The Readmission Rate That Changed Definitions Overnight
At the quarterly quality committee of a 180-bed hospital, the analyst presents a slide showing 30-day all-cause readmissions rising from 8.2% to 11.6% quarter over quarter. The board chair asks for a rapid improvement team and a report at the next meeting. You learn the prior quarter's figure was built from the Medicare FFS claims file only, because that was the only file available at the time. The current figure was built from the internal EHR using all payers, excluded observation index stays, and did not apply the planned-readmission algorithm. The all-payer result is 214/1,842. The Medicare FFS subset is 138/1,640 (8.4%). One 11-bed medical unit accounts for 22 of the 214 readmissions; last year the same unit produced 9.
1. Before the board presentation, what validation steps do you take, and in what order?
Read the worked answer
Validate the definition before anyone validates the data.
- Reconstruct the measure specification for both quarters. Payer population, index-stay inclusion (observation excluded? transferred? expired? discharged AMA?), planned-readmission handling, and the 30-day window anchor. The two figures differ on at least four of these.
- Rebuild the current quarter in the old specification using the Medicare FFS subset: 138/1,640 = 8.4%. Compare to 8.2%. The apparent 3.4-point increase disappears.
- Rebuild the prior quarter in the new specification if the EHR data exists, so both quarters sit on one definition.
- Only then mine the current all-payer data for real variation - by service line, discharge disposition, and unit.
- Check the underlying data: duplicate encounters, readmissions to other facilities not captured internally (a known all-payer undercount), and whether planned returns such as scheduled chemotherapy were counted.
The headline finding to report is that the number changed because the definition changed, not necessarily because performance changed. That is a data integrity finding, and it belongs on the slide.
2. How should the finding be presented to the board, and what do you recommend?
Read the worked answer
Present it as two clearly labeled lines, not one trend line. A merged line implies comparability that does not exist.
- "Medicare FFS 30-day all-cause readmission (consistent definition, both quarters): 8.2% → 8.4%. Movement within normal variation."
- "All-payer readmission (new measure, first measurement period): 11.6%. Baseline established this quarter; no prior-period comparison is valid."
- State explicitly that the 11.6% is a baseline, not a deterioration, and that the all-payer figure also likely understates true readmissions because it misses out-of-network returns.
Recommendation: do not charter a rapid improvement team against a definitional artifact. Charter two things instead: (a) a measure-steward assignment with a locked written specification and a version-controlled data dictionary so this cannot happen again, and (b) a targeted review of the 11-bed unit's 22 readmissions, which survives the definition change and is small enough to case-review individually.
Give the board a date for the validated baseline rather than a number you may retract.
3. What statistical caution applies to the 11-bed unit finding, and how should the unit be reviewed?
Read the worked answer
Small denominators make rates unstable and rankings misleading. Twenty-two readmissions on a unit that discharges perhaps 300 patients a year is a rate that will swing wildly with one or two patients, and last year's 9 versus this year's 22 is a difference of 13 events - well within what chance alone can produce at that volume.
Correct handling:
- Plot the unit on a p-chart with control limits rather than comparing a point estimate to last year's point estimate. If 22 sits inside the limits, there is no special-cause signal and a rapid improvement team would be chasing noise.
- Never rank units by readmission rate without volume adjustment; the smallest unit will always occupy one extreme.
- Adjust or stratify for case mix before attributing performance to the unit; an 11-bed unit may simply carry more complex discharges.
- Because the count is small, switch methods: review all 22 charts individually instead of computing a rate. Twenty-two charts can reveal preventable causes (no follow-up appointment, discharge medication not filled, no post-discharge call) without any statistical inference at all.
The judgment point: a small number is a reason to look at the individual cases, not a reason to run a control chart and declare victory or failure.