Title: Performance Reports
1Performance Reports
- Andy Bindman MD
- Department of Medicine, Epidemiology and
Biostatistics - UCSF
2Calculating an Individuals Risk
-
- Solve the multivariate model incorporating an
individuals specific characteristics - For continuous outcomes the predicted values are
the expected values - For dichotomous outcomes the sum of the derived
predictor variables produces a logit which can
be algebraically converted to a probability (enat
log odds /1 enat log odds )
3Aggregating to the group level
- Sum observed events (eg deaths) for sub-group
- Sum expected probability of events for same
- Probabilities used to calculate expected events
derived from entire data set and applied to
individuals in sub-group (eg defined by care
site) - The overall expected number of events must equal
the observed number of events but this need not
be the case at the level of subgroups
4Comparing observed and expected outcomes to
assess quality
- Observed events or rates of events
- Expected events or rates of events
- Better quality implied when observed is lower
than expected (worse quality when observed higher
than expected)
5Risk adjusted rates
- Standardizes rates across sub-groups so that they
can be directly compared with a single number - Observed rate/expected rate of subgroup
- x overall observed rate
6Observed CABG Mortality Rates, NY 1989-1992
7Calculating Expected CABG Mortality Rates in New
York by Year
- Pool all 4 years of CABG patients
- Develop risk adjustment model for CABG patients
- Apply risk adjustment model for CABG patients
sub-grouped by year to determine expected number
of deaths for each year - Divide expected number of deaths by number of
cases per year to get expected death rate
8Observed and Expected CABG Mortality Rates, NY
1989-1992
9Annual Risk Adjusted Mortality Rate for CABG in
New York
- Observed rate per year/expected rate per year
-
- X average death rate over 4 year period (3.1)
10Observed, Expected and Risk-Adjusted CABG
Mortality Rates, NY 1989-1992
11What Happened with CABG Surgery Over Time in New
York?
- Operated on sicker patients
- Observed mortality rate declined over time
- Risk adjusted mortality rate declined even more
- Did quality of CABG care improve over time?
12NY CABG Risk Adjustment Model
- Well designed model
- C index .787 Hosmer-Lemeshow chi square p.16
- Mortality is not a subjective outcome- hard to
fake - Gaming might be possible with coding some
predictors
13Interpreting Risk Adjusted CABG Outcomes
- Public reporting on hospital CABG mortality began
in 1989 - Low volume hospitals had higher mortality rates
and some stopped performing CABG over time - Process indicators of cardiac care (beta blocker
post MI) also improved in NY hospitals over time - Hospitals documented more co-morbidities over
time resulting in inflated expected death rates - Some sick NY cardiac patients operated on in NJ
14Applying Results To Providers
- Possible to aggregate observed and expected rates
of events to hospital, physician, or some other
provider level grouping - Statistical problems arise when total number of
expected events are small - Minimum of five expected events per group as a
rule of thumb
15Naming Names
- Assigning assessments of quality to specific
providers increases the stakes - Need to demonstrate validity of analytic approach
16Reporting Results
- Public reporting vs internal quality improvement
- Data users tend to want gradations of quality
along a continuum (excellent to poor) - However denoting those within a 95 confidence
interval of the expected as average is less
sensitive to noise in data
17Bootstrap ProcedureDeriving Confidence Intervals
- Multiple (e.g. 1000) random samples of same size
of original derived from original sample with
replacement - Calculate expected rate for each new sample
- Create frequency distribution of expected rates
- Empirically derive 95 CI (950 of 1000 centered
around mean)
18(No Transcript)
19Consistency in the evidence
- Differences between observed and expected may be
due to things other than quality - Are the results consistent over time
- Are results consistent with prior expectations
such as volume-outcome relationships - Confirmation through very different types of
evidence is a major goal- external validation
20Observed / Expected Mortality
21Volume-Outcome
- Relationship between high volume providers and
better outcomes - Most often studied in relationship to procedures
- Consistent with notion that practice makes perfect
22Hospital Volume and CABG Mortality in California
Hospitals Using Registry, 2000-02
23External validation of data
- Link hospital discharge data with CABG registry
data - Looking for missing cases, deaths and highly
predictive risk factors - 221 cases in discharge data not reported to
registry - 26 additional deaths (498 total)
- 63 undercodes and 123 overcodes of cardiogenic
shock - 29 overcodes of salvage (51 total)
- Direct auditing
- Deaths
- Highly weighted predictors particularly if
subjective
24Growing Number of Quality Initiatives Provide
Opportunity for Cross Comparisons
- AHRQ Quality Indicators
- JCAHO ORYX Hospital Core Performance
- CMS Hospital Quality Alliance
- National Quality Forum (NQF)
- Leapfrog
- NCQA HEDIS
25JCAHO - Hospital AMI indicators
- Process elements
- Aspirin at arrival, Aspirin at discharge, ACEI
for LVSD, Smoking cessation counseling,
Beta-blocker at discharge, Beta-blocker at
arrival, Thrombolytic within 30 minutes, PCI
within 120 minutes -
-
26Who uses these reports and how
- Patients
- Slow to catch on
- More important for those without other ways to
judge quality - Managers
- Aim to improve quality to avoid naming and
shaming - Payers (health plans)
- Selective contracting
- Pay for performance
27How much do these reports matter?
- California has lowered isolated CABG mortality by
1 (from 3 to 2 ) during public report period - Approximately 20,000 procedures per year
- Reduction from 600 to 400 deaths
- Average survival 5 years
- Even if half the change is due to gaming, 500
life years saved
28My Reflections on Performance Reports
- View the risk adjusted estimates as yellow
flags, not smoking guns - Risk models will probably improve with ability to
add more clinical data available through
electronic records - Research may not need to be perfect to bring
about public health benefits - Attempts to improve quality need to consider
unintended consequences on access/efficiency
29Outcomes Research Opportunities
- Validate risk adjustment models for new
conditions - Are health care outcomes changing over time and
if so why? - How can performance reports on health outcomes be
used to create better health care quality?