ANCOVA and logistic regression of responder rates

Download Report

Transcript ANCOVA and logistic regression of responder rates

Presentation title
Date
Responder endpoint and
continuous endpoint,
logistic regression or
ANOVA?
DSBS 24 OCT 2013
Søren Andersen
Presentation title
Date
Slide no 2
Example and problem
• HbA1c is analysed with an ANCOVA model and in
addition the ”responder rate” (HbA1c < 7%) is analysed
by a logistic regression model
• Well documented that dichotomising reduces sensitivity
• Results presented as difference in HbA1c and as odds
ratio
• Difficult to compare the results
• Difficult to interpret odds-ratio for probalities p (from
logistic regression model) in [0.2; 0.8], no
interpretation as relative risk
• Example: Old study with Liraglutide
Presentation title
Date
Slide no 3
Presentation title
Date
Slide no 4
Outline
• Comparisons on probability scale
• Show no difference between logit and probit in estimated
responder probabilities (and in treatment differences in
responder probabilities)
• Compare responder probabilites derived from ANCOVA
with responder probabilities from logit and probit
• Comparisons on continuous scale
• Compare estimates from logit and probit to estimates
from ANCOVA
Presentation title
Date
Slide no 5
Presentation title
Date
Slide no 6
Suggestion: use probit instead of logit
•
•
•
A probit model for binary data is very similar to a logit
model. Very difficult to discriminate between the two.
Pro logit:
• a logit model is very useful for retrospective studies
(not the case here)
• a logit model is convenient for calculation of
conditional probabilities
• a logit model offers interpretation in terms of oddsratio
• Technical point: simple sufficient statistics
Pro probit:
• offers interpretation in terms of a latent normal
variable (threshold model)
Presentation title
Date
Slide no 7
Comparions of logit and probit estimates
of probabilities
• Logit and probit model with effects of
•
•
•
•
Country (17)
Pre-treatment (2)
Treatment (3)
Base line HbA1c
• responder probabilities were estimated for all countries
(17) and pre-treatment (2), treatments (3) and 3
values of base line HbA1c (mean +- std)
• In all 17 x 2 x 3 x 3 = 306 probabilities
Presentation title
Date
Slide no 8
Estimated p’s of 3 treatments across subgroups
Presentation title
Date
Slide no 9
Presentation of results from probit and
logit models
• Present differences in estimated proportions between
two treatment groups, Lira and Comparator – not
constant
• Depend on proportion in the Lira (or Comparator)
Presentation title
Date
Slide no 11
Presentation title
Comparing logit and probit treatment
differences
Date
Slide no 12
Presentation title
Date
Estimated p’s of 3 treatments across
subgroups ANCOVA and probit
Slide no 14
Presentation title
Date
Slide no 15
Presentation title
Date
Slide no 16
Presentation title
Date
Slide no 17
Presentation title
Comparison on “latent scale” of
parameter estimates
Date
Slide no 18
Presentation title
Date
Slide no 19
Comparison of estimates of treatment
difference
• From ANCOVA :
0.2367 (residual s = 0.81)
• From probit:
0.3379 (”residual s = 1”)
•
0.3379*0.81 = 0.2758
• From logit:
0.5440
convert to probit: 0.5440*0.607 = 0.3302
convert to ANCOVA: 0.3302*0.81 = 0.2695
To obtain the same precision of estimate from probit and
logit as for ANCOVA twice as many observations are
needed
Presentation title
Date
Slide no 21
Conclusions
• Dichotomising reduces sensitivity (in the example
sample size doubles)
• Communicate results from logit/probit as difference in
proportions if OR markedly different from RR
• Compare results from ANCOVA and logit/probit on
probability scale and on ”latent scale”
Presentation title
Date
Slide no 22
Composite responder endpoint?
• Responder: (HbA1c < 7) & (change in weight < 0), i.e.
two binary response B1 and B2 combined
• Why composite? Why collapse 3 categories of the B1 x
B2 outcome?
• For quantitative responses we test for each parameter:
H0: no difference in HbA1c, H0: no difference in
chg_bw
• Analyse B1 and B2 separately or
• Analyse the full response pattern B1 x B2, as marginal
B1, B2 conditional on B1 (or other way round)