Eduovisual
Biostatistics & Population Health
Loss to follow-up and missing data handling
— Loss to follow-up (LTFU): participants who started a study or care episode but cannot be reached for outcome ascertainment
— Missing data: any planned measurement not obtained — covariates, exposures, or outcomes
— Both threaten internal validity (bias) and external validity (generalizability) of trials, cohort studies, and real-world quality metrics
— Attrition rate >20% in an RCT → high risk of bias per Cochrane RoB 2
— Differential dropout between arms (e.g., 5% vs 18%) — even small absolute rates matter if asymmetric
— Trial reports "per-protocol" analysis only, omitting intention-to-treat (ITT)
— Cohort with long follow-up but no sensitivity analysis for missingness
— Quality measures (HEDIS, MIPS) where denominators shrink suspiciously
— Patient no-shows for HbA1c recheck, post-MI cardiology visit, TB DOT, HIV viral load, postpartum visit, colonoscopy after positive FIT
— Cancer screening programs where positive results are not followed by diagnostic workup ("screening without follow-through" is a patient safety event)
— Transitions of care: hospital discharge → PCP visit not completed within 7–14 days
— Step 3 expects you to (1) recognize bias introduced by LTFU, (2) choose the correct analytic approach, and (3) operationalize systems to prevent LTFU in your practice
Board pearl: If an RCT's results flip when you assume all LTFU patients in the treatment arm had bad outcomes ("worst-case" sensitivity analysis), the trial's conclusion is not robust — treat findings with skepticism regardless of the p-value.
Step 3 management: When a stem describes a patient who missed a follow-up for an abnormal screening result (e.g., positive FIT, abnormal mammogram, elevated BP), the next best step is active outreach (phone, patient navigator, certified letter) — not waiting for the patient to reschedule.
— MCAR (Missing Completely At Random): probability of missingness is unrelated to any variable, observed or unobserved
– Example: lab tube dropped in the lab; survey lost in mail
– Analyses remain unbiased even with complete-case analysis, just less precise
— MAR (Missing At Random): missingness depends on observed variables but not the missing value itself
– Example: older patients less likely to complete an online quality-of-life survey, but within each age stratum missingness is random
– Can be handled validly with multiple imputation or inverse-probability weighting using observed covariates
— MNAR (Missing Not At Random): missingness depends on the unobserved value itself
– Example: depressed patients skip the depression questionnaire because they feel hopeless; sickest patients drop out because they are dying
– No purely statistical fix — requires sensitivity analyses and assumptions
— Patients lost because they moved or changed insurance → often closer to MAR (predictable by demographics)
— Patients lost because of treatment side effects or disease progression → MNAR (informative censoring)
— Random equipment failure → MCAR
— Non-differential: equal across groups → biases effect toward the null typically
— Differential: unequal across exposure/outcome groups → bias can be in either direction, unpredictable
Key distinction: MCAR vs MAR vs MNAR is not testable from the data alone for MNAR — you can rule out MCAR statistically (e.g., Little's test), but distinguishing MAR from MNAR requires subject-matter knowledge and sensitivity analysis. Board stems hinting that "sicker patients dropped out" are pointing you toward MNAR / informative censoring.
Board pearl: ITT analysis preserves randomization but requires a plan for missing outcomes — it does not magically fix missing data.
— Compare baseline characteristics of completers vs. non-completers (Table 1 stratified by follow-up status)
– If they differ on age, comorbidity, SES → missingness is not MCAR
— Examine patterns of missingness: monotone (patients drop out and never return) vs. intermittent (occasional missed visits)
– Monotone patterns are typical of LTFU; intermittent suggests logistic issues
— Plot cumulative dropout curves by arm — divergence signals differential attrition
— % missing per variable — anything >5% deserves scrutiny; >40% often unsalvageable
— % complete cases — if only 50% have all variables, complete-case analysis loses massive power and likely biases estimates
— Little's MCAR test: p<0.05 rejects MCAR (but failing to reject doesn't prove MCAR)
— Predictors of missingness: logistic regression of "missing (yes/no)" on baseline covariates identifies MAR structure
— No-show rate by clinic, by provider, by demographic group
— Time-to-follow-up after abnormal result (e.g., median days from positive FIT to colonoscopy)
— % of discharged patients with completed PCP follow-up within 14 days
— Abnormal result with no documented patient notification
— "Patient will call to schedule" with no closed loop
— Specialty referral placed but never scheduled
Step 3 management: In a CCS-style outpatient case where you ordered a test at the last visit and the patient returns without it done, re-order the test, document the prior gap, and arrange a navigator/case manager follow-up — don't simply move on.
Board pearl: A randomized trial with >5% differential LTFU between arms warrants downgrading evidence quality (GRADE framework).
— Report n missing per variable, % missing, and pattern (monotone vs intermittent)
— Tabulate baseline characteristics by missingness status
— Provide CONSORT flow diagram for trials, STROBE for observational studies
— Complete-case analysis (listwise deletion): uses only subjects with no missing data
– Valid under MCAR; biased and inefficient under MAR/MNAR
– Default in most stats software — often the wrong default
— Available-case analysis (pairwise deletion): uses all available data for each calculation
– Sample size varies by analysis → can yield non-positive-definite covariance matrices
— Single imputation (mean, median, last-observation-carried-forward [LOCF]):
– LOCF is discouraged in modern guidance — underestimates variability, biased when disease progresses (e.g., Alzheimer's trials where worsening is expected)
– Mean imputation shrinks variance artificially → underestimates standard errors → false-positive results
— Multiple imputation (MI): generates m (typically 5–20) plausibly imputed datasets, analyzes each, pools via Rubin's rules to incorporate uncertainty
– Valid under MAR; standard of care in most journals
— Inverse probability weighting (IPW): weights observed cases by inverse probability of being observed
— Maximum likelihood / mixed models: for longitudinal data, naturally handle MAR missingness without explicit imputation
— Tipping-point / sensitivity analysis for MNAR
Key distinction: Imputation does not "make up" data — it propagates uncertainty so that inference reflects what we don't know. A single imputed value treated as real falsely narrows confidence intervals.
Board pearl: LOCF in a degenerative disease trial biases toward the null in the active arm (patients who drop out look artificially stable) — favors the placebo arm spuriously.
— Build an imputation model including the outcome and auxiliary variables correlated with missingness or the missing variable itself
— Run analysis on each imputed dataset, then pool point estimates and variances via Rubin's rules:
– Total variance = within-imputation variance + (1 + 1/m) × between-imputation variance
— More imputations (m=20–100) when fraction of missing information is high
— Censoring ≠ missing outcome if assumed non-informative (administrative censoring at study end is fine)
— Informative censoring (e.g., patients withdraw because they're dying) biases Kaplan-Meier and Cox estimates
— Sensitivity: competing risks analysis (Fine-Gray) when LTFU correlates with the event
— G-methods (g-formula, marginal structural models) for time-varying confounding with dropout
— Worst-case / best-case imputation: assume all missing outcomes in one arm are failures and all in the other arm are successes — if conclusion holds, robust
— Tipping-point analysis: find the assumed difference in missing outcomes that would change the conclusion; ask whether that scenario is plausible
— Pattern-mixture models: stratify by missingness pattern, model outcomes within each
— Selection models (Heckman) when missingness depends on unobserved outcome
— FDA / ICH E9(R1) addendum on estimands: trials must prespecify the target estimand, including how intercurrent events (treatment discontinuation, death, LTFU) are handled — strategies include treatment policy, hypothetical, composite, while-on-treatment, and principal stratum
Step 3 management: When reading a trial that reports only complete-case analysis with >10% LTFU and no sensitivity analysis, do not change practice based on its conclusions — flag the missing-data risk of bias.
Board pearl: Rubin's rules inflate SEs appropriately — that's the whole point of MI vs. single imputation.
— Non-differential LTFU (equal across exposure groups, missingness unrelated to outcome): typically biases toward the null (RR/OR → 1.0)
— Differential LTFU favoring exposed group (sicker controls drop out): biases away from null, exaggerating treatment effect
— Informative censoring in survival analysis: if sickest patients are lost, KM curves overestimate survival
— Healthy-volunteer / healthy-worker effect: healthier subjects remain in observational cohorts → underestimates risk of exposure
— Even 10% differential LTFU can shift hazard ratios by 0.1–0.3 in either direction
— High fraction of missing information (FMI) in MI signals fragile inference
— Wide confidence intervals after MI honestly reflect uncertainty — narrow CIs from complete-case analysis are often falsely precise
— <5% missing AND balanced across arms AND missingness unrelated to outcome
— Sensitivity analyses (worst-case, tipping-point, MI) all yield consistent direction and significance
— Prespecified estimand strategy in trial protocol
— Robust ITT analysis as primary
— Differential LTFU favoring the "winning" arm
— Per-protocol analysis as primary, ITT as secondary or absent
— No sensitivity analysis
— LTFU correlated with baseline severity or known prognostic factors
Key distinction: Per-protocol analysis answers "efficacy if perfectly adherent" but is vulnerable to selection bias from non-random dropout. ITT answers "effectiveness as offered" and preserves randomization but can dilute effect via non-adherence. Boards favor ITT as primary for superiority trials and per-protocol as co-primary for non-inferiority trials.
Board pearl: A non-inferiority trial with high LTFU is doubly dangerous — LTFU biases toward similarity (non-inferiority), making a true difference disappear.
— Automated reminders: text, phone, email 24–72h before visit → reduce no-shows by 20–40%
— Patient navigators / community health workers: especially effective for cancer screening follow-up (positive FIT → colonoscopy), HIV retention in care, prenatal care
— Open-access scheduling for follow-up labs (no appointment needed within a window)
— Co-located services: same-day labs, on-site specialty consults reduce drop-off
— Transitional care management (TCM) billing codes (CPT 99495/99496) incentivize 7–14 day post-discharge contact
— Patient portals with result release and automated abnormal-result flags
— Low health literacy, limited English proficiency → use professional interpreters, teach-back
— Housing insecurity, transportation barriers → social work, transportation vouchers
— Substance use disorder, serious mental illness → assertive community treatment models
— Postpartum patients (notorious LTFU) → schedule visit before discharge, 3-week and 12-week ACOG visits
— Adolescents transitioning to adult care → structured transition programs
— Every abnormal result requires documented patient notification AND documented action plan
— EHR safety nets: overdue test trackers, abnormal-result dashboards, registry-based outreach
— Two-person verification for critical results (e.g., new cancer diagnosis)
Step 3 management: A 58-year-old has a FIT-positive result; the office called once, left a voicemail, and the patient never returned the call. Next step: send a certified letter and engage a patient navigator, NOT "wait for patient to reschedule." Failure to close the loop on a positive cancer screening is a sentinel-level patient safety event.
Board pearl: TCM codes require contact within 2 business days of discharge and a face-to-face visit within 7 or 14 days depending on complexity — a high-yield Step 3 systems fact.
— Plan: identify high-risk transitions (ED discharge, hospital discharge, abnormal screening, postpartum)
— Do: implement bundle (reminder + navigator + closed-loop tracking)
— Study: measure no-show rate, time-to-follow-up, % closed loops
— Act: scale or adapt
— Process: % of patients receiving reminder, % contacted within 48h of discharge, % of positive screens with diagnostic follow-up within 60–90 days
— Outcome: readmission rate, 30-day mortality, screening completion rate, HEDIS measures
— Balancing measures: staff burden, cost per patient retained
— Disease registries (DM, HTN, HF, HIV, oncology) flag overdue patients
— Outreach pyramid: portal message → text → call → letter → home visit
— Pre-visit planning: huddle reviews labs/imaging needed before patient arrives
— High-utilizer patients benefit from intensive case management
— LACE index, HOSPITAL score predict readmission risk → targeted TCM
— Hard stops in EHR for unacknowledged critical results
— Automatic referral tracking with expected close date
— Bidirectional referral platforms confirming specialist visit completion
— Stratify LTFU rates by race/ethnicity, language, insurance, ZIP code
— Disparities in LTFU drive disparities in outcomes (cancer, maternal mortality)
— Targeted interventions for groups with highest LTFU
CCS pearl: In a CCS case spanning weeks, always schedule the next visit, order needed labs to be drawn before the visit, and document patient education at each encounter — the clock advances, and unscheduled follow-up costs you points and risks "missed" diagnoses.
Board pearl: Closed-loop referral systems reduce serious diagnostic errors — a Joint Commission and AHRQ priority area.
— Multiple chronic conditions → fragmented specialty care, high visit burden
— Hearing/vision impairment limits phone and portal reminders
— Cognitive impairment → missed appointments, lost paperwork, polypharmacy errors
— Transportation dependence; loss of driving
— Caregiver burden and turnover
— Frailty → hospitalizations interrupt outpatient continuity
— Geriatric trial participants drop out more often due to death, hospitalization, or adverse events → informative censoring (MNAR)
— Cognitive testing trials especially vulnerable: patients with worsening dementia drop out → LOCF biases toward null of disease progression
— Modern Alzheimer's trials use mixed models for repeated measures (MMRM) and prespecified estimands rather than LOCF
— Identify a designated caregiver/health proxy and include in scheduling
— Consolidate visits — "annual wellness visit" bundling labs, screenings, ACP
— Home-based primary care for homebound seniors
— Pharmacist-led med reconciliation post-discharge
— Telehealth visits to reduce transportation barriers (with attention to digital literacy)
Step 3 management: An 82-year-old with CKD stage 4 has missed two nephrology appointments. Next step: engage social work, arrange transportation, and consider home-based or telehealth nephrology — not discharge from the practice. Loss of CKD follow-up risks unplanned dialysis initiation (associated with higher mortality).
Board pearl: Unplanned dialysis start (no AV fistula, central line required) carries ~2× higher 1-year mortality vs. planned start — a downstream consequence of LTFU.
— ~40% of US patients miss the 6-week postpartum visit
— Higher rates among Medicaid, Black, and rural patients
— Consequences: missed postpartum depression, untreated hypertensive disorders, missed contraception, undiagnosed diabetes after GDM
— ACOG recommendation: postpartum care as an ongoing process — contact within 3 weeks, comprehensive visit by 12 weeks
— Postpartum hypertension follow-up: BP check within 3–10 days of delivery for preeclampsia/gHTN
— Adolescent transitions to adult care: structured handoffs with shared visits
— Foster care: high LTFU; AAP recommends initial visit within 72h, comprehensive within 30 days, then per Bright Futures
— Children with special healthcare needs require medical home model
— Vaccine catch-up after missed visits per CDC catch-up schedule
— Diagnosed → linked to care → retained in care → on ART → virally suppressed
— Each step loses patients; retention (≥2 visits/year, ≥3 months apart) is HRSA metric
— Re-engagement programs (Data to Care) use surveillance data to find LTFU patients
— Directly observed therapy reduces LTFU and resistance
— Public health departments have authority to enforce treatment for active TB (legal exception to autonomy)
— High LTFU; low-barrier buprenorphine, contingency management, peer navigators improve retention
Key distinction: Postpartum visit completion is now a HEDIS quality measure and tied to value-based payment. Step 3 favors proactive scheduling before hospital discharge and early (3-week) contact rather than waiting for the 6-week visit.
Board pearl: A patient with severe-range BP postpartum needs follow-up within 72 hours — LTFU here is a major driver of postpartum maternal mortality.
— Missed diagnoses: delayed cancer dx after abnormal screening (FIT, mammogram, Pap, LDCT)
— Medication errors: unmonitored warfarin, lithium, methotrexate, immunosuppressants
— Disease progression: uncontrolled DM, HTN, HIV viral rebound, transplant rejection
— Avoidable hospitalizations and readmissions
— Maternal and neonatal mortality from postpartum gaps
— Mental health crises from missed psychiatric follow-up post-discharge (suicide risk highest in first 30 days)
— Type I error inflation (false positives) from single imputation underestimating SEs
— Biased effect estimates from complete-case analysis under MAR/MNAR
— Overestimated survival in oncology trials with informative censoring
— Replication failures when initial trials had high LTFU
— Regulatory delays/rejections (FDA citations for inadequate missing-data handling)
— Reduced quality measure performance → financial penalties (HRRP, MIPS)
— Diagnostic error claims — leading source of malpractice payouts in outpatient settings
— Health inequities perpetuated when LTFU concentrates in marginalized groups
— Erosion of trust when results are not communicated
— Failure to communicate critical test results is a Joint Commission National Patient Safety Goal
— A positive cancer screen with no follow-up for >12 months is reportable as a serious safety event in many institutions
Step 3 management: A patient discharged on warfarin for new AF misses the INR check; INR comes back 8.2 three weeks later with hematuria. The root cause is a transition-of-care failure — system-level fix (anticoag clinic auto-enrollment, scheduled INR before discharge, navigator follow-up), not blaming the patient.
Board pearl: Diagnostic error is the #1 source of paid malpractice claims in outpatient medicine, and failure to follow up on test results is a top mechanism.
— Level 1: automated reminder (text, portal, robo-call) 24–72h prior
— Level 2: live phone call from MA/RN after missed visit (within 24–48h)
— Level 3: patient navigator outreach within 1–2 weeks, including barrier assessment
— Level 4: certified letter documenting attempts and risks of non-follow-up
— Level 5: home visit (community health worker) for high-risk conditions
— Level 6: public health / legal escalation for reportable communicable disease (active TB, syphilis with neuro involvement, measles)
— Newly diagnosed cancer (any solid tumor, hematologic malignancy)
— Active TB, untreated HIV with high viral load and TB co-infection
— Critical lab values (K >6.5, glucose >500, INR >5)
— Imaging findings concerning for malignancy or aneurysm
— Severe-range BP in pregnancy/postpartum
— Suicidal ideation post-ED discharge
— Communicable diseases per state list (TB, HIV in many states, STIs, hepatitis, meningococcal, COVID per local rules)
— Child/elder/dependent adult abuse suspicion
— Certain firearm injuries, intimate partner violence (state-dependent)
— Impaired drivers in some states
— Repeated LTFU on critical results
— Patient harm event from LTFU → root cause analysis (RCA), institutional safety review
Step 3 management: A patient with newly diagnosed pulmonary TB on the smear is not returning calls. Next step: notify the local public health department, which has legal authority to locate, evaluate, initiate DOT, and in extreme cases obtain a court order for treatment/quarantine. Do not simply discharge from your clinic.
CCS pearl: In CCS, when test results are critical, "Call patient" and "Counsel patient" are clickable actions — use them and document, then schedule the follow-up.
— MCAR: equipment failure, random survey loss, administrative censoring at predetermined trial end
– Statistical fix: any method valid (complete case is unbiased)
— MAR: missingness predictable by observed variables (age, sex, baseline severity recorded before dropout)
– Statistical fix: multiple imputation, IPW, ML/mixed models with relevant covariates
— MNAR: missingness depends on unobserved values (depression score, current pain, undocumented disease worsening)
– Statistical fix: pattern-mixture, selection models, sensitivity analyses; cannot fix purely with data
— Right censoring (administrative): subject hasn't had event by end of study — non-informative, handled in KM/Cox
— Right censoring (informative): subject lost because they're getting worse — biases survival estimates
— Left censoring: event occurred before observation began (e.g., HIV seroconversion date unknown)
— Interval censoring: event occurred between two visits — common in screening studies
— Treatment discontinuation — handled by treatment policy (ITT-like), hypothetical, composite, while-on-treatment, or principal stratum strategies
— Use of rescue medication
— Death — often handled as composite outcome or principal stratum
— LTFU — typically requires hypothetical strategy with MI
— Berkson bias (hospital-based studies)
— Healthy-worker / healthy-volunteer effect
— Survivor bias (only those surviving long enough are studied)
— Immortal time bias (misclassified person-time before exposure could occur)
Key distinction: Censoring and missing outcome look similar but censoring carries the assumption (often unverified) that the subject would have continued at the same rate. If censoring is informative, treat it as a missing-data problem and run sensitivity analyses.
Board pearl: ITT analysis preserves randomization but requires outcome data — missing outcomes still need imputation or sensitivity analysis.
— Non-adherence: patient enrolled, in follow-up, but not taking assigned treatment
– ITT preserves randomization despite non-adherence; per-protocol analysis excludes them
– Complier-average causal effect (CACE) estimates effect among adherers using instrumental variables
— LTFU: patient not observed at all — outcome unknown
— Both can co-occur and compound bias
— Entry selection bias: who agreed to enroll (volunteer bias) → affects generalizability
— LTFU: who stays → affects internal validity
— Different fixes; don't conflate
— Confounding: distortion from a third variable affecting exposure and outcome — fixed by randomization, adjustment, matching, IPW
— LTFU bias: distortion from incomplete observation
— A study can have neither, either, or both
— Misclassified data is present but wrong (vs missing)
— Differential misclassification ≈ differential LTFU in spirit — both bias estimates unpredictably
— Different mechanisms but similar effect: differential information across groups
— Studies with LTFU often unpublished if results are null → meta-analyses inherit bias
— Selective outcome reporting compounds missing-data issues
— Patients who adhere to placebo also have better outcomes — placebo adherers have ~50% lower mortality than non-adherers
— Confirms that adherence is a marker of healthy behaviors, not just drug effect
— Strengthens case for ITT as primary analysis
Key distinction: Randomization protects against confounding at baseline but does not protect against post-randomization bias from differential LTFU or non-adherence — these require analytic vigilance.
Board pearl: A trial with 25% LTFU evenly distributed but where dropouts in both arms had worse baseline prognosis still has biased absolute risk estimates, even if the relative effect is preserved.
— Patient-centered medical home (PCMH): team-based care, panel management, registries
— Chronic Care Model: self-management support, decision support, delivery system design, clinical information systems
— Accountable Care Organizations (ACOs) financially incentivize population retention
— Value-based payment ties reimbursement to follow-up completion (HEDIS, MIPS)
— Scheduled follow-up appointment before discharge (PCP within 7–14d post-hospital, specialty as indicated)
— Medication reconciliation with pharmacist review for high-risk meds
— Written discharge summary in patient's language at appropriate health literacy
— Teach-back confirming understanding of warning signs and follow-up plan
— Transitional care management call within 2 business days
— Direct EHR communication to receiving PCP
— Post-MI: cardiac rehab referral (Class I); reduces mortality ~25% and improves retention in cardiology care
— Post-stroke: stroke clinic, rehab, anticoagulation clinic
— Cancer survivorship plans: explicit follow-up schedule, surveillance imaging cadence
— HIV: retention specialists, peer navigators
— CKD: multidisciplinary clinics with nephrology, dietitian, social work
— Heart failure: disease management programs, telehealth weight monitoring
— Motivational interviewing for ambivalent patients
— Shared decision-making improves engagement
— Patient activation measure (PAM) identifies low-activation patients for extra support
Step 3 management: Every hospital discharge should include a scheduled, specific follow-up date and time with the PCP, not "follow up with PCP in 1 week." Vague instructions are a leading driver of LTFU and readmission.
Board pearl: Cardiac rehab is among the most underutilized Class I therapies — only ~30% of eligible patients complete it, largely from LTFU mechanisms (cost, transportation, lack of referral).
— Post-hospital: TCM call within 2 business days, face-to-face within 7d (high complexity) or 14d (moderate)
— Post-ED: 48–72h call for high-risk discharges (chest pain, syncope, mental health)
— Postpartum: contact within 3 weeks, comprehensive visit by 12 weeks; severe-range BP follow-up within 72h
— Post-positive cancer screen: diagnostic study within 60–90 days
— New chronic disease dx (DM, HTN, HF): 2–4 week initial follow-up to assess tolerance and adherence
— Anticoagulation initiation: INR check timing per drug protocol (warfarin within days; DOACs renal/hepatic at 3–6 months)
— % no-shows by clinic/provider/demographic
— Time from abnormal result → next action
— % closed-loop referrals
— 30-day readmission rate
— % patients overdue for chronic disease monitoring (HbA1c, BP, LDL)
— Teach-back: ask patient to repeat plan in their own words
— Explicit "what to do if you can't make it" instructions
— Written summary, ideally in patient's preferred language, 5th–8th grade reading level
— Identify barriers proactively: cost, transportation, childcare, work, language
— Cardiac rehab, pulmonary rehab, stroke rehab — anchor patients in regular contact
— Diabetes self-management education and support (DSMES) — Medicare-covered
— Pulmonary rehab post-COPD exacerbation reduces readmissions
— Defined LTFU threshold per disease (e.g., HIV: no visit in 6 months)
— Outreach algorithm: portal → text → call (x3) → letter → navigator
— Document attempts; offer telehealth as low-barrier alternative
Step 3 management: A diabetic patient hasn't had an HbA1c in 14 months. Next step: outreach via call/portal, schedule visit with same-day lab, reassess social determinants — not simply order the test and assume they'll come.
Board pearl: Teach-back at discharge is associated with ~30% reduction in readmission risk.
— Must disclose risk of LTFU consequences (e.g., need for re-contact, use of vital statistics for outcome ascertainment)
— Right to withdraw at any time without penalty — but investigators may request reason and use of already-collected data
— IRBs increasingly require missing-data plans in protocols (ICH E9 R1)
— Physicians have an affirmative duty to communicate abnormal results and ensure follow-up
— Failure to follow up on a critical result is a leading cause of malpractice claims
— Documentation of outreach attempts (calls, letters, certified mail) is essential defense
— "Patient non-compliance" is not a complete defense if outreach was inadequate
— Competent patients may decline follow-up after informed refusal — document the conversation
— Exceptions: public health threats (active TB, certain STIs) — state law may compel treatment
— Pediatric/elder/dependent adult: caregiver involvement and possible reporting if neglect suspected
— Mental health: involuntary hold criteria (danger to self/others, grave disability) — state-specific
— Joint Commission National Patient Safety Goals: communicate hand-offs, reconcile medications, ensure follow-up
— Hospitalist → PCP gap is the most common site of error
— Discharge against medical advice (AMA): does NOT void physician duty; still arrange follow-up, prescriptions, education
— Disparities in LTFU contribute to outcome disparities (cancer survival, maternal mortality)
— Systems must monitor LTFU by race/ethnicity, language, insurance, geography
— Failure to provide qualified interpreters violates Title VI of Civil Rights Act and ACA Section 1557
— Selective reporting of analyses that minimize impact of LTFU is research misconduct
— Prespecified analysis plans and trial registration (clinicaltrials.gov) reduce this
Step 3 management: A patient with a 3.5 cm pulmonary nodule on CT misses the pulmonology referral and is unreachable. Next step: certified letter documenting the finding, urgency, and recommended action, with attempts logged in the chart, plus continued outreach — both an ethical duty and malpractice mitigation.
Board pearl: Documentation of three good-faith outreach attempts (including at least one in writing) is a widely accepted standard for "did your best."
— RCT attrition >20% = high risk of bias (Cochrane RoB 2)
— Differential LTFU >5% between arms → downgrade evidence quality
— Postpartum visit attendance: ~60% completion nationally
— Cardiac rehab enrollment after MI: ~30%
— HIV care continuum: ~66% retained, ~57% virally suppressed in US
— 30-day readmission after hospitalization: ~15% Medicare average
— TCM face-to-face: 7 days (high complexity, 99496) or 14 days (moderate, 99495)
— Postpartum severe BP follow-up: within 72 hours
— Positive FIT → colonoscopy target: within 60–90 days (delays >9–12 months ↑ CRC risk)
— Complete-case → valid only under MCAR
— Multiple imputation, IPW, MMRM → valid under MAR
— Pattern-mixture, selection models, sensitivity analyses → for MNAR
— LOCF → discouraged; biases vary by disease trajectory
— Rubin's rules → pool MI results, inflate SE appropriately
— Non-differential LTFU → toward null
— Differential LTFU → unpredictable
— Informative censoring (sickest drop out) → overestimates survival
— Healthy-volunteer effect → underestimates risk
— Healthy-adherer effect → exaggerates per-protocol effect
— CONSORT (RCT reporting), STROBE (observational), PRISMA (systematic reviews)
— GRADE for evidence quality
— ICH E9(R1) estimands framework
— Cochrane RoB 2 for trial risk of bias
— Bradford Hill criteria for causation
— HEDIS, MIPS, HRRP, CMS Star Ratings
— Joint Commission NPSGs (communicate critical results)
— AHRQ patient safety indicators
Board pearl: When in doubt on a missing-data MCQ, the answer is usually "intention-to-treat with multiple imputation" as primary, with sensitivity analyses, rather than complete-case or LOCF.
— Stem: RCT with 22% LTFU in treatment, 8% in control; per-protocol analysis shows benefit
— Answer: High risk of bias from differential attrition; ITT analysis with sensitivity analyses required; do not change practice
— Stem: "Patients with worsening depression stopped completing surveys"
— Answer: MNAR — informative missingness; complete-case analysis biased; sensitivity/pattern-mixture needed
— Stem: Longitudinal Alzheimer's trial uses LOCF for missing cognitive scores
— Answer: LOCF biases toward no decline, favoring placebo arm spuriously; MMRM or MI preferred
— Stem: Patient with positive FIT does not return calls
— Answer: Patient navigator + certified letter; engage social work, ensure colonoscopy within 60–90d
— Stem: Discharged HF patient readmitted in 10 days; no PCP follow-up was scheduled
— Answer: Schedule follow-up before discharge, TCM call within 2 business days, F2F within 7–14d, medication reconciliation
— Stem: Postpartum patient with preeclampsia, BP 165/110 at home 5 days post-discharge
— Answer: In-person evaluation within 72 hours of discharge was the missed standard; bring to L&D triage now, initiate or titrate antihypertensives
— Stem: Active pulmonary TB patient missing visits
— Answer: Notify public health department; DOT; possible legal compulsion
— Stem: Clinic's LTFU rate higher among Spanish-speaking patients
— Answer: Qualified interpreters, culturally tailored navigation; required by Title VI/ACA 1557
— Stem: Non-inferiority trial, high LTFU, per-protocol shows non-inferiority but ITT does not
— Answer: Non-inferiority not established; both analyses must support conclusion
— Stem: Worst-case imputation reverses trial conclusion
— Answer: Result not robust; insufficient evidence to change practice
Board pearl: When the stem includes the phrase "as-treated" or "per-protocol only," look for the trap — the primary analysis should usually be ITT.
Loss to follow-up and missing data threaten validity by introducing potentially differential, often informative bias — managed analytically with intention-to-treat plus multiple imputation/MMRM under MAR and sensitivity analyses for MNAR, and clinically with proactive scheduling, closed-loop result communication, patient navigation, and structured transitions of care.
Board pearl: If the question hinges on either "what method handles this missingness" or "what do you do next for this patient who didn't come back" — the safest, most Step 3-aligned answers are ITT + multiple imputation with sensitivity analyses (research) and proactive, documented, navigator-supported outreach with scheduled follow-up (clinical).