Full code
No documented CPR limitation in this recorded state. Missing documentation is not assumed to mean full code.
Known documented comparator
A small order.
A very big inference.
When clinicians record do not resuscitate in intensive care, what changes next? And how much can the medical record actually tell us about why?
MIMIC-IV v3.1 · 3,021 eligible first DNR orders · Observational research · Not clinical advice
01 — DEFINE THE EXPOSURE
First, name the decision.
The words on the order matter. DNR applies to CPR at cardiac arrest. It does not, by itself, direct staff to withdraw ventilation, stop medications, or provide comfort-only care.
DNR only
Do not attempt CPR at arrest. Other decisions about treatment need separate evidence.
CPR limitation
DNR + DNI
CPR and intubation limitations recorded together. It must not be pooled with isolated DNR.
CPR + intubation limitations
Comfort-measures-only is a separate treatment-goals state, not a synonym for DNR. These are recorded directives, not a complete picture of the patient’s wishes.
02 — WHERE THE COMPARISON BEGINS TO BREAK
Watch the study shrink.
Each circle represents an aggregated ICU group, never an individual. Its position shows mean recorded illness severity and the fraction of ICU time spent on invasive ventilation just before the order. Its area represents the number of orders.
ANALYTIC STAGE — NOT PATIENT TIME
The circles move because we change the eligibility rules, not because individual patients travel along these axes. Stricter overlap can make a subgroup disappear.
Read the plotted aggregate table
| ICU group | Documented order | Count | SOFA mean | Ventilation before order |
|---|---|---|---|---|
| MICU | DNR + DNI | 1,017 | 7.42 | 57.0% |
| MICU | DNR only | 737 | 7.74 | 48.2% |
| Other ICU units | DNR + DNI | 680 | 6.36 | 56.6% |
| Other ICU units | DNR only | 587 | 6.52 | 38.0% |
03 — TIME ZERO IS NOT RANDOM
The order is written. What moves?
Two aggregate clinical histories are aligned to the order—or an equivalent ICU elapsed time for eligible full-code controls. Explore four-hour observed bins. Rates can change before the order, because clinicians respond to evolving illness.
DOCUMENTED ORDER
792 orders on measured common support · CPR limitation documented without simultaneous DNI. These are selected decisions, not randomized exposures.
RECORDED FOUR-HOUR WINDOW0 to +4 hours
−24h before0 · ORDER+12h after
DOCUMENTED DNR ONLY29.2%
ELIGIBLE FULL-CODE CONTROLS35.4%
During this four-hour bin, the recorded rates differ by −6.16 percentage points. This is not a causal effect. Recorded rates after the order are also shaped by ICU exit and death.
Full 12-hour difference-in-changes: -7.78 percentage points; conditional 95% interval -10.44 to -5.12. These intervals do not apply to individual time bins.
After the order · the patient population still observed in ICU changes as people die or leave. Event time is not randomized.
These are clinical recording and treatment rates among patients still observed in ICU—not causal responses to the order. The individual four-hour bins have no published pointwise confidence intervals.
Inspect every recorded time bin
| Four-hour window | DNR only | Full-code controls |
|---|---|---|
| -24 to -20 hours | 36.3% | 36.7% |
| -20 to -16 hours | 37.0% | 37.0% |
| -16 to -12 hours | 37.0% | 37.1% |
| -12 to -8 hours | 36.2% | 36.7% |
| -8 to -4 hours | 36.4% | 36.0% |
| -4 to 0 hours | 34.5% | 34.5% |
| 0 to +4 hours | 29.2% | 35.4% |
| +4 to +8 hours | 29.2% | 37.6% |
| +8 to +12 hours | 28.9% | 37.9% |
04 — THE CLAIM IS THE DIFFICULT PART
Now try to break the finding.
Pick an estimate. Compare earlier negative-control timing, the window around the order, the primary analysis and stricter control overlap. The figures change because the analysis design changes.
RECORDED OUTCOME
Earlier pre-orderNegative-control timing
At the order−4 to +4 hours
Main−12 to +12 hours
Strict overlapControl support restricted
Invasive ventilationpercentage-point change
Vasopressor exposurepercentage-point change
MAP recordingrecords per patient-hour
SELECTED OBSERVATIONAL CONTRAST · DNR only
Vasopressor exposure · strict overlap
-0.45 pp (conditional interval -2.31 to 1.41). The interval includes no difference.
05 — THE CAUSAL QUESTION
Did the order cause the deaths?
The mortality difference is striking. It cannot answer the causal question. Follow four stages of inference and see why the conclusion stops short of attributing deaths to a CPR limitation.
OBSERVED 12-HOUR MORTALITY
Different mortality. Different underlying prognosis?
Mortality after a new DNR only order is much higher than in weighted eligible full-code risk sets. These people were not randomized to code status.
0%10%20%30%
Observed 12-hour risk difference: 17.62 percentage points (fixed-weight approximate 95% interval 14.83–20.40). This is an association, not an estimate of DNR-caused mortality. This figure uses a fixed 0–30% zero-based axis for both subtypes.
01
Recorded differences are real.
ObservedThe two observed groups have very different rates of death and treatment changes around the index. That is a description of the data.
02
Measured controls do not make patients exchangeable.
Balance failed414 of 792 case-specific risk sets have fewer than two effective controls. The stricter comparison leaves 378 cases and reaches measured balance |SMD| = 0.318, above the 0.10 diagnostic.
03
Provider practice variation is not automatically an instrument.
IV not validatedOnly 48 of 62,078 ICU-time code-status orders were entered by the recorded hospital admitting provider. Neither identifier proves quasi-random assignment of the ICU decision-maker.
04
We cannot attribute deaths to the DNR order.
Not identifiedUnknown neurological prognosis, treatment intentions, patient preferences and simultaneous clinical decisions may explain large parts of the contrast. A DNR-only directive concerns CPR at arrest, not blanket withdrawal of life support.
THE IDENTIFICATION BOUNDARYWe observe decisions.
We observe decisions.
We do not observe every reason.
Non-CPR treatment and physiological monitoring change around documented DNR decisions.
Measured case-mix similarity cannot remove unrecorded prognosis, patient wishes or clinical intentions.
The records do not establish how many deaths, if any, were caused by the isolated DNR order.
DATA: MIMIC-IV v3.1 · AGGREGATED STATISTICS ONLY
Data extracted from the existing secure ICU cohort. Aggregates below groups of 30 are suppressed from the cohort map. Conditional intervals do not include all model-selection or weight-estimation uncertainty.
Read the full DNR research release ↗ Research design and limitations ↗