When Did the Clock Start for These People? — How Observational Studies Manufacture Effects

In 2001, a study published in a medical journal drew major attention. It claimed that actors who won Academy Awards lived nearly 4 years longer than fellow actors who appeared in the same movies but did not win. Articles poured out. Interpretations followed that receiving recognition is good for the body, and that status extends life.

Five years later, other researchers recalculated the same data and pointed out a strange quirk in the original analysis. To win an award, you have to be alive on the day the award is given. Yet the original calculation counted the years the winners lived before they won the award toward the "period lived as a winner." During those years, those people could not possibly die. If they had died in the meantime, they would not have made the winner's list in the first place.

Once that period was stripped away and the numbers were rerun, the difference in lifespan shrank to just over a year. A difference of that magnitude could easily arise purely by chance. Cases like this are described as "not statistically significant" (PMID 16954361).

The only thing that changed was when each person started being counted. Neither the actors' lifespans nor the data changed.

Distortions like this do not happen only when comparing the lifespans of actors. The authors of the reanalysis noted that the same distortion frequently happens in observational studies asking whether a given treatment prolongs life (PMID 16954361). An observational study is one where researchers do not decide who gets which treatment, but instead retrace what already happened from records. Most studies using hospital records, insurance claims data, and national registries fall into this category.

---

This piece uses three terms learned in "What Is a Hazard Ratio of 0.7 the 0.7 Of? — How to Read a Survival Curve". The first is the hazard ratio. The hazard ratio is the value obtained by dividing the rate at which events occur at each time point in two groups. The second is censoring. People who do not experience the event by the end of follow-up or whose contact is lost midway are counted only up to the point known and then dropped from the calculation. This handling is called censoring. The third is the risk ratio, which was covered in "An Odds Ratio Is Not a Risk Ratio — What the 'Times More Likely' Is a Multiple Of". The risk ratio is the value obtained by dividing the proportion of people who experienced the event when followed to the end between two groups. The hazard ratio measures when events happen, whereas the risk ratio measures how many people experience them.

This piece covers what must be upheld for these mechanisms to produce valid answers. The conditions that must be upheld are two: whom we count starting from when, and who leaves the pool of subjects and why. Even if one of the two goes awry, curves are still drawn smoothly, and numbers come out down to two decimal places. Only, those numbers will end up measuring the flaws of the counting method instead of the effect of the drug.

The Period That Could Not Die — Immortality Time Bias

Let us look at the case where the first condition is misaligned. Bias refers to results leaning to one side due to the calculation method. Immortal time bias is a bias that occurs when the period spent waiting to receive treatment is mistakenly counted as the survival time of the treatment group. For a person to be grouped as a "treated person," they must stay alive until the day they receive treatment. Yet if that waiting period is included in the treatment group's survival records, the treatment group appears to have lived longer than they actually did.

The name "immortal time" comes from the fact that not a single person in the treatment group can die during that period. Anyone who died in the meantime would not have been able to receive treatment, and someone who did not receive treatment does not enter the treatment group.

An athlete study is the easiest example to understand. In several observational studies, results showed that "athletes with less training volume actually get injured more." This result came about at the point when training volume was measured. It happened because training volume was measured all at once at the end of follow-up. A player who got hurt early could not train afterward, so their training volume is recorded as low. They had low training volume because they got hurt, not because they got hurt due to low training volume (PMID 39447997).

In orthopedic research, a common condition like "including only patients followed for at least a certain number of years" does the exact same thing. Let us take 2 years, a commonly used criterion. The number itself is just an example to show the format. This condition quietly erases people who could not fill the 2 years and died or underwent re-operation. And it adds those 2 years of the survivors onto the scorecard (PMID 34253442).

In all three examples, the misaligned spot is the same. Time zero is the date set as "we count this person starting today," which is the moment their clock begins to tick. In a randomized controlled trial, which is a test where groups are divided by a draw of lots, time zero is set automatically. This is because the moment groups are assigned by a draw of lots is the exact moment the clock starts ticking. Observational studies have no such moment. Researchers dig through data and attach name tags later, saying "this person is the treatment group," "this person is the control group." If that name tag is attached after the clock, the time in between becomes free survival time wholesale.

Three things determine the magnitude of the bias (PMID 31155675). ① On average, how long does it take to receive treatment? ② What percentage of the total are treated people? ③ How long is the follow-up? The direction follows naturally from the definition. The longer the wait and the shorter the follow-up, the larger the bias becomes.

Immortal time can also make the life expectancy of the same people appear different across different periods. Life expectancy is the number of years one is expected to live further. When calculating the life expectancy of 33,867 people with intellectual disabilities from British medical records, the immortal time was simply included as the observation period. As a result, the value for 2000–2004 came out as 65.6 years, and the value for 2010–2014 using the same method came out as 58.0 years. The health of these people did not suddenly worsen during those 10 years. It was simply that the earlier the period, the longer the time "they had to stay alive until their diagnosis was recorded" was captured (PMID 35350993).

How Common Is It?

Immortal time bias is not rare in observational studies. Among 162 observational studies evaluating drug efficacy using medical and claims records, 35 (21.6%) were classified as being at high risk for immortal time bias. When 16 of them that could be re-calculated were adjusted, 4 (25%) had adjusted values that were statistically significantly different from the original estimates. For another 4 (25%), the significance determination itself flipped. In other words, the verdict changed between "there is an effect" and "there is no difference." Yet, out of the 35 studies at high risk, only 5 (14.3%) mentioned that their results might have been affected by this bias (PMID 41382023). The authors concluded that this bias can substantially distort conclusions.

There are also counts from other fields. Out of 60 papers evaluating post-surgical radiation therapy using a US cancer database, 23 (38.3%) failed to adjust for this bias (PMID 31404580). Among 20 observational studies covering drug use during pregnancy, 11 (55%) had this bias (PMID 35753061). Among 13 follow-up studies covering metformin and dementia, 8 (61.5%) either had this bias or did not address it clearly. And the studies that most frequently addressed biases arising from time calculation, such as immortal time bias, concluded with "no association" (PMID 36186728).

How Large Is the Bias?

Real drug names appear in this section. The drug names appearing here are examples to show the magnitude of bias, not judgments on whether those drugs are good or bad. Each number has meaning only within the design, subjects, and time points of those studies. Judgments on whether to use or not use a certain treatment are not in this article, and those judgments are left to the clinical evidence and practice guidelines of that field.

There is one way to read this. Place calculations with bias and without bias side by side, and look at the gap between them. The larger the gap, the more the bias moved the numbers. A risk ratio or hazard ratio of 1 means there is no difference between the two groups. The two numbers in parentheses are the confidence interval, which is the range where the true value is likely to lie. If this range encompasses 1, that result is indistinguishable from "no difference."

• Metformin and gastric cancer. Pooling 6 follow-up studies with immortal time bias made the risk look 33% lower (risk ratio 0.67, 95% confidence interval 0.59–0.77). Pooling 8 studies without bias showed no association (0.95, 0.85–1.05) (PMID 33453371).
• Beta-blockers and cancer prognosis. Pooling all 30 studies, the hazard ratio for all-cause mortality was 0.88 (0.79–0.97) and for cancer-specific mortality was 0.75 (0.64–0.88). When studies with a risk of bias were removed, the all-cause mortality side lost the difference, becoming 1.00 (0.93–1.07). The cancer-specific mortality side remained at 0.90 (0.83–0.98). However, the authors' own conclusion was that after removing the studies at risk of bias, they found no clinically meaningful evidence (PMID 27179912). The authors add that previous suggestions that beta-blockers benefit cancer survival may have relied on this bias (PMID 27179912).
• Atrial fibrillation procedures and stroke. 8 out of 10 studies (80%) had bias. When pooling only studies without bias, the reduction in stroke was not statistically significant (0.75, 0.49–1.02). This is because the interval stretches across and beyond 1 (PMID 32979508).
• Immunoglobulin and toxic shock syndrome. Toxic shock syndrome is a severe condition where bacterial infection causes the whole body to collapse in a matter of hours, and a drug called immunoglobulin is used for it. When recalculated using a tool that computes the magnitude of bias, the previously reported survival benefit was explained at least in part by this bias. The authors' conclusion was "it remains uncertain whether it is effective." They did not conclude that it is ineffective (PMID 40256040).
• SGLT-2 inhibitors (a class of diabetes drugs). 7 randomized trials and 23 observational studies were placed side by side. Some observational studies were reporting a 50% reduction in mortality. The results fluctuated wildly from study to study, and that fluctuation was explained by whether they were randomized trials or observational studies (all-cause mortality and cardiovascular mortality both p<0.001). However, hospitalization for heart failure was not the case (p=0.18) (PMID 33991127). The authors considered that the inflated results of the observational studies may have come from methodological problems such as immortal time bias.

Bias does not always make a drug look favorable. When observational studies dealing with gastric acid suppressants and cardiovascular events were pooled, the hazard ratio for cardiovascular death was 1.27 (1.11–1.44). When publication bias—the tendency for striking results to be published more easily—and the design bias of observational studies were adjusted for together, this value became 1.06 (0.96–1.16) (PMID 36195531). Here, bias was making the drug look dangerous.

Who Drops Out and Why — Informative Censoring

The second condition is who drops out of the counted group and why. As we reviewed earlier, the process of stopping the count at the exact point where we no longer know what happened to a person is censoring. Informative censoring is when the fact that a person drops out of follow-up itself contains information about that person's future. Simply put, it means dropping out was not a coincidence.

The reason dropping out must be a coincidence lies in the way a survival curve is drawn. A survival curve borrows the subsequent story of the person who dropped out from the people who remain. Therefore, for censoring to work properly, the subsequent fate of the person who dropped out of follow-up must be the same as that of the similar people who stay.

Insurance disenrollment is a good example. Among adults enrolled in private U.S. insurance, the proportion who disenroll midway through a year has remained steady at 13 to 14% every year. The factors predicting who drops out mid-year were health status. These included age, comorbidities, frailty, hospitalization, emergency room visits, and prescription drug use. In contrast, the factors predicting people dropping out all at once at the end of the year were administrative things like insurance product type or region (PMID 30788887). The data covered 2,050,000 person-years. A person-year is a unit that counts the same person again each year. Even for the same "disenrollment," mid-year disenrollment contains information about health, while end-year disenrollment does not. Therefore, in studies tracking people using insurance data, this kind of disenrollment can bias the calculations.

Even if the dropout rates in the two groups are equal, the result can be biased. It is commonly believed that if dropout rates are equal it is safe, but who drops out determines the result more than the rate. Let us think about a cancer drug trial. In one group, people quit because the side effects are severe. Then the frailest people drop out first in that group, and the curve drawn with the remaining people looks good. In the other group, people quit because they are disappointed with the assignment result. Here, people who are well-off and healthy enough to seek out other treatments outside drop out. The dropout rate is identical, but the type of people dropping out is exact opposites. When we model this situation, changing the fate of just 15% of the patients who dropped out early in each group wiped out the benefit in the period spent without the disease worsening (PMID 38382151).

Even in actual trials, there are traces of dropouts shaking up the results. Phase 3 trials are large-scale trials going up to the final stage involving humans. An intermediate indicator is an indicator used in place of survival, such as time until the disease worsens. Among Phase 3 trials published in the top three medical journals from 2010 to 2020, we looked at 73 papers that showed significant results in intermediate indicators. Among these, 33 papers (45%) also reported an overall survival benefit together, while 40 papers (55%) did not. Cases where dropouts were significantly clustered on the experimental group side appeared only in trials without an overall survival benefit (15% versus 0%). When partially reversing the dropout imbalance between the two groups and recalculating, 50% of the trials that had no overall survival benefit lost significance. For trials that had a benefit, it was 15% (PMID 34186504). The authors considered that informative censoring could explain the discrepancy where intermediate indicators improved but there was no survival benefit.

Methods to spot informative censoring are also being studied. One recent method looks at how early dropout patterns change as follow-up gets longer. In one trial, patients who dropped out early gradually disappeared as time passed. This means data was filled in belatedly, so it is not informative. In another trial, early dropouts remained as they were (PMID 41274169), and in such cases, concerns about informative censoring grow.

The direction of bias created by dropouts is also not predetermined. In a computer simulation modeled after infant vaccination research, cutting off follow-up at a specific time point could create bias in the direction making the vaccine look harmful. Conversely, the method of removing the early follow-up interval could create bias in the direction of undermining the beneficial effect (PMID 31071211). This means that even in the same data, the direction of bias splits depending on where you cut. This study investigated which direction bias occurs through computer simulation, and did not measure the effect of the vaccine.

Tools that "match comparison subjects so they are similar," like propensity score matching, do not solve this problem. Propensity score matching is a method of matching and comparing treated people with untreated people who have similar characteristics. This tool only handles the skewness of "who received treatment," and does not handle the skewness of "who dropped out of follow-up" (PMID 28992275). What propensity scores do and do not do was covered in "What Is Left After Adjustment — What E-values, Negative Controls and Propensity Scores Do and Do Not Do".

If you do not adjust for dropouts, the numbers actually move. When re-analyzing convalescent plasma therapy data for COVID-19 hospitalized patients, the curves unadjusted for dropouts were overestimating overall mortality (PMID 36676159).

How to Fix It — Resetting the Clock

Landmark analysis is a method that takes a single baseline time point as a new time zero and recounts from there. It is the oldest prescription, and the method is simple. You pre-specify a baseline time point, such as "6 months," and keep only the people who have survived up to that point. Then you divide the treatment group and the control group based on their status at that time point, and restart the clock from there. Doing this chops away the free survival that was previously tagging along.

Landmark analysis sometimes reveals hidden risks. When kidney transplant patients were calculated simply, the overall deaths of those who experienced a transplant failure and those who did not seemed similar. But when 30 months and 60 months were set as baseline time points, transplant failure was revealed as a clear risk factor (PMID 29024071). Here, the bias was masking the real risk instead of making the treatment look good.

Even with landmark analysis, bias may not all be washed away. In data looking at the effect of antiviral drugs in critically ill influenza patients, three standard analysis methods including landmark analysis all said there was some protective effect. But in a model that treats treatment as a value changing over time, the evidence for a protective effect disappeared (PMID 27546771). This model counts a person as untreated until they actually receive the drug. In a simulation from the same study, landmark analysis was also able to leave bias behind.

That is why the place recent methodology has reached is target trial emulation. Target trial emulation is a method where you first write down on paper, "If we were to actually conduct a randomized controlled trial answering this question, how would we design it?", and then realign observational data to that design. The items to write down are common-sense ones. Who is eligible to participate, what are the two strategies being compared, when does the clock start, and which group does each person fall into at that time point. If an observational study starts without writing these items down, the clock and the name tags get misaligned. If you write them down and start, the places where they would be misaligned are revealed at the design stage.

When you actually apply this, the answer changes.

Antibiotics during pregnancy and preterm birth. In this case, the direction of the effect flipped. The data are 111,403 birth records from Botswana. The researchers sequentially emulated 13 trials, one for each week from 24 to 37 weeks of pregnancy. The approach was to compare "people who started antibiotics right now" and "people who had not started yet" every week. In this method, the risk ratio for preterm birth was 1.11 (1.06–1.15). On the other hand, in a calculation where time zero was fixed at 24 weeks and "anyone who used antibiotics anytime between 24 and 37 weeks" was lumped together entirely into the antibiotic group, it came out as 0.90 (0.86–0.94). In this calculation, antibiotics appeared to prevent preterm birth (PMID 36805380).

Surgery for elderly lung cancer patients. We re-examined the benefit of surgery in early-stage lung cancer patients aged 70 to 89 using observational data. A simple survival curve with no adjustments calculated the 1-year survival difference as 22%. When we aligned the clock but ignored the informativeness of dropout, it became 17%, and when we even handled the informativeness of dropout, it became 11%. The weights used in the final step also adjusted for differences in patient characteristics between the two groups, so the shrunk magnitude is not entirely the fault of the clock (PMID 32386426). The important point is that the benefit did not disappear. It halved, but it remained.

Air pollution and lung cancer survival. We looked at fine particulate matter exposure and survival using California cancer registry data. In a model that averaged and entered exposure across the entire follow-up period, the hazard ratio per 5 µg/m³ of fine particulate matter was 1.38 (1.36–1.40). This is a value that looks like a strong risk. In the method that aligned the exposure measurement time point and the clock start, it was 0.99 (0.98–1.00) (the value from this method is an odds ratio). The authors concluded that the strong effect earlier likely came from a misalignment between the clock start and exposure assignment (PMID 37400995). The longer a person was followed, the longer the period that was averaged. Therefore, the value averaged across the entire follow-up period already has "how long this person was followed" mixed into it. The result has seeped into the exposure measurement.

Common Misconceptions

The first misconception is that "if there is bias, the entire effect is fake." Even when bias exists, a true effect can remain alongside it. A meta-analysis of 29 papers examining the relationship between side effects and treatment response in immuno-oncology is an example. In a simple calculation, the hazard ratio for overall mortality was 0.41. When recalculated using the landmark method, it decreased to 0.61 but remained significant. This means it remained of a size that is hard to dismiss as a coincidence. The authors' conclusion was that both the effect of bias and the real effect were present (PMID 33225800).

A meta-analysis on statins and breast cancer mortality is similar. This meta-analysis reported an overall protective association (hazard ratio 0.81, 0.75–0.87). At the same time, it also reported a tendency for the association to appear stronger in studies with a potential for bias (PMID 40500317).

The second misconception is that "this bias always makes the treatment look good." While it is a common tendency, it is not a rule. In the kidney transplant case we saw earlier, the bias masked a true risk factor (PMID 29024071), and in the stomach acid suppressant case, it made the drug look dangerous (PMID 36195531). The direction comes from how the clock and the nametag are misaligned, and it is not predetermined.

The third misconception is that "if it says landmark analysis was used, the problem is solved." There are two problems here. First, the answer changes depending on how many months are chosen as the landmark time. Yet in the US cancer database survey we saw earlier, 23 out of 31 papers (74.2%) that used landmarks did not explain why they chose that time point (PMID 31404580).

Second, landmark analysis itself has a structural limitation. The "untreated group" at the landmark time point gets mixed with people who will never receive treatment and people who will receive it later. This mixing makes the effect appear smaller than it actually is. Therefore, the distortion of this method is mostly toward "no effect," and the power to detect a true difference drops along with it (PMID 42493425).

An Honest Unresolved

The prescriptions introduced in this part are not finished tools.

Bias remains even in papers that use adjustment methods. When 113 papers that used specific adjustment methods to clear away immortal time bias were reviewed, 22 of them (19.5%) still harbored the potential for immortal time bias. Of these 113 papers, 91% claimed to emulate a target trial. Most of these were cases where participant eligibility was filtered using information after landmark time 0 (PMID 42600711). This means who entered the study was determined using information that could only be known after the clock started ticking. Naming a method and applying a method are different things.

It is also not yet agreed how large bias actually is in the literature. The rate at which conclusions flipped before and after adjustment was 64.3% in studies designed to demonstrate the method, and 32.8% in studies addressing real clinical questions. The rate at which the direction of the effect reversed was also split at 40.5% versus 15.5% (comparing 67 related studies, PMID 40865584). This means dramatic examples that make it into textbooks may not be the average of the actual literature.

Limitations of the data itself also remain. The British medical records study we saw earlier also tried an adjustment method using "the date the doctor entered the diagnosis" as landmark time 0. However, that date could not be trusted, so it could not be used (PMID 35350993). Even if you know where to set the clock, the data telling you that time might not exist.

And clearing away bias does not mean you have found the answer. In the toxic shock syndrome study we saw earlier, the answer that remained after clearing away bias was "we do not know." It was not even "no effect" (PMID 40256040).

How to Find This in a Paper

You just need to check four things.

1. When did the clock start ticking? In the methods section, find the sentence defining "start of follow-up," or time zero. See whether that point is the same as when treatment groups were assigned, or later than that.
2. Is the criterion for dividing groups based on information after time zero? Conditions like "took for 3 months or more," "administered 2 times or more," or "including only patients followed for 1 year or more" are all red flags.
3. Is the reason for missing people stated? If the reasons for dropping out differ between the two groups, it is a problem even if the proportions are the same.
4. If bias was addressed, how was it addressed? If landmark analysis was used, check whether it states why that specific time point was chosen.

---

The conclusion of this piece is one sentence. In an observational study, the statement "people who took the drug lived longer" is a statement about how things were counted before it is a statement about the drug. The counting method must be honest first for the subsequent sentence to have meaning.

Even after correcting start points and dropouts, problems remain. People who switch their assigned treatment midway, and people who pass away for other reasons before the event of interest occurs. Those problems are handled by "When People Don't Stay in Line — The Ones Who Switched, and the Ones Who Left for Another Reason".

Source Papers

Cases That Opened the Door

• "Do OSCAR winners live longer than less successful peers? A reanalysis of the evidence." (2006) — PMID 16954361 — The 4-year difference in the 2001 original study dropped to just over 1 year and became nonsignificant upon reanalysis.

Immortal Time Bias—Defining the Concept and Magnitude

• "Point: Immortal Time Bias—Determinants of Its Magnitude?" (2019) — PMID 31155675 — Entry-to-exposure interval, proportion of exposed individuals, follow-up length.
• "Exposure misclassification: an 'immortal time' bias … training load and injury." (2024) — PMID 39447997 — Measuring training load at the end of follow-up makes low load look like a hazard.
• "Immortal Time Bias … in Orthopedics." (2021) — PMID 34253442 — Requiring a minimum follow-up period creates immortal time (the "2 years" in the main text is an example).
• "Immortal time bias for life-long conditions … electronic health records." (2022) — PMID 35350993 — 33,867 individuals with intellectual disabilities. Unadjusted remaining life expectancy 65.6 years (2000–2004) versus 58.0 years (2010–2014). Adjustment by diagnosis entry date is unreliable.
• "An unjustified benefit: immortal time bias …" (2018) — PMID 29024071 — Kidney transplantation. After landmarks (30 and 60 months), transplant failure is revealed as a risk factor.
• "Immortal time bias in observational studies of time-to-event outcomes." (2016) — PMID 27546771 — Critically ill influenza. Treating exposure as a time-varying covariate makes the effect disappear.

How Common Is It?

• "Identifying, handling and impact of immortal time bias …" (2025) — PMID 41382023 — 35 high-risk out of 162 papers (21.6%). Out of 16 recalculated papers, 4 showed a significant difference and 4 had their conclusions reversed. Mentioned in authors' own papers 5 times (14.3%).
• "Immortal Time Bias in National Cancer Database Studies." (2020) — PMID 31404580 — 23 unadjusted out of 60 papers (38.3%). Out of 31 landmark papers, 23 (74.2%) did not state the rationale for landmark time selection.
• "Time-related biases in perinatal pharmacoepidemiology." (2022) — PMID 35753061 — 11 out of 20 papers (55%).
• "Metformin and Dementia Risk … Time Related Biases." (2022) — PMID 36186728 — 8 out of 13 follow-up studies (61.5%).

The Size of Bias — Calculations With and Without It

• "Immortal time bias exaggerates the effect of metformin on the risk of gastric cancer." (2021) — PMID 33453371 — hazard ratio 0.67 (0.59–0.77) versus 0.95 (0.85–1.05).
• "Beta blockers and cancer prognosis — The role of immortal time bias." (2016) — PMID 27179912 — all-cause mortality 0.88→1.00, cancer-specific mortality 0.75→0.90 (0.83–0.98). Authors' conclusion: "no clinically meaningful evidence after excluding studies at risk of bias".
• "Distorting Effect of Immortal Time Bias … Catheter Ablation." (2021) — PMID 32979508 — 8 out of 10 studies biased. Unbiased studies only 0.75 (0.49–1.02).
• "Quantifying Potential Immortal Time Bias … Acute Severe Infection." (2025) — PMID 40256040 — reported benefit of immunoglobulins explained at least in part by bias. Whether an effect exists remains uncertain.
• "The effect of SGLT-2 inhibitors … randomized trials versus observational studies." (2021) — PMID 33991127 — erratic nature of results explained by study type (mortality p<0.001), but not heart failure hospitalization (p=0.18).
• "Proton pump inhibitors and the risk of cardiovascular events …" (2022) — PMID 36195531 — cardiovascular mortality 1.27 (1.11–1.44) → 1.06 (0.96–1.16) after adjusting for publication bias and design bias.

Informative Censoring

• "Informative censoring by health plan disenrollment …" (2019) — PMID 30788887 — 2,050,000 person-years. Mid-year disenrollment predictors are health status, and end-of-year disenrollment is administrative factors.
• "Equal censoring but still informative …" (2024) — PMID 38382151 — Changing the fate of just 15% of early dropouts makes the benefit disappear.
• "Informative censoring of surrogate end-point data in phase 3 oncology trials." (2021) — PMID 34186504 — 73 articles. After restoring dropout balance, 50% of trials that had no overall survival benefit lost statistical significance (15% of those that did).
• "Identifying informative censoring from censoring patterns …" (2026) — PMID 41274169 — Discriminating by changes in early dropout patterns following follow-up extension.
• "Cautionary note: propensity score matching does not account for bias due to censoring." (2018) — PMID 28992275 — Confounding adjustment tools do not handle dropout bias.
• "Informative Censoring—A Cause of Bias in Estimating COVID-19 Mortality …" (2023) — PMID 36676159 — Unadjusted curves overestimate overall mortality.
• "Selection bias introduced by informative censoring … vaccination in infancy." (2019) — PMID 31071211 — This study investigated the direction of bias through simulation, and is not a study measuring the effectiveness of the vaccine. Depending on the cutoff location, the direction of the bias split in both directions.

Target Trial Emulation

• "Emulating Target Trials to Avoid Immortal Time Bias — antibiotics and preterm delivery." (2023) — PMID 36805380 — 13 sequential emulations risk ratio 1.11(1.06–1.15) versus biased calculation 0.90(0.86–0.94).
• "Reflection on modern methods: trial emulation …" (2020) — PMID 32386426 — elderly lung cancer surgery 1-year survival difference 22% → 17% → 11%.
• "Immortal Time Bias With Time-Varying Exposures in Environmental Epidemiology." (2023) — PMID 37400995 — particulate matter hazard ratio 1.38(1.36–1.40) versus time-aligned approach odds ratio 0.99(0.98–1.00).
• "Use of clone-censor-weight to avoid immortal-time bias …" (2026) — PMID 42600711 — among 113 papers using clone-censor-weight (91.2% claiming target trial emulation), 22 papers (19.5%) had remaining potential for bias.
• "On the Statistical Limitations of Landmark Analysis …" (2026) — PMID 42493425 — late-treated individuals mixed into the non-treatment group at the landmark time, shrinking the effect toward "none" and losing statistical power.

Remaining Debates

• "Immortal time bias tends to be more pronounced in methodological studies …" (2025) — PMID 40865584 — 67 papers. Discrepancy in conclusions 64.3% versus 32.8%, direction reversal 40.5% versus 15.5%.
• "Immortal time bias in the association between toxicity and response …" (2021) — PMID 33225800 — Overall mortality 0.41 → landmark 0.61. Both bias and true effect exist.
• "Statin use and breast cancer-specific mortality and recurrence …" (2025) — PMID 40500317 — Hazard ratio 0.81 (0.75–0.87). A tendency for associations to appear stronger in studies with potential for bias is also reported.