Title: Compliance and Causal Analysis Lecture 2 Randomisationbased methods 1: likelihood estimation
1Compliance and Causal AnalysisLecture 2
Randomisation-based methods 1 likelihood
estimation
- Ian White
- MRC Biostatistics Unit, Cambridge, UK
- ian.white_at_mrc-bsu.cam.ac.uk
2Causal analysis in randomised trials
- A randomised controlled trial is used to evaluate
an intervention - No confounding because randomised group R is
independent of all predictors of outcome - But we get departures from randomised
intervention actual treatment D differs from
allocated treatment R - Want to infer causal effect of treatment D
3Departures from randomised intervention
- Sometimes called non-compliance
- I prefer departures
- avoids implicit value judgements
- more precise includes both
- non-adherence (randomised to X, clinician
prescribes X, patient does Y) - changes in prescribed treatment (randomised to
X, clinician prescribes Y, patient does Y)
4Types of departure from randomised intervention
- Switches to other trial treatment
- Changes to non-trial treatment
- Includes changes to nothingin a comparative
trial - Departures may be
- all-or-nothing (either always get A or always get
B) - or quantitative (e.g. dose changes)
- or time-dependent (e.g. emergency operation)
? FOCUS ON SWITCHES
5Motivating example MASS trial
- Abdominal aortic aneurysms are often fatal if
they rupture - May be repaired if detected before rupture
- Reliably detected by ultrasound screening
- MASS trial (Lancet, 2002) 67,800 men were
randomised to invitation to screening or control - Main outcome was aneurysm-related mortality
6MASS trial Aneurysm-related mortality
7Intention-to-treat analysis
- Intention-to-treat analysis compares groups as
randomised, ignoring any departures - respects the randomisation
- avoids selection bias
- Now the standard analysis and rightly so
- Answers an important pragmatic question e.g. what
is the public health impact of prescribing X? - Disadvantage this may be the wrong question!
8Disadvantage of ITT
- Doctor doctor, how much will taking this tablet
reduce my risk of heart disease? - I dont know, but prescribing it reduces
disease risk by 10 - on average
- thats on average over whether you take it or not
9Example MASS trial (ctd)
- Intention-to-treat (ITT) analysis invitation
to screening reduced aneurysm-related death - hazard ratio 0.58 (95 CI, 0.42 to 0.78), P0.002
- 20 of invited group didnt attend for screening
(all-or-nothing non-compliance) - ITT measures the average benefit of screening in
invitees - What is the benefit of screening in attenders?
10Plan for this lecture
- Basic idea for all-or-nothing compliance
- Binary outcome estimation
- Normal outcome
- Simple vs. ML estimation
- Negative weights method
- Back door method
111. Basic idea for all-or-nothing compliance
12Notation
- Randomise to E (experimental) or S (standard)
treatment - S could be nothing / placebo
- Potential treatments are E and S assume everyone
gets either all E or all S (all-or-nothing
compliance) - Ri 1/0 indicates randomisation of ith
individual to E/S - Di 1/0 indicates receipt of E/S
13Theoretical framework
- Ri 1/0 indicates randomisation of ith
individual to E/S - Di 1/0 indicates actual receipt of E/S
- Di(0) 1/0 indicates receipt of E/S if
randomised to S observed in S arm,
counterfactual in E arm - Di(1) 1/0 indicates receipt of E/S if
randomised to E observed in E arm,
counterfactual in S arm - So actual Di Di(Ri)
- The pair Di(0), Di(1) define an individuals
compliance type.
14Compliance types
- Always-takers (A) Di(0) Di(1) 1 always take
E regardless of randomisation - Compliers (C) Di(0) 0, Di(1) 1 take whatever
they were allocated to - Never-takers (N) Di(0) Di(1) 0 never take E
regardless of randomisation - Defiers (D) Di(0) 1, Di(1) 0 take the
opposite of what they were allocated to
15Using compliance types
- Note that compliance-type is a characteristic
that is inherent to the individual before
randomisation - unlike actual compliance, which is affected by
randomisation - This means we can meaningfully adjust/stratify by
compliance-type, but not by actual compliance - But unfortunately compliance-type is incompletely
observed requires careful statistical methods - Compliance types are an example of Principal
strata (Frangakis Rubin, 2002)
16What do we observe about compliance-types?
17What do we observe about compliance-types?
Simplification if S arm cant get E (no
always-takers and no defiers)
18Defining true treatment effect
- Could be defined in various ways
- Use potential outcomes Y(r,d) outcome if
randomised to r and received d - Y(r,d) may depend on r
- e.g. never-takers of a counselling intervention
might do worse in the E arm (where they refused
the counselling) than in the S arm (where they
werent offered the counselling) - but in many settings it wont (exclusion
restriction)
19Defining true treatment effect
- Average effect of treatment
- EY(1,1)-Y(0,0)
- Average effect of treatment among the treated
- EY(1,1)-Y(0,0) D(1)1
- For an always-taker, we will observe
- Y(0,1) if we randomise to S
- Y(1,1) if we randomise to E
- but we cannot observe Y(0,0)
- Is there a measure of causal effect that involves
potentially observable outcomes?
20Complier average causal effect (CACE)
- ITT effect EY(1,D(1)) Y(0,D(0))
- Define average causal effect of treatment
assignment in compliance-type t as EY(1,D(1))
Y(0,D(0)) Tt - Complier Average Causal Effect (CACE)
- EY(1,1) Y(0,0) TC
- also Local Average Treatment Effect (LATE)
(Angrist et al, 1996) - Never-taker average causal effect (NACE?)
- EY(1,0) Y(0,0) TN
- Always-taker average causal effect (AACE?)
- EY(1,1) Y(0,1) TA
- CACE measures treatment efficacy but only
involves potentially observable outcomes (Imbens
Rubin, 1997)
21CACE (2)
- Ignore defiers (mainly for simplicity)
- Can write ITT wC CACE wN NACE wA AACE
- where wC P(compliance-type C) etc.
- Its often reasonable to assume NACEAACE0, so
that ITT wC CACE - Give a simple estimate of the CACE
22CACE (3)
- For binary outcomes, we can define the CACE on
different scales. - Risk difference scale
- CACE EY(1,1) TC EY(0,0) TC
- Risk ratio scale
- CACE EY(1,1) TC / EY(0,0) TC
- Odds ratio scale
- CACE OY(1,1) TC / OY(0,0) TC
- where OXEX/(1-EX)
232. Estimation
24Vitamin A trial(Sommer and Zeger, 1991)
- Vitamin A vs. control in Indonesian children
- randomise villages (24000 children)
- outcome is mortality
- about 20 of villages didnt get their vitamin A
supply - this analysis ignores clustering by village
25ITT analysis
ITT odds ratio ? 46/77 0.60
26Observed compliance (Vitamin A arm)
27Inferred compliance (Control arm)
CACE odds ratio ? 12/43 0.28
28Estimation method of subtraction
nNE
if nE nS
29Formally
Hence write down and maximise (log-)likelihood. No
te 1-w, w previous wC, wN
30MLE
In this simple case, pCE is estimated as
(dCE/nCE) pCS is estimated as (dS dNE
nS/nE) / (nS nNE nS/nE) provided both
terms 0 (Ill assume this is true)
31MLE
- Use estimates of pCE and pCS to estimate CACE
- as pCE-pCS on RD scale
- as pCE/pCS on RR scale
- as pCE/(1-pCE) / pCS /(1-pCS) on OR scale
- On RD scale only, can show that MLEs obey CACE
ITT / (1-w)
32CACE compared with other quantities
33Vitamin A vs. control summary
- ITT 0.38 vs. 0.64, RR0.60
- CACE 0.12 vs. 0.45, RR0.28
- Per-protocol 0.12 vs. 0.64, RR0.19
- As-treated 0.12 vs. 0.77, RR0.16
- On-treatment and as-treated are too extreme
because of strong selection effect 1.41
(untreated in treatment arm) vs. 0.64 (untreated
in control arm)
34Comparison of assumptions
- Per-protocol and as-treated analyses assume
random non-compliance - no association between compliance-type and
outcome, once treatment effect is taken into
account - CACE analysis assumes exclusion restriction
- randomisation doesnt affect mean outcome for
never-takers and always-takers - no assumption of comparability of different
compliance-types - usually much more plausible
35Extensions to CACE model (1)
- Above we had all the S arm getting S
- Easy to allow for S arm possibly getting E
(Cuzick et al, 1997) - CACE is again estimable under
- 2 exclusion restrictions NACEAACE0
- Either no defiers or same causal effect in
defiers as in compliers (DACE-CACE)
36Extensions to CACE model (2)
- Introduce covariates
- Covariates that predict Y improve precision (as
in ITT analysis) - Covariates that predict D
- also improve precision (unlike in ITT analysis)
(Jo, 2002) - enable estimation of NACE etc. as well as CACE
(Hirano et al, 2000)
37Extensions to CACE model (3)
- Define g difference in outcome between
never-takers and compliers after allowing for
their differences in actual treatment - selection effect
- difference in counterfactual outcomes
- As-treated analysis assumes g0 (random
non-compliance) - CACE and ITT analyses make no assumption about g
- Could estimate g from data leads to CACE
analysis - Instead, introduce appropriate prior information
about g (White, 2005)
38Model for log odds of death (observed risk)
Use informative prior gN(0,s2) for various
values of s
39Vitamin A trial Bayesian CACE analyses
as-treated
CACE
403. Normal outcome
41Normal outcome
- Can simply modify the method of subtraction
work with means instead of proportions - Link to instrumental variables (IV) method
(lecture 5)
42Model
CACE mCE-mCS
43Likelihood
CACE mCE-mCS
44ML estimation
- No closed form solution
- EM algorithm is easy compliance type as the
missing data - usual problems in estimating standard errors
- Newton-Raphson also fairly straightforward
- No directly available software in Stata, but
gllamm can be used see lecture 6.
454. Method comparisons
46Comparison of CACE estimators
- Weve looked at
- Simple estimation using ITT wC CACE (for
difference of means, not risk ratio or odds
ratio) - Method of subtraction
- Maximum likelihood
- For binary outcomes, they all give the same
answer - For continuous outcomes, ML estimation is
different (potentially more efficient see later)
47Simple CACE vs. ITT
- They estimate different parameters
- But they test the same null hypothesis
- Significance levels are equal
- obvious from CACE ITT / (1-w)
- binary case explored in detail by Branson and
Whitehead (2003) significance levels are equal
when likelihood ratio test is performed
48Vitamin A trial profile likelihoods
Likelihood ratio test has same value for both
models
49.. and quadratic approximation (dotted)
Likelihood ratio test has same value for both
models but Wald test doesnt.
50ML estimation of CACE vs. ITT
- For binary outcome, significance levels are equal
- For Normal outcome, significance levels arent
equal - CACE is more efficient whenever theres a
non-zero selection effect or a non-zero treatment
effect - the next slides are thanks to Taeko Becque
51Asymptotic relative efficiencyof CACE vs. ITT
(approximate)
q1 CACE, q2 selection effect, outcome SD 1
52Power of CACE and ITT analyses
Compliance rate 50 Selection
effect-0.5 CACE-0.5 Standard deviation1.5
53Power and compliance rate
Sample size 300 Selection effect-0.5 CACE-0.5
Standard deviation1.5
54Power including covariateweak predictor of
compliance
55Power including covariatestrong predictor of
compliance
56Summary
575. Negative weights method(Kim and White, 2004)
58Negative weights method
- For simplicity take nEnS
- Recall that we subtracted the number of
events/people in the NE cell (never-takers
randomised to E) from the number of events/people
in the S cell (all randomised to S) - Can also achieve this by including them in the S
arm but with a weight -1 - nS/nE in general
59Negative weights in Stata
- Unfortunately many Stata commands are too
sensible to allow negative weights - Exceptions are regress, logistic, cox
- Illustration uses MASS data pretending outcome is
binary
60- . use mass, clear
- . l
- rand screen event n
- 1 1 0 27104
- 1 1 1 43
- 1 0 0 6670
- 1 0 1 22
- 0 0 0 33848
- 0 0 1 113
- . tab rand screen fwn, sum(event) mean
- Means and Number of Observations of
AAA-related death? - Invited to Screened?
- screening? 0 1 Total
- -------------------------------------------
- 0 .00332735 . .00332735
61- . CACE via negative weights
- . tab rand fwn
- Invited to
- screening? Freq. Percent Cum.
- -----------------------------------------------
- 0 33,961 50.09 50.09
- 1 33,839 49.91 100.00
- -----------------------------------------------
- Total 67,800 100.00
- . gen w 1
- . replace w -33961/33839 if randgtscreen
- (2 real changes made)
- . l
- rand screen event n w
62- . logistic event rand fwn, coef ITT analysis
- Logistic regression Number
of obs 67800 - LR
chi2(1) 12.97 - Prob gt
chi2 0.0003 - Log likelihood -1229.0537 Pseudo
R2 0.0052 - --------------------------------------------------
--------------------- - event Coef. Std. Err. z Pgtz
95 Conf. Interval - -------------------------------------------------
--------------------- - rand -.5508119 .1558631 -3.53 0.000
-.856298 -.2453258 - _cons -5.702247 .094229 -60.51 0.000
-5.886933 -5.517562 - --------------------------------------------------
--------------------- - . logistic event screen iwnw, coef CACE
analysis - --------------------------------------------------
--------------------- - event Coef. Std. Err. z Pgtz
95 Conf. Interval - -------------------------------------------------
---------------------
63Bootstrap standard error
- Now we bundle the previous commands into a file
negwt.ado - To speed things up, I use only 1/30 of the
controls
64- . . negwt
- Logistic regression
Number of obs 1210 - LR
chi2(1) 15.43 - Prob
gt chi2 0.0001 - Log likelihood -410.28453
Pseudo R2 0.0185 - --------------------------------------------------
---------------------- - event Coef. Std. Err. z Pgtz
95 Conf. Interval - -------------------------------------------------
---------------------- - screen -.7461635 .1952982 -3.82 0.000
-1.128941 -.363386 - _cons -1.787902 .1141955 -15.66 0.000
-2.011721 -1.564083 - --------------------------------------------------
---------------------- - . bootstrap _bscreen, reps(1000) negwt
- --------------------------------------------------
---------------------- - Var Reps Observed Bias Std. Err.
95 Conf. Interval - -------------------------------------------------
----------------------
65Negative weights summary
- Agrees exactly with direct method in this simple
case - Easy to generalise e.g. to situations with
covariates (weights would have to depend on
covariates) - Naïve standard errors are too small bootstrap
needed - Formal rationale is via unbiased estimating
equations (Abadie 2002)
666. Back-door method (Nagelkerke 2000)
67Back door method
- Idea would like to regress Y on D, adjusting for
U - Its enough to adjust for E that blocks every
indirect path from D to Y (back door criterion) - Approximate E by the residual from a linear
regression of D on R
68Properties
- Nagelkerke et al showed that the method agrees
exactly with the instrumental variables method
for a linear model - For non-linear models it only approximately
agrees with the method of subtraction - Not clear whether standard errors are adequate or
whether bootstrapping is needed - Easy to generalise to more complex settings
69Back door method for MASS
- . reg screen rand fwn
- Source SS df MS
Number of obs 67800 - -------------------------------------------
F( 1, 67798) . - Model 10908.799 1 10908.799
Prob gt F 0.0000 - Residual 5368.59021 67798 .079185082
R-squared 0.6702 - -------------------------------------------
Adj R-squared 0.6702 - Total 16277.3892 67799 .240083028
Root MSE .2814 - --------------------------------------------------
---------------------------- - screen Coef. Std. Err. t
Pgtt 95 Conf. Interval - -------------------------------------------------
---------------------------- - rand .80224 .0021614 371.16
0.000 .7980037 .8064764 - _cons 5.55e-17 .001527 0.00
1.000 -.0029929 .0029929 - --------------------------------------------------
---------------------------- - . predict E, residual
- . tab rand E fwn
70Back door method results
- . logistic event screen E fwn
- Logistic regression
Number of obs 67800 - LR
chi2(2) 20.04 - Prob
gt chi2 0.0000 - Log likelihood -1225.5162
Pseudo R2 0.0081 - --------------------------------------------------
---------------------- - event Odds Ratio Std. Err. z Pgtz
95 Conf. Interval - -------------------------------------------------
---------------------- - screen .4738008 .094594 -3.74 0.000
.3203717 .7007087 - E 1.015178 .295373 0.05 0.959
.5739573 1.795582 - --------------------------------------------------
---------------------- - Note tiny effect of E no evidence of selection
in these data
71Method comparison
- . logistic event rand fwn ITT analysis
- --------------------------------------------------
---------------------------- - event Odds Ratio Std. Err. z
Pgtz 95 Conf. Interval - -------------------------------------------------
---------------------------- - rand .5764816 .0898522 -3.53
0.000 .4247315 .7824496 - --------------------------------------------------
---------------------------- - . logistic event screen fwn As-treated
analysis - --------------------------------------------------
---------------------------- - event Odds Ratio Std. Err. z
Pgtz 95 Conf. Interval - -------------------------------------------------
---------------------------- - screen .476156 .0834625 -4.23
0.000 .3377127 .6713534 - --------------------------------------------------
---------------------------- - . logistic event screen E fwn Approximate
CACE analysis - --------------------------------------------------
---------------------------- - event Odds Ratio Std. Err. z
Pgtz 95 Conf. Interval - -------------------------------------------------
----------------------------
72Summary
- Weve explored a variety of methods for
all-or-nothing treatment switches - randomise to E or S everyone gets all E or all S
- In lecture 4, we will extend to much more complex
patterns of switching - e.g. get E just for 3 months, then S
- Another problem is where some participants get no
treatment at all (or a treatment other than E/S) - ITT difference depends on 2 effects (E vs.
nothing, S vs. nothing) - Walter (2006) adapted the compliance-type
approach but assumed equality between some
compliance-types