mean medication effect to treated values was adopted only in four studies. One study used this method for handling medication use in the exposure. No study used any of the more advanced methods suggested in the literature, such as quantile regression, 3 censored normal regression 2 or Heckman's treatment model. 4,5 In total, only 10 studies (8%) explicitly provided justification for the chosen methods for handling medication use. Given justifications, however, may not reflect the validity of the methods used. Sensitivity analyses were performed in 21 studies (16%) in total. A list of methods used in the sensitivity analyses can be found in Supplementary material S5.

| Assessment of research aim-analysis match and validity of the methods used
The results of the assessment of the methods used for handling medication use are summarized in  We found that invalid methods were especially prevalent when the affected variable was the outcome. Often the analysis was performed conditional on medication use. Although the bias due to selection on events related to the outcome has been discussed extensively in the literature, 2,21,22,[24][25][26] it seemed that such consideration was often not taken into account. We also observed that the research aim was most often ambiguously reported for confounding variables affected by medication use. This is not surprising since confounders are mostly not the variables of main interests. However, inadequately handling medication use in confounding variables can lead to bias. 10 We noticed that recommendations in methodological papers were seldom applied. For example, Tobin et al. 2  We only included studies that used linear regression, logistic regression and mixed linear models. However, potential bias due to measurements affected by medication use is present in any study where a mixed study population of medication users and non-users exists. 14

| CONCLUSION
Our review has shown that potential bias due to medication use is often overlooked and that decisions on handling medication use are frequently made without valid justification. We urge researchers to provide clear information on medication use, consciously decide on a method for handling medication use based on their research question and communicate the rationale behind their decision.