Case-Crossover Design

In plain language

A case-crossover study asks: was the patient more likely to be taking a short-acting drug in the days just before their emergency hospitalization than they were in calmer periods months earlier? Instead of comparing the patient to other people, the design compares the same person to themselves — the hazard window (a brief period right before the event) versus earlier referent windows when no event occurred. Because both windows come from the same person, anything that is stable about them — their chronic conditions, genetics, income, general frailty — automatically cancels out. The trade-off is that it only works when the exposure is intermittent (taken sometimes but not always) and the event is sudden and well-dated, like a hospitalization.

The case-crossover design studies whether a transient exposure triggers an abrupt event by comparing, within each case, the exposure status during a short window just before the event (the case/hazard window) with exposure during one or more referent/control windows in that same individual at an earlier time. There is no separate control group: each case is matched to itself. Analytically it is a matched case-control study where the matched sets are person-periods, so the natural estimator is conditional (stratum = person) logistic regression, and the estimand is the odds ratio for the transient effect of exposure on event onset within the hazard window (an approximation of the rate ratio under the usual rare-event/incidence-density logic of Mittleman, Maclure & Robins).

Core conceptual distinction — why self-controlled

Because both windows come from the same person, every time-invariant characteristic is matched out by design: genetics, sex, baseline comorbidity, frailty, socioeconomic status, chronic channeling, and any other stable confounder — measured or not — cancels in the within-person contrast. This is the design's unique selling point relative to cohort or conventional case-control designs, which must measure and adjust for those confounders. The price is that the design controls only time-invariant confounding; anything that varies within a person over the spacing between windows (acute illness, season, day of week, secular trends in exposure prevalence) is not controlled and can bias the estimate badly.

Estimand and assumptions

Under (i) a transient exposure with a well-defined effect (induction/hazard) period, (ii) a stable distribution of exposure over time within persons (the "no exposure-time trend" assumption), (iii) no carryover of effect from the case window into the referent windows (and vice versa), and (iv) the event not itself altering subsequent exposure, the conditional-logistic OR estimates the incidence rate ratio for the acute triggering effect. Violations of (ii) and (iii) are the dominant failure modes and motivate the case-time-control and case-case-time-control extensions below.

Pros, cons, and trade-offs

- vs cohort / new-user designs: Case-crossover needs no comparator cohort and is immune to all stable confounding, which is decisive when the suspected confounders are chronic and hard to measure in claims (frailty, lifestyle). Cost: it answers only the transient-trigger question, gives an OR not an absolute risk, has no information on chronic-exposure effects, and is vulnerable to within-person time trends that a cohort with an external comparator can absorb. Prefer for short-acting drugs and acute events where confounding-by-indication and unmeasured frailty would wreck a cohort. - vs conventional / nested case-control: Removes between-person confounding entirely and needs no control selection, eliminating one source of selection bias. Cost: it cannot study non-time-varying exposures, is less efficient when exposure is rare or near-constant within persons, and shifts the threat from confounding to exposure-trend bias. - vs self-controlled case series (SCCS): Both are self-controlled and remove time-invariant confounding. SCCS models the full observation time with a Poisson/conditional-Poisson model, handles recurrent events naturally, requires the event not to censor/curtail observation (or a modified SCCS), and is well suited to vaccine-safety risk-interval analyses. Case-crossover samples discrete referent windows, needs only the cases, and is simpler when the exposure is sharply intermittent and the event is acute and rare. Prefer SCCS for recurrent events or when modeling age/time effects explicitly; prefer case-crossover for a single acute event triggered by a sharply transient exposure. - vs case-time-control / case-case-time-control: When background exposure prevalence is changing over calendar time (e.g., a drug's use is rising), the plain case-crossover OR is biased by that exposure-time trend. Case-time-control (Suissa 1995) adds a separate control group to estimate and subtract the trend; case-case-time-control (Wang, Schneeweiss) uses future cases as the control series to also absorb within-person trends from disease progression. These are the correct fixes, not heavier covariate adjustment.

When to use

(1) The exposure is intermittent/transient with a plausibly short induction period (an acute drug effect, a single dose, a behavioral trigger, an environmental spike). (2) The outcome is an abrupt, acute, well-dated event (MI, GI bleed, motor-vehicle crash, fracture, anaphylaxis, arrhythmia, seizure). (3) Suspected confounders are largely stable within persons and hard to measure. (4) You want a design that needs only the cases and is robust to unmeasured chronic confounding.

When NOT to use — and when it is actively misleading or dangerous

- Chronic / continuous exposure. If exposure is essentially constant across the case and referent windows (a maintenance statin taken daily for years), there is no within-person contrast and the design is uninformative or biased toward the null. Use a cohort/new-user design instead. - Exposure with a secular time trend. If the drug's prevalence is rising (or falling) over calendar time, exposure is systematically more (or less) likely in the more recent case window than in earlier referent windows for reasons unrelated to the event, producing a spurious OR. This is the canonical case-crossover failure and is exactly what case-time-control repairs. Bidirectional sampling (referent windows both before and after the event) partly mitigates it but assumes the event does not affect future exposure. - Carryover / long or ill-defined effect window. If the effect persists across both windows, or the induction period is long relative to window spacing, exposure "bleeds" into the referent window and attenuates the contrast. Case-crossover demands a short, well-characterized hazard period. - Event alters subsequent exposure (within-person reverse causation). If a prodrome of the event changes exposure (a patient with early symptoms stops or starts a drug), unidirectional (before-only) sampling is biased; bidirectional sampling is invalid because post-event exposure is affected by the event. - Depletion of susceptibles / first-event-only chronic use. For drugs where susceptible patients have their event early in treatment, the within-person exposure-event association is distorted; SCCS or a new-user cohort is safer.

Operational sampling choices (these are design decisions, not nuisances)

Unidirectional sampling draws referent windows only before the event (mandatory when the event can affect future exposure or cause death); it is the more conservative default in pharmacoepidemiology. Bidirectional sampling draws referents both before and after the event and controls for within-person time trends, but is valid only when post-event exposure is unaffected by the event. The number and spacing of referent windows trade efficiency against the exposure-trend and carryover assumptions: more referent windows improve precision (Mittleman, Maclure & Robins show the relative efficiency gains) but widen the calendar span over which the no-trend assumption must hold. A further choice is the point-in-time approach (was the person exposed on the index day of each window?) vs the window/hazard-period approach (was there any exposure within a multi-day hazard period?); in claims the window approach with a `days_supply`-defined exposure interval is usually required because dispensing is bursty (Hallas et al. 2024).

Data-source operational depth

- Claims (FFS): Exposure is reconstructed from pharmacy fills — a person is "exposed" on a given day if a dispensing's `[fill_date, fill_date + days_supply)` interval covers that day (apply a grace period and stockpiling rule, and account for 90-day mail-order and free samples that distort `days_supply`). The event date is the index claim for the acute outcome. Require continuous medical + pharmacy enrollment spanning the entire span from the earliest referent window through the event so that absence of a fill is true non-exposure, not unobserved dispensing. Failure modes: Medicare Advantage and capitated person-time lack fee-for-service claims, so exposure in a referent window can be missing rather than absent — restrict to enrollees with the relevant medical + Part D / pharmacy benefit and exclude MA-only person-time across every window. Outpatient-only acute events may be miscoded or undated; prefer inpatient/ED principal-diagnosis events with a clean admission date as the anchor. - EHR: Exposure is the order/administration, not the fill, and is visit-driven, so the referent windows are populated only when the patient interacts with the system — differential visit intensity around the event (more contact just before an MI) manufactures a spurious within-person exposure trend. Link to dispensing where possible and restrict to exposures routinely captured (e.g., inpatient MAR) rather than self-managed OTC drugs. - Registry / linked: Registries give well-adjudicated, well-dated acute events (the ideal anchor) but rarely complete longitudinal exposure; link to claims for the fill history and to a death index, since unidirectional sampling is mandatory when the event is fatal (no post-event windows exist).

Worked claims example

Question: does short-term NSAID exposure trigger acute upper-GI bleeding among older adults in a Medicare FFS + Part D database? (1) Cases: the first hospitalization with a principal ICD-10 diagnosis of acute upper-GI hemorrhage; the event date is the admission date. (2) Enrollment: require continuous Parts A/B/D and exclude any MA-only person-time over the 6 months preceding the event so every window is FFS-observable. (3) Exposure: a day is NSAID-exposed if any non-aspirin NSAID dispensing's `[fill_date, fill_date + days_supply + 3-day grace)` interval covers it. (4) Hazard (case) window: the 7 days immediately before the event date. (5) Referent (control) windows: four non-overlapping 7-day windows ending 30, 60, 90, and 120 days before the event (unidirectional, because the bleed and any hospitalization can alter subsequent drug use). (6) Classify each window as exposed/unexposed by the rule in (3), creating one stratum per person with one case-window row and four referent-window rows. (7) Fit conditional logistic regression stratified on `person_id`; the exponentiated coefficient is the OR for the transient NSAID-triggering effect. (8) Sensitivity: vary the hazard-window length (3/7/14 days) and grace period; run a case-time-control analysis to net out the secular rise in NSAID use; and test a negative-control exposure (a drug with no plausible acute GI-bleed effect) — its OR should be ~1.

Worked example

Scenario

Two older adults were each hospitalized with an acute upper-gastrointestinal bleed. We want to know whether taking ibuprofen in the days just before the bleed raised their risk. For each patient we define one hazard window (the 7 days immediately before admission) and four referent windows (the 7-day periods ending 30, 60, 90, and 120 days earlier). We then look up whether a pharmacy claim for ibuprofen covered any day in each window, using the rule: a fill is 'on board' for every calendar day from the fill date through the fill date plus the days supplied. If the drug turns up in the hazard window more often than in the referent windows — across many patients like these two — that is evidence of a triggering effect.

Dataset

Pharmacy fill records for two patients. Each row is one prescription dispensed; days_supply tells us how many days that bottle was meant to last. A day is considered 'ibuprofen-exposed' if it falls within [fill_date, fill_date + days_supply).

Steps

• Patient 1001 had a GI bleed admission on 2023-09-15. Their ibuprofen fill on Sep 5 covers Sep 5 through Sep 18 (14 days). The hazard window runs Sep 8–Sep 14 — all seven days fall inside the fill's coverage, so this window is EXPOSED.

• Patient 1001's four referent windows (ending Aug 16, Jul 17, Jun 17, May 17) all fall outside any fill. No ibuprofen was on board during any of them, so all four are UNEXPOSED.

• Patient 1002 had a GI bleed admission on 2023-09-28. Their ibuprofen fill on Aug 18 covers Aug 18 through Aug 24 (7 days). The hazard window runs Sep 21–Sep 27 — well after the fill expired — so this window is UNEXPOSED.

• Patient 1002's referent window ending Aug 29 spans Aug 23–Aug 29. The fill's last covered days are Aug 23 and Aug 24, which fall inside this window, so referent window 1 is EXPOSED. The remaining three referent windows (ending Jul 30, Jun 29, May 30) have no fill coverage and are UNEXPOSED.

• Now count the informative pairs. A 'discordant pair' is any (hazard window, referent window) combination where one is exposed and the other is not — these pairs carry all the information about whether hazard-window exposure is unusually high.

• Patient 1001 contributes 4 discordant pairs: hazard=EXPOSED paired with each of 4 UNEXPOSED referent windows. Patient 1002 contributes 1 discordant pair: hazard=UNEXPOSED paired with referent window 1=EXPOSED. The other three of Patient 1002's referent windows are concordant (both unexposed) and contribute nothing.

• Estimated OR = (pairs where hazard exposed, referent not) ÷ (pairs where referent exposed, hazard not) = 4 ÷ 1 = 4.0.

Result

OR = 4 / 1 = 4.0 — across these two patients the ibuprofen-exposure odds were four times higher in the hazard windows than in the matched referent windows, consistent with NSAID use triggering acute GI bleeding. (In a real study, hundreds of cases would be pooled via conditional logistic regression to get a stable estimate with a confidence interval.)

Timeline Spec

Title

Case-crossover windows for Patient 1001 — NSAID exposure vs. GI bleed event

Window

Start

2023-05-11

End

2023-09-15

Label

Observable person-time spanning all four referent windows through the event date

Events

• Label

  Ibuprofen fill

  Start

  2023-09-05

  Length Days

  14

  Quantity

  14 days_supply

• Label

  GI bleed admission

  Start

  2023-09-15

  Length Days

  1

  Quantity

  event

Spans

• Kind

  unexposed

  Start

  2023-05-11

  End

  2023-05-17

  Label

  Referent 4 (ends 120 d before event): UNEXPOSED

• Kind

  followup

  Start

  2023-05-18

  End

  2023-06-10

  Label

  Between referent windows (not assessed)

• Kind

  unexposed

  Start

  2023-06-11

  End

  2023-06-17

  Label

  Referent 3 (ends 90 d before event): UNEXPOSED

• Kind

  followup

  Start

  2023-06-18

  End

  2023-07-10

  Label

  Between referent windows (not assessed)

• Kind

  unexposed

  Start

  2023-07-11

  End

  2023-07-17

  Label

  Referent 2 (ends 60 d before event): UNEXPOSED

• Kind

  followup

  Start

  2023-07-18

  End

  2023-08-09

  Label

  Between referent windows (not assessed)

• Kind

  unexposed

  Start

  2023-08-10

  End

  2023-08-16

  Label

  Referent 1 (ends 30 d before event): UNEXPOSED

• Kind

  followup

  Start

  2023-08-17

  End

  2023-09-07

  Label

  Between last referent and fill start (not assessed)

• Kind

  exposed

  Start

  2023-09-08

  End

  2023-09-14

  Label

  Hazard window (7 days before event): EXPOSED — ibuprofen fill on board

Result

Label

Hazard window: EXPOSED | Referent windows: all UNEXPOSED → this patient's windows are all discordant in the same direction, contributing to OR > 1

Value

4.0

Caption

Each horizontal band is one 7-day window assessed for ibuprofen exposure in Patient 1001. The four referent windows (blue, UNEXPOSED) represent calm periods months earlier; the hazard window (orange, EXPOSED) sits right before the admission. Because this one patient's hazard window is exposed while all four referent windows are not, the comparison screams 'the drug was unusually present just before the event.' Pooling this pattern across many patients — and dividing by the rare opposite pattern — produces the odds ratio.

Alt Text

Timeline for Patient 1001 running from May 2023 to September 15 2023. Four 7-day bands labeled Referent 4 through Referent 1 are shaded blue (unexposed) at 120, 90, 60, and 30 days before the event. A 14-day ibuprofen fill bar starts September 5. The hazard window (September 8-14) is shaded orange (exposed) immediately before the GI bleed admission marker on September 15.

Runnable example

import numpy as np
import pandas as pd
import statsmodels.api as sm

HAZARD_DAYS = 7          # length of each window
REFERENT_LAGS = [30, 60, 90, 120]  # days before event_date at which each referent window ENDS
GRACE = 3                # days_supply grace to bridge late refills

def _covered(days_set: set, win_end, win_len=HAZARD_DAYS) -> int:
    """1 if any exposed day falls inside the window [win_end - win_len + 1, win_end]."""
    win = {win_end - pd.Timedelta(days=d) for d in range(win_len)}
    return int(bool(days_set & win))

def build_case_crossover(events: pd.DataFrame, rx: pd.DataFrame,
                         enroll: pd.DataFrame) -> pd.DataFrame:
    # Earliest day any window can reach back to; used for the enrollment-observability check.
    earliest_lag = max(REFERENT_LAGS) + HAZARD_DAYS
    ev = events.merge(enroll, on="person_id")
    ev["covers"] = ((ev["enroll_start"] <= ev["event_date"] - pd.Timedelta(days=earliest_lag)) &
                    (ev["enroll_end"]   >= ev["event_date"]) &
                    (~ev["ma_only"]))                 # every window must be FFS-observable
    ok = ev.loc[ev["covers"], "person_id"].unique()
    ev = events[events["person_id"].isin(ok)].copy()

    # Per person, the set of calendar days covered by the exposure of interest.
    rx = rx[rx["person_id"].isin(ok)].copy()
    rx["start"] = rx["fill_date"]
    rx["end"]   = rx["fill_date"] + pd.to_timedelta(rx["days_supply"] + GRACE, unit="D")
    exposed_days = {}
    for pid, g in rx.groupby("person_id"):
        days = set()
        for _, r in g.iterrows():
            days |= {r["start"] + pd.Timedelta(days=d)
                     for d in range((r["end"] - r["start"]).days)}
        exposed_days[pid] = days

    rows = []
    for _, r in ev.iterrows():
        pid, e = r["person_id"], r["event_date"]
        days = exposed_days.get(pid, set())
        # Case window: the HAZARD_DAYS immediately before (and including) the event date.
        rows.append({"person_id": pid, "is_case_window": 1,
                     "exposed": _covered(days, e)})
        # Referent windows: each ends `lag` days before the event date (unidirectional).
        for lag in REFERENT_LAGS:
            rows.append({"person_id": pid, "is_case_window": 0,
                         "exposed": _covered(days, e - pd.Timedelta(days=lag))})
    return pd.DataFrame(rows)

def fit_case_crossover(df: pd.DataFrame):
    # Conditional logistic = stratified-by-person logit of the case-window indicator on exposure.
    model = sm.ConditionalLogit(endog=df["is_case_window"],
                                exog=df[["exposed"]],
                                groups=df["person_id"])
    res = model.fit(disp=False)
    or_ = np.exp(res.params["exposed"])
    ci = np.exp(res.conf_int().loc["exposed"])
    return res, or_, (ci[0], ci[1])

library(data.table)
library(survival)

HAZARD_DAYS  <- 7L
REFERENT_LAGS <- c(30L, 60L, 90L, 120L)  # days before event_date at which each referent window ENDS
GRACE        <- 3L

build_case_crossover <- function(events, rx, enroll) {
  setDT(events); setDT(rx); setDT(enroll)
  earliest_lag <- max(REFERENT_LAGS) + HAZARD_DAYS

  # Keep only people FFS-observable across every window (no MA-only person-time).
  ev <- merge(events, enroll, by = "person_id")
  ok <- ev[enroll_start <= event_date - earliest_lag &
           enroll_end   >= event_date & !ma_only, unique(person_id)]
  ev <- events[person_id %in% ok]
  rx <- rx[person_id %in% ok]

  # Per-person set of exposed calendar days from [fill_date, fill_date + days_supply + grace).
  exposed_days <- rx[, .(day = unlist(Map(function(s, n) s + seq_len(n) - 1L,
                                          fill_date, days_supply + GRACE))),
                     by = person_id]

  covered <- function(pid, win_end) {
    win <- win_end - (0:(HAZARD_DAYS - 1L))
    as.integer(nrow(exposed_days[person_id == pid & day %in% win]) > 0L)
  }

  out <- rbindlist(lapply(seq_len(nrow(ev)), function(i) {
    pid <- ev$person_id[i]; e <- ev$event_date[i]
    case <- data.table(person_id = pid, is_case_window = 1L, exposed = covered(pid, e))
    refs <- rbindlist(lapply(REFERENT_LAGS, function(lag)
      data.table(person_id = pid, is_case_window = 0L, exposed = covered(pid, e - lag))))
    rbind(case, refs)
  }))
  out[]
}

fit_case_crossover <- function(df) {
  # Conditional logistic regression stratified on person; exp(coef) is the OR.
  clogit(is_case_window ~ exposed + strata(person_id), data = df)
}