Transcript document

Focusing on the variation at the higher
level: Methods in the Glance paper
1. Standard regression model
•
•
•
•
Ignore MD and run logistic regression with all
patients
Use coefficients of patient level variables and
each patient’s values to obtain patient’s
predicted probability of death
For each MD, take average probability of
death for his/her patients for expected rate
Calculate observed vs. expected mortality
ratio; standardize to state average
Methods in the Glance paper
2. Fixed effects regression model with an
indicator variables for each MD (less one)
•
•
•
Can’t just exponentiate coefficients from
indicator variables for each MD
Calculate predicted MD mortality rate from
patients predicted probability adding term for
providers coefficients weighted by provider’s
proportion of patients
Form observed versus expected ratio as with
standard model
Methods in the Glance paper
3. Random effects regression model: Two
methods from RE model
•
Used fixed effects patient coefficients as in
standard model to obtain each MD’s predicted
mortality rate and form observed/expected ratio
•
Use “shrinkage estimates” for each MD from
the RE model; exponentiate coefficient for each
MD to get individual odds ratios relative to the
average MD
Methods in the Glance paper
Ranking outliers from the statistical models:
1. Calculate upper and lower 95% CI of O/E for
each MD using standard error of each MD’s
mean assuming a Poisson distribution
2. Obtain from CI of O/E by bootstrapping,
stratified by MD
•
Only used with standard regression model
3. Exponentiate 1.96 x SE of each MD’s
shrinkage estimate to get 95% CI of OR
Methods in the Glance paper
Three methods for comparing agreement
among the models:
1. Kappa statistic
2. Calculate ICC using ANOVA
3. Compare the size of the confidence intervals
Effect of shrinkage versus fixed
effects estimates in Glance et al.
• 16 of 127 MDs outliers in all models
• 4 more outliers with FE but only 1 with RE
– 3/20 (15%) classified differently by shrinkage
• 2 of 29 hospitals outliers in all models
• 3 more outliers with FE but 0 with RE
– 3/5 (60%) classified differently by shrinkage
Conclusions in Glance
• “Shrinkage estimators...more precise...at
the cost of introducing bias...may also
have substantial limitations”
• “Fixed effects...will result in unbiased
estimates...when provider effects are
correlated with...patient risk factors”
• “Bootstrap...was...less conservative in
classifying providers as quality outliers”
Characteristics of Fixed (FE)
versus Random Effects (RE)
• FE models control for or remove the
association between the higher level variable
and the outcome
– Indicator variable gives a separate intercept for
each unit
– Conditional regression “conditions out” the effect
of the higher level variable
– In both cases, cannot look at association with
variables of higher level characteristics
Characteristics of Fixed (FE)
versus Random Effects (RE)
• RE models can have variables for higher
level characteristics
• RE model the variance distribution of the
higher level variable independently of
lower level variables:
– Model may be biased if a lower level variable
is associated with the random effect
Characteristics of Fixed (FE)
versus Random Effects (RE)
• FE models control unmeasured
confounding of higher level units
• RE more precise than FE because uses all
variance (not just within cluster variance)
• Common trade-off in statistics: more
precision more potential for bias
Example: FE vs. RE when lower level
variable associated with higher level
• Calif. CABG data set modified by adding a
patient variable “x” whose true coef. = 1.0
• Variable X is strongly associated with
hospital
• FE indicator variable model compared to
GEE and RE models
Estimate of X (true value=1.0): 3 Models
1) Fixed effects regression
xi: logit y x i.hospid
y
x
Coef.
1.03
Std. Err.
.0848
2) GEE regression
xtlogit y x, pa i(hospid)
y
x
Coef.
0.93
Std. Err.
.0778
3) Random effects regression
xtlogit y x, re i(hospid)
y
x
Coef.
1.20
Std. Err.
.0860
Estimate of X (true value=1.0), adjusting
GEE and RE models for prevalence of X
among the hospitals
4) GEE regression adj for prx
xtlogit y x prx, pa i(hospid)
y
x
Coef.
1.02
Std. Err.
.0854
5) Random effects regression adj for prx
xtlogit y x prx, re i(hospid)
y
x
Coef.
1.02
Std. Err.
.0844
Choosing a Model for Multilevel Data
1. Regression model ignoring higher level
variables
2. Regression model with an indicator
variable for each level 2 unit (minus one)
3. Conditional regression model
4. Regression model with generalized
estimating equations (GEE model)
5. Random or mixed effects regression
model
Choice of Analysis Model:
Three Main Considerations
• What is the research question?
• How many observations are there at each
level of the data?
• How important is controlling unmeasured
confounding at the higher level?
Research Question Targets
• Association of lower level measurements
with outcome
• Association of higher level variable
measurements with outcome
• Variation at higher level
• Whether coefficient of a measurement
varies signficantly at either level
What Model Would You Use?
• Clinical trial with 10 centers, 400 subjects,
characteristics of centers and subjects
• Case-control study, 200 CABG surgeons,
one patient who died and one who didn’t
from each MD, characteristics of patients
What Model Would You Use?
• Cross-sectional study of patient satisfaction
with MD, 70 MD’s, 1500 patients,
characteristics of MD’s and patients
• Cross-sectional study patient bp, 400 MD’s
and 400 patients (1 per MD),
characteristics of MD’s and patients
What Model Would You Use?
• Cross-sectional study of patient mortality
100 hospitals, 20,000 patients, patient
characteristics, RQ is whether patient age
predicts mortality differently among hospitals
• Cross-sectional study of patient days in
hospital, 50 hospitals, 4000 patients, patient
characteristics, suspect significant
unmeasured confounding by hospital