SEMINAR
Introduction to case-crossover study design for drug-drug interaction
2026 Volume 8 Issue 2 Pages 62-71
Details
2026 Volume 8 Issue 2 Pages 62-71
This article introduces the basic concept of a novel 6-parameter model case-crossover study with active comparator design. While case-crossover study design eliminates time-invariant confounding, the 6-parameter model as a modified version also allows studying the effect of drug initiation patterns between two interacting drugs. The implementation of the case-crossover study is illustrated, including the generation of the dataset for data analysis and performing the 6-parameter model using conditional logistic regression. Potential interpretation framework is proposed using eight combinations based on three scenarios shown in the results of the 6-parameter case-crossover study design.
Multimorbidity is prevalent globally with an overall prevalence of 37%1). Taking multiple medications is common in people with multimorbidity, which could lead to drug-drug interactions (DDIs). It is estimated that 5% of hospital admissions are due to DDIs in elderly patients2). Randomised controlled trials are rarely conducted to evaluate DDIs due to potential ethical and logistical concerns3). Therefore, at the time a drug is approved for routine use, many potential DDIs remain unknown. Investigating DDIs is therefore an important part of pharmacovigilance which requires extensive pharmacokinetic and pharmacodynamic studies and spontaneous reports to screen for potential DDIs. However, there are well established limitations of spontaneous reports4), in particular when the causal suspicion of the reporter may only be attributed to a single drug rather than a DDI5). Improved epidemiological methods and established electronic health databases now offer the potential to provide further evidence to assess possible causal associations and quantify the scale of the impact of DDIs. Investigating DDIs using this approach requires very large electronic population databases, as only a small proportion of the population concomitantly receives the drugs together, and the study outcomes are often rare. Most studies focus on DDIs involving two drugs, namely the drug for which the effect is altered (defined as the object drug) and the drug that triggers the interaction (defined as the precipitant drug)3). While there are many hypothetical DDIs based on known pharmacological actions, including an increase in responses of the object drug by the precipitant drug, such action may not be clinically relevant. Therefore, thorough investigation using appropriate epidemiological study designs is needed.
In this article, we aimed to 1) provide an overview of a novel 6-parameter model of case-crossover study design and the use of active comparator; 2) illustrate how to implement the 6-parameter model of case-crossover study (how to generate the dataset suitable for running the STATA algorithms and R programme that we put on GitHub; 3) present how to interpret findings using the 6-parameter model case-crossover study.
To investigate possible causal associations in DDI research, conventional observational study designs including cohort studies and case-control studies have long been used. However, these designs are susceptible to between-person confounding. For example, in cohort study design, the risk of the outcomes of interest is compared between an exposed group (i.e., people receiving the drug of interest) and a comparison group (i.e., either people not receiving the drug of interest or people receiving another drug with a similar indication)6). Therefore, the health characteristics between exposed group and comparison group could be systematically different. However, as a within-person study design, time-invariant confounding could be removed. For example, the case-crossover study design, one of the within-person study designs, compares the odds of exposure to precipitant drugs within an individual continuously exposed to object drug to investigate DDIs7–9). Thus, the case-crossover study inherently controls for all confounders that are stable within individuals over the observation period.
Advantages and LimitationsCase-crossover study is a case-only design including people with the outcome of interest only. It is an analogue of case-control study design as it begins with identifying patients with the outcome of interest and then retrospectively measuring any change in past exposures over time7),8). It is easy to implement by comparing the odds of exposure to each interacting drug in a period prior to the outcome (hazard period) to the exposure to each interacting drug during an earlier control period (referent period) within an individual. Only discordant exposure pairs in the hazard and referent periods are included (i.e., cases exposed only in hazard period or only in the referent period). This design works best when the outcome of interest is acute10). It is also not susceptible to strict assumptions that could be easily violated when investigating DDIs, such as end of observation period independent of outcome occurrence (e.g., life-threatening outcomes), subsequent exposure not affected by the outcome (but discontinuation of suspected interacting drug often take place when DDI occurs), as in the self-controlled case series study.
Notably, the case-crossover study design is susceptible to time trend bias as the change in exposure between hazard and referent periods could be simply due to a population-level upward or downward trend in prescribing of the interacting drugs over time11). It may also lead to bias when the use of precipitant drug is persistent. As concordant exposure pairs in the hazard and referent periods would be excluded in the analysis, it is more likely to create upward bias when the only pattern contributing to the analysis is exposed during the hazard period and unexposed during the referent period12). A simulation study showed that when 30% of the study cohort remaining exposed to the precipitant drug, an upward bias was observed, but the estimate was unbiased when all people were exposed to precipitant drug with duration of <90 days13). In addition, risk differences cannot be obtained in within-person designs. Other study designs (e.g., cohort design) are needed to estimate absolute risks to quantify DDIs for evaluating public health impact.
A Saturated 6-parameter ModelConcomitant use of an object and a precipitant drug can be categorised as three combinations based on the order of initiation of these two drugs3). Understanding drug initiation patterns is clinically significant because the timing and order in which medications are started can greatly affect the risk and severity of DDIs. In routine clinical practice, starting a new drug in the presence of another can alter pharmacokinetics and pharmacodynamics, potentially leading to adverse outcomes if dose adjustments are not properly managed3). For example, initiating clarithromycin in a patient already taking warfarin may elevate warfarin levels and increase bleeding risk, as the prescriber may be unaware of re-titrating the dose of warfarin. Identifying these initiation patterns can help clinicians to anticipate and mitigate risks associated with DDIs.
A novel 6-parameter model case-crossover study design has been recently developed to examine the effect of drug initiation patterns between two interacting drugs14). Studying the effects in different drug initiation patterns accounting for the possible suboptimal management of dose adjustment could help us to understand how DDIs occur. Given that case-crossover study design can eliminate time-invariant confounding, this modified case-crossover design can both yield robust estimates and assess the effect of multiple drug initiation patterns. A 3-parameter model has also been developed, but it is susceptible to bias when there are heterogeneous effects for different drug initiation patterns and cannot distinguish different drug initiation patterns14).
A 6-parameter model can be used to study how the order of drug initiation impacts on outcomes14). It includes all possible patterns of how two interacting drugs can be initiated in three different ways in routine clinical setting. Analogous to a stratified analysis, the parameters in the logistic regression model are: 1) initiation of object drug only, 2) initiation of precipitant drug only, 3) both drugs are initiated together, 4) initiation of object drug in the presence of precipitant drug, 5) initiation of precipitant drug in the presence of object drug, 6) use one drug in the hazard period and another drug in the referent period (see Table 1).
| Strata | Description | Parameter |
|---|---|---|
| 1 | Object drug only | Initiation of object drug monotherapy |
| 2 | Precipitant drug only | Initiation of precipitant drug monotherapy |
| 3 | Joint exposure | Joint initiation of object drug and precipitant drug |
| 4 | Object drug when precipitant = 1 | Initiation of object drug in the presence of precipitant drug |
| 5 | Precipitant drug when object = 1 | Initiation of precipitant drug in the presence of object drug |
| 6 | Switch | Use one drug in the hazard window and the other drug in the referent window |
Apart from studying different order of drug initiation patterns, another strength of 6-parameter model is that we could assess confounding by indication by comparing the risk of outcomes in different drug initiation patterns14). For instance, if a higher risk of outcome is only associated with the stratum indicating both drugs are initiated together, the elevated risk is likely due to confounding by indication because at least two diseases or medical conditions occur around the time when the two drugs were initiated. On the other hand, if the object drug is believed to be associated with an increased risk of outcome itself (as shown in the stratum of object drug only), we could assess the DDI using the stratum of an initiation of a precipitant drug in the presence of an object drug as this stratum keeps the object drug concordant in both hazard and referent periods. (i.e., measures the risk of the outcome associated with the precipitant drug in object drug users)14). This principle also applies to the stratum of initiation of the object drug in the presence of the precipitant drug. Below we could use an example of assessing whether there is a DDI between warfarin (an object drug) and clarithromycin (a precipitant drug). Bleeding is a well-known adverse drug reaction of warfarin. To evaluate whether an observed risk of bleeding was due to DDIs or just the object drug warfarin itself, the risk of bleeding in the stratum of an initiation of clarithromycin in long-term warfarin users (i.e., keeping warfarin exposure concordant in both hazard and referent periods) would be useful for assessment by comparing it with that in the stratum of initiation of warfarin only.
Pharmacologically, it is not anticipated that differences in drug initiation patterns between drugs would lead to different effects on DDIs. However, when a higher risk of outcomes is associated with either the stratum of initiation of precipitant drug in the presence of object drug or the stratum of an initiation of object drug in the presence of precipitant drug in 6-parameter model, it could represent a positive signal due to a DDI involving a dose-titrated object. This could occur when the object drug and precipitant drug are not started together, e.g., the object drug is started, and the precipitant drug is later added or vice versa. The prescriber may be unaware to reduce the dose of the object drug to avoid a DDI3), versus joint initiation pattern. When both drugs are initiated together, the dose of object drug would more likely be adjusted to reduce the risk of a DDI. For example, when a patient had been taking warfarin and started a clarithromycin prescription afterwards, the prescriber might not be aware of adjusting the dose of warfarin to avoid a DDI. However, if warfarin and a clarithromycin were started together, the prescriber would be more likely to adjust the dose of warfarin to avoid a DDI. Therefore, the added advantage of using the 6-parameter model is to examine whether there is a DDI involving dose-titrated object by assessing the relevant strata.
In the 6-parameter model, the stratum of switch indicates the use of one drug in the hazard period and use of another drug in the referent period. This stratum does not directly inform an assessment of a possible DDI. It is also unlikely of clinical interest14), but it is included as it represents a possible exposure pattern stratum in the analysis of two potentially interacting drugs.
Active ComparatorIn observational pharmacoepidemiology studies, biases such as confounding by indication can pose significant challenges to valid inference results. An active comparator design can be applied to mitigate such biases by comparing the drug of interest with similar indications. While the case-crossover study design effectively controls for the time-invariant confounders, an active comparator design can be considered to reduce the time-varying confounding in the within-person study design. Negative control precipitants are recognised as the active comparators in DDI research, which do not have the potential pharmacological properties for a DDI but with similar indications as the precipitant drug. Therefore, time-varying confounding could be mitigated and elucidate the effect of DDIs specific to the precipitant drug of interest. However, using a negative control precipitant should be done with caution, as the choice should be guided by clinical rationale and prescribing patterns. Inappropriate control may even lead to an increase in the bias.
There are two active comparator approaches in the case-crossover study design: a simple ratio approach and an effect modifier approach. However, they have not been proposed in the 6-parameter model of case-crossover study design for DDIs in the current literature. This article therefore proposes to use the simple ratio approach for DDI research. To implement this approach, the odds ratio for each stratum of the 6-parameter model is estimated for the investigated object drug with the investigated precipitant drug. The analysis is then repeated for the investigated object drug with the negative precipitant drug. Then the comparator-adjusted estimate can be computed as a simple ratio of the estimate for the precipitant drug of interest and the negative precipitant drug. Wald-test-based method can be used to calculate the confidence intervals for each computed ratio.
To investigate the DDIs, two drugs with prior evidence of pharmacologic interaction and associated adverse clinical outcomes are selected. For example, warfarin and clarithromycin, where warfarin is an anticoagulant metabolised by cytochrome P450 enzymes (i.e., CYP3A4), and clarithromycin is known as a strong CYP3A4 inhibitor.
Generating the DatasetsThe first step in conducting the 6-parameter model of a case-crossover study is to generate a dataset that is ready for analysis. The example of how to derive a dataset is presented below.
Variables required for the input datasets with a defined study periodBefore we run the algorithms, we need the dataset containing variables of “patid”, “eventdate”, “cohortentry”, “cohortend”, “rxst”, “rxen”, “drug”.
Using electronic health records, each patient is assigned to a de-identified unique ID (“patid”). A diagnosis dataset containing all patients’ diagnosis information is used to identify cases with the outcome of interest. Cases are extracted with the outcome event occurrence date (“eventdate”). In the example, the study period starts on 2010-01-01 and ends on 2024-12-31. While the case is identified, each individual has their corresponding cohort entry date (“cohortentry”) and cohort end date (“cohortend”), where the cohort entry date is subject to inclusion criteria and the cohort end date is the earliest date of death or study end date, as shown in Table 2. In this example, we assume that all individuals are tracked continuously from birth to death within the database, so there is no need to define specific database entry or exit points for the cohort’s entry and end dates.
| patid | eventdate | cohortentry | cohortend |
|---|---|---|---|
| cco001 | 2015-07-16 | 2012-11-13 | 2021-09-05 |
| cco002 | 2017-03-25 | 2015-05-05 | 2017-11-27 |
| cco003 | 2020-11-14 | 2014-01-24 | 2020-11-23 |
| cco004 | 2015-10-17 | 2011-11-21 | 2023-12-31 |
| cco005 | 2023-12-12 | 2023-05-08 | 2023-12-31 |
| cco006 | 2018-10-04 | 2016-10-24 | 2018-11-13 |
A prescription dataset containing all patients’ prescription information is used to capture the exposures, including the object drug (warfarin) and precipitant drug (clarithromycin). Each observation record has a prescription start date (“rxst”), prescription end date (“rxen”), and relevant drug information such as drug name (“drug”) (see Table 3).
| patid | rxst | rxen | drug |
|---|---|---|---|
| cco001 | 2015-07-16 | 2017-03-23 | warfarin |
| cco002 | 2015-05-08 | 2017-06-20 | warfarin |
| cco003 | 2014-04-22 | 2014-07-14 | warfarin |
| cco003 | 2014-08-12 | 2018-10-29 | warfarin |
| cco003 | 2020-10-24 | 2020-12-22 | warfarin |
| cco005 | 2022-01-06 | 2022-07-06 | warfarin |
| cco005 | 2022-07-14 | 2023-05-03 | warfarin |
| cco005 | 2023-05-12 | 2023-09-04 | warfarin |
| cco005 | 2023-11-09 | 2024-01-04 | warfarin |
| cco006 | 2018-09-21 | 2018-12-03 | warfarin |
| patid | rxst | rxen | drug |
|---|---|---|---|
| cco001 | 2015-05-16 | 2015-05-24 | clarithromycin |
| cco001 | 2015-05-26 | 2015-05-30 | clarithromycin |
| cco001 | 2015-06-01 | 2015-06-06 | clarithromycin |
| cco001 | 2015-06-08 | 2018-06-13 | clarithromycin |
| cco002 | 2014-03-11 | 2014-03-15 | clarithromycin |
| cco002 | 2017-02-01 | 2017-02-11 | clarithromycin |
| cco004 | 2015-10-17 | 2024-03-03 | clarithromycin |
| cco005 | 2022-01-06 | 2022-07-06 | clarithromycin |
| cco005 | 2022-07-14 | 2023-09-12 | clarithromycin |
| cco005 | 2023-11-09 | 2024-01-03 | clarithromycin |
| cco006 | 2015-03-04 | 2018-08-28 | clarithromycin |
For each individual, a hazard window and a referent window are established based on a predefined risk period (“riskperiod”), typically 30 days, 60 days, or 90 days, according to the usual course of therapy and the predicted onset of DDIs in clinical practice, with inputs from clinicians and/or current literature, e.g., case-reports. In some cases, a washout period may be added to avoid autocorrelation and carry-over effects in exposures between hazard and referent windows. In the example, a washout period (“washoutperiod”) is applied. Using a risk period of 30 days and a washout period of 30 days, the hazard window is defined as days 1–30 before the event date and the referent window as days 61–90 before the event date (see Table 4a).
| patid | eventdate | period_start | period_end |
|---|---|---|---|
| cco001 | 2015-07-16 | 2015-06-17 | 2015-07-16 |
| cco002 | 2017-03-25 | 2017-02-24 | 2017-03-25 |
| cco003 | 2020-11-14 | 2020-10-16 | 2020-11-14 |
| cco004 | 2015-10-17 | 2015-09-18 | 2015-10-17 |
| cco005 | 2023-12-12 | 2023-11-13 | 2023-12-12 |
| cco006 | 2018-10-04 | 2018-09-05 | 2018-10-04 |
| patid | eventdate | period_start | period_end | period |
|---|---|---|---|---|
| cco001 | 2015-07-16 | 2015-04-20 | 2015-05-19 | 0 (referent window) |
| cco001 | 2015-07-16 | 2015-06-17 | 2015-07-16 | 1 (hazard window) |
| cco002 | 2017-03-25 | 2016-12-28 | 2017-01-26 | 0 (referent window) |
| cco002 | 2017-03-25 | 2017-02-24 | 2017-03-25 | 1 (hazard window) |
| cco003 | 2020-11-14 | 2020-08-19 | 2020-09-17 | 0 (referent window) |
| cco003 | 2020-11-14 | 2020-10-16 | 2020-11-14 | 1 (hazard window) |
| cco004 | 2015-10-17 | 2015-07-22 | 2015-08-20 | 0 (referent window) |
| cco004 | 2015-10-17 | 2015-09-18 | 2015-10-17 | 1 (hazard window) |
| cco005 | 2023-12-12 | 2023-09-16 | 2023-10-15 | 0 (referent window) |
| cco005 | 2023-12-12 | 2023-11-13 | 2023-12-12 | 1 (hazard window) |
| cco006 | 2018-10-04 | 2018-07-09 | 2018-08-07 | 0 (referent window) |
| cco006 | 2018-10-04 | 2018-09-05 | 2018-10-04 | 1 (hazard window) |
Next, two observation records are created for each patient, labelled by a variable (“period”) to distinguish hazard and referent windows. Using the “replace” command in STATA to redefine the period start and end dates. Patients whose risk periods fall outside the valid study period are excluded (see Table 4b).
2. Identify drug exposure in the risk periodsIn this step, both diagnosis and prescription datasets of the object drug (warfarin) and precipitant drug (clarithromycin) are used. In STATA, the “joinby” command merges the diagnosis dataset (Table 4b) with the prescription datasets (Table 3a and Table 3b). In the merged dataset, variables such as “rxst”, “rxen”, “period_start”, and “period_end” are used to determine whether an object drug or precipitant drug was prescribed during the hazard window or referent window. Example for one patient (cco001) is illustrated in Table 5.
| patid | eventdate | period_start | period_end | period | rxst | rxen | object |
|---|---|---|---|---|---|---|---|
| cco001 | 2015-07-16 | 2015-04-20 | 2015-05-19 | 0 | 2015-07-16 | 2017-03-23 | 0 |
| cco001 | 2015-07-16 | 2015-06-17 | 2015-07-16 | 1 | 2015-07-16 | 2017-03-23 | 1 |
| patid | eventdate | period_start | period_end | period | rxst | rxen | precipitant |
|---|---|---|---|---|---|---|---|
| cco001 | 2015-07-16 | 2015-04-20 | 2015-05-19 | 0 | 2015-05-16 | 2015-05-24 | 1 |
| cco001 | 2015-07-16 | 2015-06-17 | 2015-07-16 | 1 | 2015-05-16 | 2015-05-24 | 0 |
| cco001 | 2015-07-16 | 2015-04-20 | 2015-05-19 | 0 | 2015-05-26 | 2015-05-30 | 0 |
| cco001 | 2015-07-16 | 2015-06-17 | 2015-07-16 | 1 | 2015-05-26 | 2015-05-30 | 0 |
| cco001 | 2015-07-16 | 2015-04-20 | 2015-05-19 | 0 | 2015-06-01 | 2015-06-06 | 0 |
| cco001 | 2015-07-16 | 2015-06-17 | 2015-07-16 | 1 | 2015-06-01 | 2015-06-06 | 0 |
| cco001 | 2015-07-16 | 2015-04-20 | 2015-05-19 | 0 | 2015-06-08 | 2018-06-13 | 0 |
| cco001 | 2015-07-16 | 2015-06-17 | 2015-07-16 | 1 | 2015-06-08 | 2018-06-13 | 1 |
| patid | eventdate | period_start | period_end | period | object | precipitant |
|---|---|---|---|---|---|---|
| cco001 | 2015-07-16 | 2015-04-20 | 2015-05-19 | 0 | 0 | 1 |
| cco001 | 2015-07-16 | 2015-06-17 | 2015-07-16 | 1 | 1 | 1 |
Drug exposure is coded as a binary variable, indicating whether the drug was prescribed during the risk period. For example, if patient “cco001” had a prescription of object drug (warfarin) in the hazard window, the variable “object” was coded as “1” to indicate exposure (see Table 5). Duplicate records per patient and period are removed to keep one record per patient per period.
3. Generate variables in the 6-parameter modelThis step involves carefully constructing the variables needed to fit the saturated 6-parameter model of case-crossover design. The final dataset is generated as Table 6.
| patid | period | object | precipitant | objectonly | preciponly | joint | object_onprecip | precip_onobject | switch |
|---|---|---|---|---|---|---|---|---|---|
| cco001 | 0 | 0 | 1 | 0 | 0 | 0 | 0 | 0 | 0 |
| cco001 | 1 | 1 | 1 | 0 | 0 | 0 | 1 | 0 | 0 |
| cco002 | 0 | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| cco002 | 1 | 1 | 1 | 0 | 0 | 0 | 0 | 1 | 0 |
| cco003 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| cco003 | 1 | 1 | 0 | 1 | 0 | 0 | 0 | 0 | 0 |
| cco004 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| cco004 | 1 | 0 | 1 | 0 | 1 | 0 | 0 | 0 | 0 |
| cco005 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| cco005 | 1 | 1 | 1 | 0 | 0 | 1 | 0 | 0 | 0 |
| cco006 | 0 | 0 | 1 | 0 | 0 | 0 | 0 | 0 | 0 |
| cco006 | 1 | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 1 |
The six parameters on the specific exposure categories can be defined, as follows:
1) “objectonly”: Coded as 1 if the patient was exposed to the object drug in at least one period but never exposed to the precipitant drug (i.e., object drug exposure varies between periods, precipitant drug never exposed). Coded as 0 if the record shows no object drug exposure or if the conditions are not met. Example: patient cco003
2) “preciponly: Coded as 1 if the patient was exposed to the precipitant drug in at least one period but never exposed to the object drug (i.e., precipitant drug exposure varies between periods, object drug never exposed). Coded as 0 if the record shows no precipitant drug exposure or if the conditions are not met. Example: patient cco004
3) “joint”: Coded as 1 if the patient was concurrently exposed to both drugs during at least one period, and exposed to neither of the interacting drugs across all periods. Coded as 0 if otherwise. Example: patient cco005
4) “object_onprecip”: Coded as 1 if the patient was exposed to the precipitant drug in both hazard and referent period and the object drug exposure varied between different periods. Coded as 0 if otherwise. Example: patient cco001
5) “precip_onobject”: Coded as 1 if the patient was exposed to the object drug in both hazard and referent period and the precipitant drug exposure varied between different periods. Coded as 0 if otherwise. Example: patient cco002
6) “switch”: Coded as 1 if the patient switched exposure status for both drugs between periods (i.e., both drugs’ exposure changed from 0 to 1 or vice versa), indicating a change in exposure pattern. Coded as 0 if the conditions are not met. Example: patient cco006
4. Remove patients without discordant pairs of exposure between periodsIn the final step, patients who do not have a discordant period of exposure between the hazard period and referent period are removed.
The final dataset should contain one record per patient per period, with the six parameters encoded, ready for running conditional logistic regression. If an active comparator design is applied, a dataset for the object drug with a negative precipitant drug should be generated, following the same process as mentioned above.
Analysis Using Conditional Logistic RegressionConditional logistic regression is used to estimate the odds ratio for each stratum by comparing the odds of exposure during the hazard window to the odds of exposure in the referent window. If an active comparator design is applied, we should repeat the analysis for the active comparator and compute the comparator-adjusted estimate using a simple ratio approach. The Wald-test-based method is used to calculate the confidence interval. The STATA and R programme are available for implementing the analysis on GitHub (https://github.com/yqyqyy/CaseCrossover_DDI).
This framework is developed assuming that investigators use high quality data, for example, accurate recording of exposure, covariates and outcomes.
Currently, there is no guidance for interpreting results using the 6-parameter model of case-crossover study design for a DDI study. Thus, we propose a framework that can be used to facilitate the interpretation of findings using the 6-parameter model of case-crossover study design (see Box 1).
| First, set the pre-defined minimum clinically relevant effect size based on clinical context and expert clinical knowledge. | |
| In the 6-parameter model, there is a positive signal when one of the following scenarios occurs | |
| A. | higher risk of the outcome is observed in the stratum of initiation of precipitant drug in the presence of object drug (i.e., keeping exposure to object drug concordant in both hazard and referent periods); so that the point estimate is above the pre-defined minimum clinically relevant effect size and, this magnitude of risk is larger than the one associated with the stratum of initiation of the precipitant drug only; and Wald test suggests difference between these two strata; or |
| B. | higher risk of the outcome is observed in the stratum of initiation of object drug in the presence of precipitant drug (i.e., keeping exposure to precipitant drug concordant in both hazard and referent periods); so that the point estimate is above the pre-defined minimum clinically relevant effect size and, the magnitude of risk is larger than the one associated with the stratum of initiation of object drug only; and Wald test suggests difference between these two strata; or |
| C. | higher risk of the outcome is observed in the stratum of joint initiation of precipitant drug and object drug; so that the point estimate is above the pre-defined minimum clinically relevant effect size and, the magnitude of the risk is larger than the one associated with the stratum of initiation of object drug only or initiation of precipitant drug only; and Wald test suggests difference between these two strata |
Based on the three scenarios, we presented eight possible combinations in the 6-parameter case-crossover study design for interpreting the findings (see Table 7).
| Combination | Scenario A | Scenario B | Scenario C | Possible Interpretation |
|---|---|---|---|---|
| 1 | + | − | − | DDIs could be due to a poorly dose-titrated object drug, e.g., if a prescriber was unaware of reducing the dose of the object drug when initiating the precipitant drug to avoid a DDI. |
| 2 | − | + | − | DDIs could be due to a poorly dose-titrated precipitant drug, e.g., if a prescriber was unaware of reducing the dose of the precipitant drug when initiating the object drug to avoid a DDI. |
| 3 | + | + | − | Indicate a DDI. |
| 4 | + | − | + | Indicate a DDI. |
| 5 | − | + | + | Indicate a DDI. |
| 6 | + | + | + | Indicate a DDI. |
| 7 | − | − | + | Suggest confounding by indication, as the joint initiation of the two drugs may imply multiple medical conditions requiring separate treatments were present at the time point, and this multimorbidity rather than the drugs themselves could have driven poorer outcomes. |
| 8 | − | − | − | Indicate no DDIs. |
Abbreviation: DDIs, drug-drug interactions
+ positive signal; − no positive signal
Combination 1: Positive signals involving dose-titrated object drug
This combination shows findings of potential DDIs on the outcomes. However, the DDIs could be due to a poorly dose-titrated object drug, as only Scenario A is observed. In this case, the prescriber was unaware of reducing the dose of the object drug when initiating the precipitant drug to avoid a DDI.
Combination 2: Positive signals involving dose-titrated precipitant drug
This combination shows findings of potential DDIs on the outcomes. However, the DDIs could be due to a poorly dose-titrated precipitant drug, as only Scenario B is observed. In this case, the prescriber was unaware of reducing the dose of the precipitant drug when initiating the object drug to avoid a DDI.
Combination 3, 4, 5, 6: Positive signals
The combination of 3, 4, 5, and 6 suggests potential DDIs on the outcomes. In these combinations, at least two scenarios were observed.
Combination 7: Positive signal with confounding
This combination shows findings of potential DDIs on the outcomes. However, it may suggest confounding by indication, as only Scenario C is observed. The joint initiation of the two drugs may imply multiple medical conditions requiring separate treatments were present at the time point, and this multimorbidity rather than the drugs themselves could have driven poorer outcomes.
Combination 8: No positive signals
Given adequate power to detect the pre-defined minimum clinically relevant effect size in the strata of 6-parameter model case-crossover design (stratum of initiation of precipitant drug in the presence of object drug, and stratum of the initiation of object drug in the presence of precipitant drug), this scenario shows consistent findings of no harmful effect of DDIs on the outcomes.
Further considerations of the 6-parameter model and future development
In order to avoid solely using statistical significance to determine signals, we include minimum clinically relevant effect size to incorporate clinical relevance of the interaction into the interpretation framework. Similar to power calculation, investigators should justify the reason of how the value is set which should be context specific and guided by clinical knowledge.
Further, it is unclear how to directly assess the effect of persistent use of precipitant drug causing upward bias in 6-parameter model. Further studies are warranted to determine the minimum proportion of prescriptions >90 days that could lead to upward bias for practicality, as many medications are of long-term use.
We introduce the basic concept, advantages and limitations of a novel 6-parameter model case-crossover study, along with the implementation and interpretation, to assess the associations between concomitant drug use and adverse outcomes in DDI research to reduce time-invariant confounding and assess order of drug initiation patterns. Future design development should consider the impact of sustained use of interacting drugs to minimise bias.
This work was supported by the Laboratory of Data Discovery for Health (D24H) funded by AIR@InnoHK administered by Innovation and Technology Commission, the Government of the Hong Kong Special Administrative Region, China.
Funders had no role in the study design, collection, analysis, and interpretation of data; in the writing of the report; and in the decision to submit the article for publication.
This article solely provides an overview of the novel case-crossover study design and discusses its strengths and limitations. Therefore, ethics approval is not required.
Contributions are as follows:
Conceptualization AYSW
Writing (original draft) QY, AYSW;
All authors were involved in design and conceptual development and reviewed and approved the final manuscript.
We thank the AIR@InnoHK for providing funding to investigate drug interactions with oral anticoagulants, which has offered a valuable context for exploring methodologies in drug interaction-related pharmacoepidemiology studies.
AYSW has received honoraria from the 6th Annual Meeting of the Society for Clinical Epidemiology in Tokyo and 28th Annual Meeting of the Japanese Society for Pharmacoepidemiology in Kyoto in November 2023 and Henry Stewart Talks in 2024, outside the submitted work. All other authors have no conflicts of interest to disclose.
Angel YS Wong is one of the Editorial Board members of Annals of Clinical Epidemiology (ACE). This author was not involved in the peer-review or decision-making process for this paper.
