Design-Specific &
Temporal Biases
Evaluating Epidemiological Research — HSCI 230
Dr. Kiffer G. Card, Faculty of Health Sciences, Simon Fraser University
Learning objectives for this lesson:
- Identify biases unique to randomized controlled trials, including placebo effects, Hawthorne effects, and contamination
- Explain how inadequate allocation concealment and blinding inflate treatment effect estimates
- Define immortal time bias and describe how misclassifying person-time produces spurious protective associations
- Distinguish lead-time bias from genuine survival benefit in cancer screening evaluations
- Recognize time-window bias in case-control studies of medication effects
- Differentiate period effects from cohort effects in repeated cross-sectional data
- Critically evaluate whether epidemiological studies have adequately addressed design-specific and temporal bias threats
Glossary — Key Terms, People & Concepts
📚 Reference page — available throughout the lesson
This glossary collects the key concepts, people, and ideas you will meet in this lesson. Use it as a reference while you work through the material, or as a review before assessments. Type in the search box to filter entries.
Randomized Trial Biases
Introduction and Overview
Lessons 7–9 catalogued the three classical sources of bias one at a time — causal-specification, selection, and information bias. Lesson 10 picks up where Lesson 9 left off: rather than introducing a fourth category, it works through biases that are specific to particular study designs or that arise from the way time is handled in the analysis. These usually combine elements from the three classical categories in characteristic ways. The three content sections move from RCT-specific biases (Section 1: allocation concealment, blinding, placebo, Hawthorne, adherence) to time-related biases that haunt observational pharmacoepidemiology and screening evaluation (Section 2: immortal time bias, lead-time bias, length-biased sampling, overdiagnosis) to time-window and age–period–cohort effects that arise in any analysis of trends (Section 3). Lesson 11 will then bring the lesson on confounding to a close, and Lesson 12 will integrate the entire course.
Learning Objectives
- Explain why even properly randomized trials can yield biased effect estimates without adequate allocation concealment and blinding.
- Distinguish performance, detection, and reporting biases that operate after randomization.
- Describe how placebo effects, Hawthorne effects, contamination, and noncompliance distort intention-to-treat versus per-protocol estimates.
- Use meta-epidemiologic evidence to predict the direction and magnitude of bias from specific RCT design flaws.
Biases in Randomized Controlled Trials
Randomized controlled trials (RCTs) are often described as the “gold standard” of causal inference. However, RCTs are not immune to bias. Design-specific biases can systematically distort treatment effect estimates even within a randomized framework. Understanding these biases is critical for interpreting trial results and for designing robust studies.
Key Concept: Why RCTs Can Still Be Biased
Randomization addresses confounding by balancing known and unknown prognostic factors between groups at baseline. However, biases can arise after randomization through inadequate allocation concealment, lack of blinding, differential compliance, contamination, or the psychological effects of being in a trial. These biases affect internal validity even when randomization itself is successful.
Allocation Concealment and Blinding
Meta-epidemiologic studies—studies of studies—have documented that trials without adequate allocation concealment or blinding report systematically larger treatment effect sizes than trials with these safeguards. The Cochrane Collaboration has repeatedly confirmed these findings across hundreds of meta-analyses.
Schulz et al. (1995) analyzed 250 controlled trials from 33 meta-analyses and found that trials with inadequate allocation concealment yielded odds ratios that were exaggerated by an average of 30–40% compared to adequately concealed trials. Trials that were not double-blinded showed similar inflation. Wood et al. (2008) extended these findings across 146 meta-analyses, confirming that lack of blinding particularly inflated subjective outcome estimates (e.g., pain, functional status) while having less impact on objective outcomes like mortality.
Allocation concealment refers to procedures that prevent those enrolling participants from knowing upcoming treatment assignments. When concealment is inadequate, recruiters can selectively enroll sicker patients into the treatment arm or healthier patients into the control arm—or vice versa—introducing selection bias after randomization.
- Adequate methods: Central telephone randomization, sequentially numbered sealed opaque envelopes, pharmacy-controlled allocation
- Inadequate methods: Open random-number tables, unsealed envelopes, alternation by day of week or bed number
Blinding (masking) prevents participants, clinicians, or outcome assessors from knowing group assignment. Lack of blinding introduces several biases:
- Performance bias: Clinicians may provide differential co-interventions to unblinded groups
- Reporting bias: Participants may report outcomes differently based on perceived treatment allocation
- Detection bias: Assessors may interpret ambiguous outcomes differently when they know group assignment
Empirical evidence of bias magnitude:
| Methodological Flaw | Average Effect Inflation | Outcome Type Most Affected |
|---|---|---|
| Inadequate allocation concealment | 30–40% exaggerated OR | All outcomes |
| Lack of double-blinding | 15–25% exaggerated OR | Subjective outcomes (pain, function) |
| Both flaws combined | Up to 50% exaggerated OR | Subjective outcomes |
Allocation concealment and blinding address mechanical and observer-driven biases. The next family of trial-specific biases comes from the participants themselves — from what they expect to feel, how being studied changes their behaviour, and whether they actually take the treatment they were assigned.
Placebo Effects
The placebo effect refers to measurable improvements in participants who receive an inert treatment, driven by expectation, conditioning, and the therapeutic context (catalogued among the classical biases by Sackett, 1979). Placebo effects are particularly pronounced in trials involving pain and depression, where substantial improvement is routinely observed in placebo arms.
The magnitude of placebo response in pain trials has been increasing over time, particularly in U.S.-based trials, making it progressively harder to demonstrate superiority of active treatments over placebo.
')">This does not mean antidepressants are ineffective—but it demonstrates the enormous contribution of placebo response in psychiatric trials and the importance of placebo-controlled designs for establishing true drug efficacy.
')">Nocebo effects complicate the interpretation of adverse event reporting in trials and may inflate discontinuation rates in active treatment arms when participants know the expected side-effect profile.
')">Hawthorne Effect
The Hawthorne effect (Wikipedia) describes the phenomenon in which participants modify their behavior simply because they know they are being observed or studied. This effect is named after the famous Western Electric studies in the 1920s–1930s, where factory workers increased productivity regardless of which workplace change was introduced, apparently because they were being monitored. McCambridge, Witton, & Elbourne (2014) systematically reviewed evidence for the effect and emphasized that “research participation effects” are heterogeneous and design-dependent.
Srigley et al. (2014) measured hand hygiene compliance among healthcare workers using both direct observation (where workers knew they were being watched) and electronic monitoring (covert). Compliance rates were nearly three times higher when workers knew they were observed (estimated 70%+ vs. ~25% with electronic monitoring). This finding has profound implications for infection control studies: if the Hawthorne effect inflates compliance in all trial arms, the true baseline behavior is obscured, and interventions may appear less effective than they would be in unmonitored settings.
Compliance and Adherence Bias
Even in well-designed RCTs, not all participants adhere to their assigned treatment. Compliance bias arises when adherent participants differ systematically from non-adherent participants in ways that affect outcomes—a phenomenon sometimes called the “healthy adherer effect.”
Intention-to-treat (ITT) analysis includes all randomized participants in their assigned groups regardless of compliance. Per-protocol (PP) analysis includes only participants who adhered to the study protocol. PP analyses can introduce bias because compliant participants are systematically different from non-compliant ones (healthier, more motivated, higher socioeconomic status). The Coronary Drug Project Research Group (1980) demonstrated this dramatically: placebo adherent patients had 15% lower mortality than placebo non-adherent patients, confirming that adherence itself is a marker of overall health behavior.
Simpson et al. (2006) conducted a meta-analysis showing that good adherence to placebo was associated with lower mortality (pooled OR = 0.56, 95% CI: 0.43–0.74). This means that adherence is a proxy for a constellation of health-promoting behaviors. Per-protocol analyses that compare “adherers to drug” vs. “all controls” conflate drug effects with the healthy adherer effect, inflating apparent treatment benefits.
Contamination occurs when control group participants partially receive the intervention. This is particularly common in community intervention trials and pragmatic trials. For example, in the COMMIT (Community Intervention Trial for Smoking Cessation; COMMIT Research Group, 1995), control communities were exposed to national anti-smoking campaigns occurring simultaneously, which diluted the contrast between intervention and control and made the community-level intervention appear ineffective. Contamination biases results toward the null, reducing the apparent effect of the intervention.
Summary: RCT-Specific Biases
| Bias | Mechanism | Likely Direction | Primary Safeguard |
|---|---|---|---|
| Allocation concealment failure | Selective enrollment post-randomization | Away from null | Centralized randomization |
| Lack of blinding | Differential co-interventions, reporting, detection | Away from null | Double-blind, placebo control |
| Placebo effect | Expectation and conditioning | Inflates control improvement | Placebo-controlled design |
| Hawthorne effect | Behavior change from observation | Toward null (if equal in arms) | Covert measurement when ethical |
| Healthy adherer effect | Adherence confounded with health behaviors | Away from null (PP analysis) | ITT analysis |
| Contamination | Controls exposed to intervention | Toward null | Cluster randomization, geographic separation |
1. Meta-epidemiologic studies by Schulz et al. and Wood et al. found that trials without adequate allocation concealment report treatment effects that are:
2. In the Coronary Drug Project, participants who adhered to the placebo regimen had substantially lower mortality than placebo non-adherers. This finding demonstrates:
3. In the COMMIT smoking cessation trial, control communities were simultaneously exposed to national anti-smoking campaigns. This is an example of:
Immortal Time & Lead-Time Bias
Introduction and Overview
Section 1 covered biases that arise inside the controlled environment of an RCT. This section turns to a family of biases that haunt the messier observational designs — especially pharmacoepidemiology and screening evaluations — where time is handled poorly. All four biases in this section share a common engine: a comparison group that has somehow been guaranteed extra survival or extra opportunity to be exposed, by virtue of how the data were assembled rather than anything about the underlying biology. We start with the most consequential of them in observational drug research.
Learning Objectives
- Define immortal time bias and identify it in pharmacoepidemiologic study designs.
- Apply time-dependent exposure classification and landmark designs to remove immortal time bias.
- Distinguish lead-time bias, length-biased sampling, and overdiagnosis in screening studies, and explain why mortality is the only outcome immune to all three.
- Critique an observational drug or screening study for the role of time-related biases in its reported effect.
Immortal Time Bias
Immortal time bias occurs when a period of follow-up during which the outcome cannot occur is misclassified or improperly handled in the analysis (Lévesque, Hanley, Kezouh, & Suissa, 2010). The term “immortal” refers to the fact that participants must survive (remain event-free) long enough to be classified as exposed. When this survival requirement is not properly accounted for, the exposed group appears to have artificially better outcomes.
Walk through how a survival window before treatment can fake a treatment benefit. Next ▶ advances scenes.
A 6-scene visualization of immortal time: a patient timeline from diagnosis to death, the misclassified pre-treatment window, the spuriously divergent survival curves, the ghost of immortal time itself, and the time-varying-exposure fix.
Why “Immortal” Time?
Consider a study of statin use and mortality. To be classified as a “statin user,” a patient must survive long enough after hospital discharge to fill a prescription. The time between cohort entry and the first prescription is “immortal time”—by definition, the patient could not have died during this period and still been counted as exposed. If this time is credited to the exposed group (or excluded from the unexposed group), the result is a spurious survival advantage for statin users.
Suissa (2008) demonstrated that several widely cited observational studies reporting dramatic mortality reductions from post-MI statin use (reductions of 25–50%) contained immortal time bias. These studies classified patients as “statin users” based on prescriptions filled after hospital discharge. Patients who died before filling a prescription were automatically classified as “non-users,” and the survival time before the first prescription was either misattributed to the exposed group or excluded entirely. When Suissa reanalyzed the data using time-dependent exposure classification—correctly assigning person-time before the first prescription to the unexposed period—the dramatic survival benefit was substantially reduced or eliminated.
What you'll do: simulate a 2,000-patient cohort in which the drug has zero survival benefit, then run two analyses — one that ignores immortal time and one that handles it correctly. What to take away: the dramatic mortality reduction reported in some early observational drug studies is exactly the kind of artefact this code reproduces, and the fix is bookkeeping rather than biology.
Build a cohort where statins truly have no survival benefit, then misclassify person-time the wrong way. Watch a fake survival benefit appear out of nothing.
set.seed(230)
N <- 2000
# Each patient survives Exp(rate=1/3) years; 25% eventually fill a Rx.
# If they die before filling, they go in the unexposed group (no benefit).
death_t <- rexp(N, rate = 1/3)
rx_t <- rexp(N, rate = 1/2)
ever_rx <- rx_t < death_t
## (1) WRONG analysis: classify by ever_rx, ignore immortal time
mean_surv_wrong <- tapply(death_t, ever_rx, mean)
mean_surv_wrong # exposed group looks much better
## (2) RIGHT analysis: split person-time at the prescription date
unexp_pt <- pmin(death_t, rx_t) # every pt's unexposed person-time
exp_pt <- pmax(death_t - rx_t, 0) * ever_rx # only after Rx, only if they got it
unexp_events <- sum(!ever_rx)
exp_events <- sum(ever_rx)
c(unexp_rate = unexp_events / sum(unexp_pt),
exp_rate = exp_events / sum(exp_pt))The bias is engineered, not measured. Wrong person-time bookkeeping — not a real treatment effect — produced the apparent benefit. Time-dependent exposure assignment is the fix.
R Reflect on what you just ran
Use the questions below to interpret the output you produced. Look at your console before answering.
1. In the WRONG analysis, mean survival times printed as ~1.43 for the unexposed (FALSE) and ~4.27 for the exposed (TRUE). The simulation built in zero treatment effect — so where did the apparent 3-year survival advantage come from? Be specific about which person-time is being miscounted.
2. In the RIGHT analysis, the unexposed rate (0.337) and exposed rate (0.343) are nearly identical. Why is the rate of ~0.33 close to the simulated true death-rate parameter 1/3, and what does that confirm about the time-dependent classification you implemented?
rexp(n, rate = 1/3), so the true death rate is 1/3 per unit time ≈ 0.333. Both the corrected exposed rate (0.343) and unexposed rate (0.337) sit right at that value, confirming that the time-dependent classification — assigning person-time to unexposed until the moment of treatment, and to exposed only after — has eliminated the artefact. The agreement of both groups with the simulated truth is the cleanest test that the analytic fix works; the IRR is 1.02 instead of 3.0.3. Suissa's reanalysis of post-MI statin studies showed the same pattern. If you were peer-reviewing an observational drug study reporting a 25-50% mortality reduction, what specific feature of the methods section would you check first, based on what this simulation showed?
Immortal time bias introduces a systematic distortion through three related mechanisms:
- Survival requirement: Exposed participants must survive long enough to receive the exposure (e.g., fill a prescription, receive a transplant, attend a rehabilitation program)
- Time misclassification: The pre-exposure survival time is either credited to the exposed group (inflating their person-time denominator) or excluded from analysis
- Asymmetric comparison: The unexposed group includes all person-time from cohort entry, including deaths that occurred before exposed participants had the opportunity to be classified
The result is a comparison between a group with guaranteed survival during one period vs. a group with no such guarantee, producing a spurious protective effect.
Nobel Prize longevity study: A widely reported analysis suggested Nobel Prize winners live longer than nominees. However, the “exposure” (winning the prize) requires surviving to the date of the award. Nominees who died before the award ceremony contribute person-time only to the “non-winner” group. Correcting for immortal time eliminated the longevity advantage.
Metformin and cancer: Early observational studies reported that metformin users had dramatically reduced cancer risk compared to users of other diabetes drugs. Suissa and Azoulay (2012) showed that time-related biases, including immortal time bias, explained much of this apparent benefit.
Correcting immortal time bias:
- Time-dependent exposure classification: Model exposure as a time-varying covariate, so person-time before the first prescription is correctly assigned to the unexposed period
- Landmark analysis: Define a fixed time point after cohort entry and classify exposure status at that landmark; exclude patients who die or are lost before the landmark
- Nested case-control or case-crossover designs: Match on time at risk to ensure comparable exposure windows
- Target trial emulation: Design the analysis to mimic a hypothetical RCT, specifying time zero, eligibility, and treatment assignment simultaneously (Hernán & Robins, 2016)
Immortal time bias is the most common time-related bias in observational drug studies. The closely related family of biases that affect the evaluation of screening programs share the same fundamental issue — the way time enters the comparison — but they manifest differently. The next three subsections take them in turn.
Lead-Time Bias
Lead-time bias occurs when screening or early detection appears to improve survival simply because the diagnosis is made earlier, without actually changing the time of death. The “lead time” is the interval between screen detection and the time the disease would have been diagnosed clinically. If survival is measured from diagnosis, screened patients will appear to live longer even if screening does not change mortality.
Key Concept: Lead-Time Bias Illustrated
Imagine two identical patients with the same cancer developing at age 55 and causing death at age 70. Patient A is diagnosed by screening at age 58 (15-year survival from diagnosis). Patient B is diagnosed clinically at age 65 (5-year survival from diagnosis). Screening “added” 10 years of survival—but both patients lived to exactly 70. The 10-year difference is entirely lead time, not a genuine survival benefit. Only mortality-based outcomes (not survival from diagnosis) can distinguish true benefit from lead-time bias.
Following the introduction of PSA (prostate-specific antigen) screening in the late 1980s, 5-year survival rates for prostate cancer in the U.S. rose from approximately 75% to over 99%. However, prostate cancer mortality rates declined only modestly. Etzioni et al. (2002) estimated that lead-time bias accounted for the majority of the apparent survival improvement. Many screen-detected prostate cancers are slow-growing tumors that would never have caused symptoms or death (overdiagnosis), further inflating apparent survival statistics.
Lead-time bias inflates apparent survival without changing the time of death. The next bias goes one step further: it changes which tumors are even available to be detected.
Length-Biased Sampling
Length-biased sampling (or length bias) occurs when screening programs preferentially detect slower-growing, less aggressive tumors. Because slow-growing tumors have a longer detectable preclinical phase, they are more likely to be present during a screening examination. Fast-growing tumors, which have a shorter preclinical window, are more likely to present as “interval cancers” between screenings.
Summary: Temporal Biases in Screening and Pharmacoepidemiology
| Bias | Setting | Mechanism | Correction Strategy |
|---|---|---|---|
| Immortal time bias | Drug/exposure studies | Pre-exposure survival time misclassified | Time-dependent analysis, landmark analysis |
| Lead-time bias | Screening evaluation | Earlier diagnosis inflates survival from diagnosis | Use mortality, not survival, as endpoint |
| Length-biased sampling | Screening evaluation | Screening detects slower-growing tumors | Randomized screening trials, mortality endpoints |
| Overdiagnosis | Screening programs | Detection of nonprogressive disease | Compare incidence in screened vs. unscreened populations |
Reflection
A pharmaceutical company publishes an observational study showing that patients who start their new cholesterol medication within 30 days of a heart attack have 45% lower mortality than those who do not. The study classifies patients based on whether they filled a prescription within 30 days. Identify at least two temporal biases that could explain this finding, and propose an analytical approach that would produce a less biased estimate.
1. In observational studies of post-MI statin use and mortality, immortal time bias produces a spurious protective effect because:
2. After PSA screening was introduced, prostate cancer 5-year survival rates rose from ~75% to >99%, yet mortality rates declined only modestly. The most important explanation for this discrepancy is:
3. Length-biased sampling in cancer screening means that:
Time-Window, Period & Cohort Effects
Introduction and Overview
Sections 1 and 2 covered biases tied to specific designs (RCTs and observational pharmacoepidemiology / screening). This section covers two final time-related issues that can affect almost any observational analysis: time-window bias in case-control designs (a close cousin of immortal time bias from Section 2) and the age-period-cohort identification problem that arises whenever we try to interpret time trends in disease rates.
Learning Objectives
- Recognize time-window bias in case-control studies where exposure ascertainment periods differ between cases and controls.
- Distinguish age, period, and cohort effects and explain why the three cannot be separately identified without additional constraints.
- Interpret a time-trend graph using birth-cohort versus calendar-period framings, and articulate the assumptions each framing imposes.
- Propose design or analytic strategies (matching on exposure-window length, restriction, theory-driven APC constraints) to address temporal biases.
Time-Window Bias
Time-window bias occurs in case-control studies when the exposure opportunity period differs between cases and controls. If cases have a systematically shorter (or longer) time window during which exposure could be ascertained, comparisons of exposure prevalence will be distorted.
Key Concept: Time-Window Bias
In a case-control study of medication use and an acute outcome (e.g., myocardial infarction), cases who die shortly after the event have a truncated exposure window compared to surviving controls. If exposure is defined as “any use in the prior year,” controls have a full year to accumulate medication use while cases who die within months may not. Alternatively, cases with longer pre-event periods may accumulate more exposure. The mismatch in observation time creates differential exposure opportunity, biasing the association.
Suissa et al. (2006) showed that several case-control studies reporting protective effects of statins on cancer risk were affected by time-window bias. Controls were matched to cases on calendar date but had systematically longer exposure opportunity periods. Because statin use increased rapidly over time, controls—who were observed over longer and more recent periods—had a higher probability of statin exposure than cases. This differential created an artificial protective association. When exposure windows were properly aligned between cases and controls, the protective effect was substantially attenuated or disappeared.
Both biases involve time-related distortions, but they operate through different mechanisms:
- Immortal time bias occurs in cohort studies when pre-exposure person-time is misclassified, giving the exposed group a guaranteed survival advantage
- Time-window bias occurs in case-control studies when cases and controls have unequal opportunities to be classified as exposed, biasing exposure prevalence comparisons
Both can produce spurious associations, but the correction strategies differ: immortal time bias requires time-dependent analysis, while time-window bias requires matching or restricting on exposure opportunity period.
Strategies for addressing time-window bias include:
- Match on exposure opportunity: Ensure cases and controls have the same length of observation time for exposure assessment
- Use incidence density sampling: Select controls from the risk set at the time of each case’s event, ensuring comparable exposure windows
- Define fixed exposure windows: Restrict exposure assessment to a fixed period (e.g., 90 days before the index date) that is identical for cases and controls
- Sensitivity analyses: Vary the exposure window length and assess stability of results
Time-window bias is about the design of a single study. The last topic of the lesson zooms out to time-trend analyses themselves — surveillance data, cross-sectional surveys repeated across decades — and the famous identification problem that arises whenever we try to attribute change to a particular cause.
Period and Cohort Effects
When analyzing trends in disease rates or health behaviors over time, it is essential to distinguish between age effects, period effects, and cohort effects—the classic age-period-cohort (APC) problem.
Repeated cross-sectional surveys of smoking in the U.S. and U.K. show that overall smoking prevalence has declined steadily since the 1960s. However, examining trends by birth cohort reveals a more complex picture. Cohorts born in the 1920s–1940s had peak smoking rates above 50% among men and experienced gradual declines as anti-smoking campaigns took effect (a period effect). In contrast, cohorts born after 1970 never reached such high peak prevalence—reflecting a cohort effect of growing up in an environment where smoking was increasingly stigmatized and regulated. Failing to distinguish these effects leads to incorrect projections: a simple period model might predict convergence of all cohorts toward the same low rate, while a cohort model reveals that some older cohorts will maintain elevated rates until they exit the population through mortality.
The APC Identification Problem
Age, period, and cohort are mathematically collinear: Cohort = Period − Age. This means that given any two of these factors, the third is determined. This linear dependency creates a fundamental identification problem—it is impossible to simultaneously estimate all three effects without making additional assumptions or imposing constraints. Researchers must use theory, external data, or constrained models to disentangle these intertwined effects.
Summary: Time-Related Study Design Biases
| Bias / Effect | Study Design | Core Issue | Solution |
|---|---|---|---|
| Time-window bias | Case-control | Unequal exposure opportunity | Match on observation time, incidence density sampling |
| Period effect | Cross-sectional, ecological | Calendar-time changes affecting all ages | Age-period-cohort modeling |
| Cohort effect | Cross-sectional, ecological | Birth-year-specific risk trajectories | Age-period-cohort modeling |
| APC confounding | Any time-trend analysis | Mathematical collinearity of age, period, cohort | Theory-driven constraints, external validation |
Reflection
A researcher examines lung cancer rates over time and observes declining rates in men but rising rates in women. They attribute this to a period effect of anti-smoking campaigns being more effective for men. Using your knowledge of cohort effects, propose an alternative explanation and describe what data analysis you would need to distinguish between these two interpretations.
1. Time-window bias in case-control studies of medication use occurs when:
2. In repeated cross-sectional surveys, smoking prevalence among 50-year-olds has declined from 40% in 1980 to 15% in 2020. Which statement best distinguishes between period and cohort explanations?
3. The age-period-cohort identification problem exists because:
4. A researcher finds that statin use appears protective against cancer in a case-control study. However, controls had an average of 4.2 years of pharmacy records while cases (who died) had an average of 2.1 years. This discrepancy most likely introduces:
Final Assessment
Bringing It All Together
This lesson covered design-specific and temporal biases in three groups: the biases that operate inside RCTs even when randomization succeeds (Section 1), the biases that haunt observational pharmacoepidemiology and screening evaluations (Section 2), and the biases that arise from how time itself is handled in case-control comparisons and time-trend analyses (Section 3). Each bias here typically combines elements from the three classical categories of Lessons 7–9 — for example, immortal time bias is partly a measurement problem (how exposure status is assigned over time) and partly a selection problem (who survives long enough to be classified as exposed).
The unifying thread is that design choices and the way time enters the analysis can manufacture associations that look causal but are not. The corresponding repairs are also tied to design: blinding and allocation concealment for trials; time-dependent exposure classification and landmark designs for pharmacoepidemiology; mortality endpoints for screening evaluation; matched exposure windows and theory-driven constraints for time-trend analyses. The final reflection asks you to put the full inventory to work on a single hypothetical screening claim; the assessment then tests the conceptual material across all three sections before Lesson 11 closes the loop on confounding and Lesson 12 integrates the entire course.
Key Takeaways from Lesson 10
- RCT biases: Inadequate allocation concealment and blinding inflate effect estimates by 30–50%. Placebo effects, Hawthorne effects, compliance bias, and contamination further distort trial results.
- Immortal time bias: Misclassifying pre-exposure survival time creates spurious protective associations in pharmacoepidemiologic studies. Time-dependent analysis and landmark designs correct this bias.
- Lead-time bias: Earlier detection through screening inflates survival time from diagnosis without necessarily reducing mortality. Only mortality-based endpoints avoid this artifact.
- Length-biased sampling: Screening preferentially detects indolent tumors, making screen-detected cancers appear to have better prognosis regardless of treatment.
- Time-window bias: Unequal exposure ascertainment periods in case-control studies create differential exposure opportunity, biasing association estimates.
- Period vs. cohort effects: Changes in disease rates over time may reflect calendar-period influences on all ages or birth-cohort-specific risk trajectories. The APC identification problem requires theory-driven constraints to disentangle these effects.
The companion R script r-activities/HSCI_230_Lesson_10_Design_Specific_and_Temporal_Biases.R constructs a cohort in which a prescription has no real effect on survival, then contrasts a naive analysis (classify everyone by whether they ever filled the Rx, and compare mean survival) against a correct analysis that splits person-time at the prescription date. The wrong analysis manufactures a large survival advantage out of thin air; the right analysis recovers the truth that the two rates are essentially equal.
set.seed(230)
N <- 2000
# Each patient survives Exp(rate=1/3) years; 25% eventually fill a Rx.
# If they die before filling, they end up classified as unexposed.
death_t <- rexp(N, rate = 1/3)
rx_t <- rexp(N, rate = 1/2)
ever_rx <- rx_t < death_t
## (1) WRONG analysis: classify by ever_rx, ignore immortal time
mean_surv_wrong <- tapply(death_t, ever_rx, mean)
mean_surv_wrong # exposed group looks much better
## (2) RIGHT analysis: split person-time at the prescription date
unexp_pt <- pmin(death_t, rx_t) # unexposed person-time
exp_pt <- pmax(death_t - rx_t, 0) * ever_rx # exposed person-time
unexp_events <- sum(!ever_rx)
exp_events <- sum( ever_rx)
c(unexp_rate = unexp_events / sum(unexp_pt),
exp_rate = exp_events / sum(exp_pt))Final Reflection
Consider the following scenario: A national screening program reports that 5-year survival for a particular cancer has improved from 50% to 85% over 15 years, and attributes this improvement to the screening program. Drawing on everything you have learned in this lesson about design-specific and temporal biases, identify all the biases that could contribute to this apparent improvement, explain how each one operates, and describe what evidence you would need to determine whether the screening program truly reduced mortality.
Final Quiz
This 15-question assessment covers all sections of Lesson 10. You must score 100% to complete the lesson. Review the explanations for any questions you miss, then retry.
1. A meta-analysis combines results from 20 RCTs of a surgical procedure. Trials that used sham surgery controls showed a pooled OR of 1.1, while trials using usual care controls (no blinding) showed a pooled OR of 2.4. This pattern is best explained by:
2. In a workplace wellness RCT, the intervention group receives health coaching while the control group receives no contact. Both groups show improved health behaviors. The most likely explanation for improvement in the control group is:
3. The Coronary Drug Project found that placebo-adherent patients had 15% lower 5-year mortality than placebo non-adherent patients. This finding implies that:
4. A community-randomized trial of a dietary intervention finds no significant difference in cardiovascular events between intervention and control communities. However, process data show that 40% of control community residents adopted the dietary changes through media coverage of the trial. The true intervention effect is most likely:
5. An observational study reports that patients who complete cardiac rehabilitation after MI have 50% lower mortality than those who do not. This dramatic benefit could be partly explained by immortal time bias because:
6. A landmark analysis of statin use and mortality sets the landmark at 6 months post-MI and classifies patients based on statin use by that point. What is the primary advantage of this approach?
7. A screening program for neuroblastoma in infants detects many tumors, but a randomized trial shows no mortality reduction. The most complete explanation involves:
8. In a case-control study of NSAIDs and colorectal cancer, cases had pharmacy records covering a mean of 3.5 years while controls had records covering 6.2 years. If NSAIDs are truly unrelated to colorectal cancer, this study would most likely find:
9. Lung cancer rates among women aged 55–64 are higher in 2020 than in 1980, while rates among men of the same age declined. A researcher claims this reflects a period effect of anti-smoking campaigns being less effective for women. An alternative cohort-based explanation is:
10. A researcher wants to study whether a new diabetes drug reduces cancer risk using a healthcare database. Which study design feature most directly prevents immortal time bias?
11. In a trial of acupuncture for chronic pain, patients receiving real acupuncture improve by 40%, while patients receiving sham acupuncture (needles placed at non-acupuncture points) improve by 35%. Patients on a waitlist improve by 10%. These results suggest:
12. A cancer screening study reports that patients whose cancers were detected by screening have a 10-year survival rate of 90%, while patients whose cancers were detected clinically have a 10-year survival rate of 40%. This comparison is invalid for evaluating screening effectiveness because:
13. An investigator studying the effect of beta-blockers on heart failure hospitalization in a case-control study ensures that both cases and controls have exactly 365 days of pharmacy data before the index date. This design feature primarily addresses:
14. Obesity rates among 40-year-olds in 2020 are much higher than among 40-year-olds in 1980. If this is primarily a cohort effect rather than a period effect, you would expect to find:
15. A systematic review finds that observational studies of a medication report a 45% risk reduction, while the single large RCT reports only a 12% risk reduction. Considering design-specific biases, which combination of biases most likely explains the discrepancy?