The dispute began when the authors of a letter to the editor pointed out a subtle but potentially consequential flaw in how the original study assigned patients to treatment groups. In the study, patients were classified as receiving prophylaxis only if they were seizure-free from the moment their symptoms began until they reached the hospital and treatment was started. That classification, the letter writers argued, creates a hidden asymmetry. A patient who seized in the minutes or hours between symptom onset and hospital arrival could never be counted in the prophylaxis group, because the treatment window had already closed. Such patients would automatically land in the no-prophylaxis group, inflating that group’s seizure rate and making the medication appear more protective than it truly is. This is the classic structure of immortal time bias, dressed in the specific clothing of emergency stroke care.

The research team, led by Margaret Banker of the Department of Preventive Medicine and Andrew M. Naidech of the Department of Neurology at Northwestern University Feinberg School of Medicine, responded with a combination of empirical data and formal sensitivity analysis. Their first line of defense was timing. The team had prospectively recorded the interval between symptom onset and initial presentation, whether that presentation occurred at their own facility or at a referring hospital. The median interval turned out to be 4.3 hours, with an interquartile range of 1.1 to 11.5 hours. In other words, the window during which an early seizure could occur before any treatment decision was possible is, for most patients, relatively brief. That brevity matters, because the magnitude of immortal time bias depends directly on how much unobservable time exists between the true start of follow-up and the moment treatment could logically begin.

Still, the researchers acknowledged that the critics’ proposed remedies, including a target trial emulation with clone-censor weighting, would be the gold standard approach for this kind of question. Such methods construct hypothetical randomized trials from observational data by cloning every patient into multiple treatment strategies and censoring each clone at the moment it deviates from its assigned strategy. The problem, the team explained, is practical rather than conceptual: those analyses require knowing the exact time of the first seizure, and that information was not captured. Time of symptom onset is a standard NIH Common Data Element for stroke research, but time of seizure is not a common data element in epilepsy care, so it never appeared on the case report forms. The researchers agreed that future studies should record both timestamps to enable the more rigorous analytic frameworks their critics favor.

Unable to run the ideal analysis, the team did the next best thing: they systematically dismantled the bias from both directions. The potential misclassification, they reasoned, could arise in two distinct ways. First, a seizure might occur almost immediately after symptom onset, before the patient could possibly reach a hospital and be offered prophylaxis. Second, a seizure might occur after a longer delay, but the patient might arrive so late that the treatment decision window had effectively passed. Each scenario implies a different pattern of misclassification, and each can be probed by excluding the patients most likely to embody it.

To address the first scenario, the researchers re-ran their analysis after excluding patients who presented very rapidly, within one, two, or four hours of symptom onset. These exclusions remove precisely the patients who could have seized before prophylaxis was even a possibility. The results barely moved. Across all three exclusion windows, the association between prophylactic antiseizure medication and reduced seizure risk remained essentially unchanged, suggesting that immediate pre-hospital seizures were not driving the observed effect. To address the second scenario, the team performed the mirror-image analysis, excluding patients who presented late, after six, eight, or ten hours. If delayed presentations were allowing extra time for early seizures to accumulate in the no-prophylaxis group, removing those patients should have attenuated the association. It did not. The direction and magnitude of the effect held steady, indicating that the findings are robust to how quickly or slowly patients reached medical care.

The researchers then went a step further, deploying a quantitative bias analysis, a technique that asks a deceptively simple question: how wrong would the data have to be to erase the finding? They calculated how many patients with seizures, falsely classified as not receiving prophylaxis, would need to be hypothetically reassigned to the prophylaxis group before the adjusted odds ratio crossed the null and the association vanished. The answer was striking: roughly forty percent. Only when approximately two in five of the seizure patients in the no-prophylaxis group were assumed to be misclassified did the protective signal disappear entirely. A misclassification rate of that scale, arising solely from seizures occurring in the narrow pre-hospital window, struck the authors as implausible given the observed presentation times. Their conclusion was unambiguous: the results are robust.

Beyond the statistical sparring, the exchange touches on a genuine clinical dilemma. Prophylactic antiseizure medication after lobar intracerebral hemorrhage remains controversial, because antiseizure drugs carry their own risks, and the evidence base for preventive use has long been thin. The researchers expressed appreciation for the call to conduct a prospective, randomized trial, but they were candid about its prospects. They themselves proposed exactly such a study, the Seizure Prophylaxis in Lobar ICH Treatment trial, known by the acronym SPLIT, but it was not given sufficient priority for funding. Meanwhile, the PEACH trial, a randomized, double-blind, placebo-controlled phase 3 study of prophylactic levetiracetam in acute intracerebral hemorrhage conducted in France, was halted due to low recruitment. The randomized evidence that clinicians and guideline writers crave may simply never materialize, leaving well-designed observational analyses as the best available evidence.

That reality shapes the research agenda going forward. The authors suggested that future work will likely focus on identifying which patients are most likely to experience a subsequent seizure, so that prophylaxis can be targeted to those who stand to benefit rather than applied indiscriminately. Their own data point to one such criterion: acute lobar intracerebral hemorrhage, in which bleeding occurs in the brain’s outer lobes rather than deep structures, appears to be a reasonable trigger for considering antiseizure prophylaxis. Lobar hemorrhages are known to carry a higher risk of post-hemorrhagic seizures than deep hemorrhages, plausibly because cortical blood irritates the brain’s electrical circuitry in ways that deep bleeds do not.

The broader lesson of the exchange extends well beyond this single clinical question. Immortal time bias has quietly distorted findings across the medical literature, from studies of cancer screening to intensive care interventions, and the letter-and-response format on display here represents peer review functioning as intended. Critics identified a real methodological vulnerability; the original authors responded not with dismissal but with layered sensitivity analyses, quantitative bias assessment, and an honest accounting of what their data could and could not support. The work was supported by the National Institutes of Health under grant R01NS117608, and both the original study and the rebuttal are open access, allowing clinicians, statisticians, and patients to examine the evidence directly. For now, the association between prophylactic antiseizure medication and fewer seizures after lobar brain hemorrhage stands, tested against one of epidemiology’s most persistent traps and emerging, by the authors’ account, intact.

Subject of Research: Immortal time bias in observational studies of prophylactic antiseizure medication after lobar intracerebral hemorrhage

Article Title: Response to: “Addressing Immortal Time Bias in Estimating the Effect of Prophylactic Antiseizure Medication After Lobar Intracerebral Hemorrhage”

Article References: Banker, M., & Naidech, A. M. (2026). Response to: “Addressing Immortal Time Bias in Estimating the Effect of Prophylactic Antiseizure Medication After Lobar Intracerebral Hemorrhage”. Neurocritical Care. https://doi.org/10.1007/s12028-026-02642-9

Image Credits: AI Generated

DOI: 10.1007/s12028-026-02642-9

Keywords: immortal time bias, intracerebral hemorrhage, antiseizure medication, seizure prophylaxis, neurocritical care, lobar hemorrhage, observational study, sensitivity analysis, quantitative bias analysis, levetiracetam, PEACH trial, epilepsy

The study under scrutiny, led by Banker and colleagues and published online in December 2025, examined patients with lobar intracerebral hemorrhage, a form of bleeding stroke that occurs in the lobes of the brain rather than deep structures such as the basal ganglia or thalamus. Lobar hemorrhages are strongly associated with cerebral amyloid angiopathy, an age-related condition in which amyloid protein deposits weaken vessel walls, and they carry a particularly high risk of subsequent seizures. Because clinical seizures can worsen outcomes after hemorrhagic stroke, physicians have long debated whether to prescribe preventive antiseizure drugs even in patients who have never seized, despite growing concern that these medications may themselves be harmful.

That concern was echoed in 2025 clinical practice guidelines from the Neurocritical Care Society, authored by Frontera and colleagues, which recommended against routine prophylactic antiseizure medication for most patients with nontraumatic intracerebral hemorrhage, citing trials suggesting possible harm, particularly with older agents such as phenytoin. Against this backdrop, observational studies comparing treated and untreated patients carry enormous weight. If such studies are methodologically sound, they can suggest when cautious prophylaxis might still help selected patients; if they are flawed, they can mislead clinicians and guideline developers alike.

The letter by Kobayashi and colleagues centers on the mechanics of immortal time bias, a problem first articulated in detail by epidemiologist Samy Suissa in a landmark 2008 paper in the American Journal of Epidemiology. The bias arises whenever a study classifies patients by a treatment that can only be started after they have survived for some period, yet counts that survival period as if it belonged to the treated group. Patients who die early, before ever receiving the drug, are automatically assigned to the untreated comparison group, even if they would have received the medication had they lived longer. The treated group thus appears to be populated by people who, by construction, could not die during the time window in question. They are, in a literal sense, immortal within that interval, and any apparent survival advantage may reflect nothing more than the way time was allocated.

In the context of lobar intracerebral hemorrhage, the problem is especially acute because the timing of antiseizure medication initiation is tightly linked to disease trajectory. Sicker patients, those with larger hemorrhages, deeper depressions of consciousness, or impending withdrawal of life-sustaining therapy, often die or transition to comfort care before prophylaxis is ever considered. Survivors who stabilize, by contrast, are the ones who live long enough to be started on medication. If the original analysis grouped patients from the moment of admission according to whether they ultimately received prophylaxis, while allowing the days before treatment to count toward the treated group’s follow-up time, the comparison would systematically favor the treated cohort for reasons unrelated to any pharmacological benefit.

Recent work has clarified just how pervasive and structurally varied this bias can be. In 2025, Hernán, Sterne, Higgins, Shrier, and Hernández-Díaz published a formal structural description in the journal Epidemiology showing how different causal structures generate immortal time, emphasizing that the bias is not a statistical nuisance to be adjusted away but a design-level misalignment between the question being asked and the analysis being performed. Their companion concept, the target trial, introduced by Hernán and colleagues in the Journal of Clinical Epidemiology in 2016, urges observational researchers to first specify the randomized trial they would ideally conduct, including a clearly defined time zero at which eligibility and treatment assignment both begin, and then emulate that design as faithfully as the data allow.

Applying that framework to seizure prophylaxis after lobar hemorrhage reveals the core difficulty: treatment is not assigned at a single moment but initiated at heterogeneous times, often days into the hospital course. Modern methods exist to handle this. The clone-censor-weight approach, increasingly used in pharmacoepidemiology and described in accessible detail by Gaber and colleagues in Cancer Medicine in 2024, creates artificial copies of each patient at time zero, one assigned to each treatment strategy, and then censors the copies when they deviate from their assigned strategy, using inverse probability weighting to correct for the informative censoring that results. This preserves the integrity of time zero and prevents the spurious survival advantage that plagues naive comparisons.

The exchange between the letter writers and the authors of the original study, with a formal response also published in Neurocritical Care, illustrates a broader and increasingly visible shift in how the medical literature polices itself. Methodological letters were once relegated to statistical footnotes; today, as large observational datasets drive clinical guidelines in neurocritical care, they sit at the center of practice-defining debates. The authors of the letter report no funding and no conflicts of interest and disclose that language-model tools were used only to refine phrasing and check citation consistency, with all content verified by the authors themselves. Their intervention does not claim that prophylactic antiseizure medication is definitively harmful or beneficial after lobar hemorrhage; it claims something more fundamental, that the existing estimate cannot be trusted until the time-zero problem is resolved.

For clinicians, the practical message is one of humility: observational evidence suggesting that prophylaxis is safe or effective in lobar hemorrhage should be weighed carefully against the randomized-trial signals of harm that shaped current guidelines. For researchers, the letter is a template for how to interrogate pharmacoepidemiologic analyses in acute neurology, where treatment timing, mortality, and disease severity are deeply entangled. As big-data studies continue to fill the evidence gaps left by trials that are difficult to conduct in critically ill stroke patients, the battle against immortal time bias, fought in the pages of methodological letters like this one, may prove as consequential for patient care as any new drug.

Subject of Research: Immortal time bias in observational studies of prophylactic antiseizure medication after lobar intracerebral hemorrhage

Article Title: Addressing Immortal Time Bias in Estimating the Effect of Prophylactic Antiseizure Medication after Lobar Intracerebral Hemorrhage

Article References: Kobayashi, G., Shibata, T., & Shiraishi, A. (2026). Addressing Immortal Time Bias in Estimating the Effect of Prophylactic Antiseizure Medication after Lobar Intracerebral Hemorrhage. Neurocritical Care. https://doi.org/10.1007/s12028-026-02643-8

Image Credits: AI Generated

DOI: 10.1007/s12028-026-02643-8

Keywords: immortal time bias, lobar intracerebral hemorrhage, antiseizure medication, seizure prophylaxis, neurocritical care, observational study design, target trial emulation, clone-censor-weight method, epidemiology, stroke, cerebral amyloid angiopathy, clinical guidelines