Potential Bias in Self‐Controlled Case Series Design: Impact of Incident Versus Prevalent Scenarios of Exposure and Outcome

ABSTRACT

Purpose

To assess changes in study results in self‐controlled case series (SCCS) studies and explore potential biases when applying incident and prevalent exposure and outcome scenarios.

Methods

We used the National Health Insurance Service database in South Korea to conduct two SCCS studies: (1) the risk of retinal detachment following fluoroquinolone use (expected elevated risk), and (2) the risk of acute cardiovascular disease following ranibizumab use (expected null association). Exposure and outcome scenarios were classified based on a washout period of 2 and 1 year, respectively, before the observation period: incident exposure‐incident outcome, incident exposure‐prevalent outcome, prevalent exposure‐incident outcome, and prevalent exposure‐prevalent outcome. For each scenario, conditional Poisson regression models, adjusted for time‐varying age, estimated incidence rate ratios (IRRs) with 95% confidence intervals (CIs).

Results

In the fluoroquinolone study, IRRs of retinal detachment were as follows: 1.83 (95% CI 1.72–1.96) for incident exposure‐incident outcome, 1.83 (1.71–1.95) for incident exposure‐prevalent outcome, 1.71 (1.62–1.80) for prevalent exposure‐incident outcome, and 1.70 (1.62–1.79) for prevalent exposure‐prevalent outcome scenario. In the ranibizumab study, IRRs of acute cardiovascular disease were 0.91 (0.84–0.99), 0.89 (0.82–0.97), 1.02 (0.94–1.10), and 1.00 (0.93–1.08), respectively.

Conclusions

In prevalent exposure scenarios, IRRs in both studies were attenuated, consistent with prevalent user bias. In prevalent outcome scenarios, where survivor bias was anticipated, no notable shifts in IRRs were observed, possibly due to the high severity and low recurrence of the outcomes associated with the respective exposures. Further consideration of prevalent user and survivor bias is important when interpreting SCCS results. Additional research is needed to quantify the impact of varying exposure and outcome definitions on SCCS estimates.

Summary.

• Two SCCS studies assessed the impact of incident and prevalent exposure and outcome scenarios on study estimates, investigating potential biases associated with different scenarios.

• In prevalent exposure scenarios, IRRs were attenuated, consistent with prevalent user bias.

• In prevalent outcome scenarios, no notable IRR shifts were observed, despite expected survivor bias, possibly due to the high severity and low recurrence of outcomes associated with the respective exposures.

• Further studies on incident and prevalent scenarios of various exposures and outcomes, and research to quantify the impact of varying exposure and outcome definitions on SCCS study estimates, are needed.

1. Purpose

The self‐controlled case series (SCCS) method is a case‐based design well suited for examining an association between transient exposures and acute outcomes. By eliminating the need for separate control groups and controlling for time‐invariant confounders, the SCCS design offers a valuable alternative to traditional cohort study designs [1]. Although several studies have investigated bias introduced by including prevalent users and those with prior outcomes in cohort studies [2, 3, 4], similar considerations have yet to be thoroughly explored in the context of SCCS studies.

This study aimed to explore the impact of incident and prevalent scenarios of exposure and outcome on the results of SCCS studies. We selected two previously conducted SCCS studies on the association between transient exposure and acute outcome, both involving new users with no prior diagnosis of the outcome [5, 6]. By varying exposure and outcome settings as incident or prevalent, we aimed to examine how results differ across the four scenarios and identify potential biases, such as prevalent user bias and survival bias, associated with each scenario.

2. Methods

2.1. Data Source

We used the National Health Insurance Service (NHIS) database in South Korea. The NHIS covers the entire population of > 50 million and provides comprehensive healthcare utilization information including sociodemographic information, diagnoses, prescriptions, and procedures. Diagnosis records are coded according to the International Classification of Disease 10th Revision, and prescriptions can be identified by national drug codes based on the Anatomic Therapeutic Chemical classification of the World Health Organization.

2.2. Incident and Prevalent Scenarios of Exposure and Outcome in SCCS Designs

We conducted two SCCS studies on (1) the risk of retinal detachment following fluoroquinolone use, where an increased association was expected, and (2) the risk of acute cardiovascular disease following ranibizumab use, where no association was expected. We applied a wash‐out period before the observation window to define four scenarios, in which the presence of at least one prescription or diagnosis record during the wash‐out period indicated prevalent exposure and prevalent outcome, respectively. The four scenarios explored in our study were: (1) incident exposure–incident outcome, (2) incident exposure–prevalent outcome, (3) prevalent exposure–incident outcome, and (4) prevalent exposure–prevalent outcome. To assess the impact of incident versus prevalent definitions, we used the incident exposure–incident outcome scenario as the reference. The detailed study design of incident and prevalent scenarios of exposure and outcome is shown in Figure S1.

2.3. Case Definition and Exposure Assessment

2.3.1. Study 1: Risk of Retinal Detachment Following Fluoroquinolone Use

We included patients aged 19 years and older who had at least one prescription for oral or injectable fluoroquinolones and a diagnosis of retinal detachment during the observation period from 2004 to 2015, with a wash‐out period from 2002 to 2003. Only patients who survived throughout the observation period were included to comply with the key assumption of the SCCS design [1]. Exclusion criteria included patients with a prescription of fluoroquinolone and a diagnosis of retinal detachment recorded on the same day, and patients with any prior diagnosis related to retinal disease prior to the outcome event (Table S1).

The observation period was divided into four periods: a 60‐day pre‐exposure risk period, an exposure risk period, a 60‐day post‐exposure risk period, and a baseline period [6]. If multiple diagnoses of retinal detachment were recorded within 30 days, only the first event was included to avoid potential bias from subsequent events being influenced by the initial event (Figure 1).

FIGURE 1.

Scheme of two self‐controlled case series designs. Abbreviations: FQ, fluoroquinolone.

2.3.2. Study 2: Risk of Acute Cardiovascular Disease Following Ranibizumab Use

We included patients aged 45 years and older who were diagnosed with acute cardiovascular disease (stroke or myocardial infarction) and received at least one ranibizumab injection for exudative age‐related macular degeneration during the observation period from 2009 to 2016, with a wash‐out period in 2008 (Table s2). Only patients who survived throughout the observation period were included.

Given that the standard treatment regimen for ranibizumab involves three injections at four‐week intervals, patients were considered exposed for 28 days following each injection. The observation period was divided into four periods: an 84‐day pre‐exposure risk period, an exposure risk period, a 28‐day post‐exposure risk period, and a baseline period [5]. As with the first case study, only the initial diagnosis was included when multiple cardiovascular events were recorded within 30 days (Figure 1).

2.4. Statistical Analysis

For each scenario, we calculated the incidence rates of the outcome during pre‐exposure risk periods and exposure risk periods and compared these rates with those during baseline periods. Incidence rate ratios (IRRs), adjusted for time‐varying age, and 95% confidence intervals (CIs) were estimated using conditional Poisson regression models. In Study 2, additional time‐varying adjustments were made for prescriptions of nonsteroidal anti‐inflammatory drugs, antithrombotic agents, and antihypertensive medications within 12 months prior to the outcome event. Sensitivity analyses were conducted by extending the wash‐out period by 1 year, which covered 2002–2004 for Study 1 (i.e., 3‐year washout period) and 2008–2009 for Study 2 (i.e., 2‐year washout period).

3. Results

3.1. Study 1: Risk of Retinal Detachment Following Fluoroquinolone Use

The IRRs of retinal detachment following fluoroquinolone use across the four scenarios were as follows: 1.83 (95% CI, 1.72–1.96) in 15 134 patients for incident exposure‐incident outcome, 1.83 (1.71–1.95) in 15 178 patients for incident exposure‐prevalent outcome, 1.71 (1.62–1.80) in 21 376 patients for prevalent exposure‐incident outcome, and 1.70 (1.62–1.79) in 21 455 patients for prevalent exposure‐prevalent outcome scenario. Extending the washout period by 1 year shifted the IRRs further from the null and showed a similar directional pattern observed in the main analysis across all four scenarios (Table 1).

TABLE 1.

Results of main analysis and sensitivity analysis extending a wash‐out period by 1 year across four exposure‐outcome scenarios.

Study	Incidence rate ratio (95% confidence intervals)
	Incident exposure‐incident outcome	Incident exposure‐prevalent outcome	Prevalent exposure‐incident outcome	Prevalent exposure‐prevalent outcome
Study 1: Risk of retinal detachment following fluoroquinolone use	N = 15 134	N = 15 178	N = 21 376	N = 21 455
Main analysis	1.83 (1.72–1.96)	1.83 (1.71–1.95)	1.71 (1.62–1.80)	1.70 (1.62–1.79)
Sensitivity analysis	2.31 (2.14–2.49)	2.30 (2.13–2.48)	1.96 (1.86–2.06)	1.93 (1.83–2.03)
Study 2: Risk of acute cardiovascular disease following ranibizumab use	N = 2039	N = 2219	N = 2211	N = 2407
Main analysis	0.91 (0.84–0.99)	0.89 (0.82–0.97)	1.02 (0.94–1.10)	1.00 (0.93–1.08)
Sensitivity analysis	0.89 (0.81–0.97)	0.82 (0.75–0.89)	0.87 (0.80–0.95)	0.82 (0.75–0.89)

3.2. Study 2: Risk of Acute Cardiovascular Disease Following Ranibizumab Use

The IRRs of acute cardiovascular disease following ranibizumab use across the four scenarios were as follows: 0.91 (95% CI, 0.84–0.99) in 2039 patients for incident exposure‐incident outcome, 0.89 (0.82–0.97) in 2219 patients for incident exposure‐prevalent outcome, 1.02 (0.94–1.10) in 2211 patients for prevalent exposure‐incident outcome, and 1.00 (0.93–1.08) in 2407 patients for prevalent exposure‐prevalent outcome scenario. Similarly, results from the sensitivity analysis were shifted further from the null (Table 1). Table s3 presents the IRRs for pre‐exposure risk periods.

4. Conclusions

Our study conducted two analyses to evaluate the impact of incident and prevalent scenarios of exposure and outcome on estimates in SCCS studies. Across the four scenarios, we found that variations in exposure scenarios had a more pronounced impact on the results, with bias observed in both directions, whereas variations in outcome scenarios had minimal impact.

In prevalent exposure scenarios, IRRs in both studies were attenuated, likely due to prevalent user bias, which may complicate the accurate assessment of the drug's true effect. A major concern with including prevalent users is depletion of susceptibles, where patients less likely to experience the outcome event are included [2]. In the fluoroquinolone study, where an increased risk was anticipated, the IRRs for prevalent exposure were attenuated, as expected. In the ranibizumab study, where no association was anticipated, the IRRs for prevalent exposure were slightly elevated. In studies where no association is expected, the direction of the IRR may change in either direction, influenced by factors such as exposure duration and risk periods, which can vary depending on study drugs. Therefore, further studies involving a broader range of drugs are essential to gain a more comprehensive understanding of incident and prevalent exposure scenarios.

Regarding prevalent outcome scenarios, survivor bias may arise because patients who die after outcome occurrence during the wash‐out period are excluded from the study. Survival bias can either underestimate or overestimate the drug's true effect. However, no notable shifts in IRRs were observed in our study, which may be attributed to the high severity of the study outcomes, likely leading to a low recurrence rate following exposure to the study drugs. Additionally, our study results should be interpreted with caution, as changing the definition of incident or prevalent outcomes did not largely affect the number of patients included in each respective scenario.

We applied a wash‐out period to determine the incident and prevalent scenarios. In the fluoroquinolone study, a 2‐year wash‐out period was applied because retinal detachment is an acute outcome with a low recurrence rate of 2.7% [7], necessitating a longer wash‐out period to accurately identify truly incident cases. Conversely, in the ranibizumab study, a 1‐year wash‐out period was considered sufficient, as myocardial infarction and stroke generally have an increased risk of recurrence after the first event [8, 9], making a 1‐year wash‐out period adequate to define truly incident cases. Additionally, in sensitivity analyses, we extended the wash‐out period to better capture truly incident cases and reduce the influence of prevalent user bias and survival bias. Although further studies examining different exposures and outcomes are warranted, the optimal length of the wash‐out period should be determined based on the clinical context of the condition under investigation.

Several limitations should be considered when interpreting our study results. First, due to the observational nature of the study, there is a possibility of unmeasured confounding. Additionally, time‐varying confounders, such as disease status, comorbidities, and comedications, may have influenced the results. However, this is likely to be non‐differential across the four scenarios. Second, the SCCS assumption that exposure is not affected by preceding outcome events may not always be met and could have influenced the results across the four scenarios. Furthermore, although our study compared the four scenarios based on IRRs in the exposure risk period, further comparisons of IRRs in the pre‐exposure periods, exposure risk periods stratified by time since initiation, and washout periods may be needed to fully understand the changes in IRRs across the four scenarios. Third, we assumed that prescription records accurately reflected drug consumption when defining the exposure risk period, which could lead to exposure misclassification. Moreover, in the absence of a gold standard, we used the incident exposure–incident outcome scenario as the reference, which limited our interpretation to the direction of the estimates. Further research on different exposures and outcomes, as well as studies that quantify the impact of varying exposure and outcome definitions on estimates in SCCS studies, is needed.

4.1. Plain Language Summary

This study aimed to understand how different ways of defining exposure and outcomes affect results in self‐controlled case series (SCCS) studies. Researchers conducted two SCCS studies using national claims data in South Korea: (1) the risk of retinal detachment after fluoroquinolone use (expected to show an increased risk), and (2) the risk of acute cardiovascular disease after ranibizumab use (expected to show no association). Incident and prevalent scenarios for both exposure and outcome were defined using washout periods, and the study examined the results for four scenarios: incident exposure–incident outcome, incident exposure–prevalent outcome, prevalent exposure–incident outcome, and prevalent exposure–prevalent outcome. In scenarios with prevalent exposure, the risk estimates were attenuated, as expected. In scenarios with prevalent outcome, there were no significant changes despite the anticipated survival bias, likely because the outcomes were severe and had a low chance of recurrence due to the study drugs. Further studies on incident and prevalent scenarios of various exposures and outcomes, and research to quantify the impact of varying exposure and outcome definitions on estimates in SCCS studies, are needed.

Ethics Statement

This study was approved by the Institutional Review Board of Sungkyunkwan University and performed in accordance with the principles of the Declaration of Helsinki. The need for informed consent was waived as this study was performed using anonymized claims data.

Conflicts of Interest

Ju‐Young Shin received grants from the Ministry of Food and Drug Safety, the Ministry of Health and Welfare, the National Research Foundation of Korea, and the Government‐wide R&D Fund for Infectious Disease Research and Pharmaceutical Companies, including Pfizer, Union Chimique Belge, and LG Chemical.

Supporting information

Table S1: Case definition of the study on the risk of retinal detachment following fluoroquinolone use.

Table S2: Case definition of the study on the risk of acute cardiovascular disease following ranibizumab use.

Table S3: Results in pre‐exposure risk period and exposure risk period.

Figure S1: Definition of four incident and prevalent scenarios of exposure and outcome.

Jung K., Byun S. J., Lim G., et al., “Potential Bias in Self‐Controlled Case Series Design: Impact of Incident Versus Prevalent Scenarios of Exposure and Outcome,” Pharmacoepidemiology and Drug Safety 34, no. 10 (2025): e70234, 10.1002/pds.70234.