Comment by Professor Home on ACCORD,VADT and ADVANCE
Could I just pick up on the ACCORD, VADT, ADVANCE results and the comments in ENDODIABOLOGY please.Firstly there is twice a suggestion that CV deaths were increased in VADT. Though the results have not been published, my memory from the presentations is that there was no different between groups. CV deaths were numerically higher, but the power was way too low for this to be meaningful - indeed the very analysis is dangerous.
Secondly ACCORD did not show increased deaths. It was stopped because of a signal of increased death, which generated a p value of 0.04. This interim observational test should not be regarded as statistically significant. Indeed most of us would not have stopped the study on the basis of the results presented in the paper.
Thirdly you will have seen the UKPDS extension study published since. Attached for those who have not. It is important to realize that ACCORD, ADVANCE and VADT are all consistent with the UKPDS findings, but of course not statistically significant largely because underpowered (VADT always was, ACCORD because stopped early). The consistency and power problems can be seem by considering the headline CV risk reduction in each study :
study HR/RR p ACCORD 0.90 (0.78, 1.04) NS
VADT 0.86 (0.73, 1.04) NS
Advance 0.94 (0.84, 1.06) NS
UKPDS'08 0.84 (0.75, 0.97) 0.01
Someone will no doubt do a meta-analysis on this lot when VADT appears in print - my bet it that it will be statistically significant.Comments, as ever, welcome
Philip Home
Secondly ACCORD did not show increased deaths. It was stopped because of a signal of increased death, which generated a p value of 0.04. This interim observational test should not be regarded as statistically significant. Indeed most of us would not have stopped the study on the basis of the results presented in the paper.
Thirdly you will have seen the UKPDS extension study published since. Attached for those who have not. It is important to realize that ACCORD, ADVANCE and VADT are all consistent with the UKPDS findings, but of course not statistically significant largely because underpowered (VADT always was, ACCORD because stopped early). The consistency and power problems can be seem by considering the headline CV risk reduction in each study :
study HR/RR p ACCORD 0.90 (0.78, 1.04) NS
VADT 0.86 (0.73, 1.04) NS
Advance 0.94 (0.84, 1.06) NS
UKPDS'08 0.84 (0.75, 0.97) 0.01
Someone will no doubt do a meta-analysis on this lot when VADT appears in print - my bet it that it will be statistically significant.Comments, as ever, welcome
Philip Home
posted by Arutchelvam at 11:53 PM
2 Comments:
Terence Aspray commented:
Like everyone, I am concerned about the message promoted by reporting of the
ACCORD trial. However, the ADA statement on ACCORD is
"...intensive participants in ACCORD are now being switched to the standard
treatment program because of an INCREASED DEATH RATE {my capitals} in the
intensive treatment program"
http://www.diabetes.org/for-media/pr-ada-statement-related-to-accord-trail-a
nnouncement-020608.jsp
Has something changed? Philip, the term "signal of increased deaths" is
clearly a technical term; but one which I do not understand. I do understand
that, in a study of an intervention one would HOPE to decrease overall death
rate, the intervention was associated with a higher mortality (14 dths/1000
patients/yr versus 11 dths/1000 patients/yr in the standard treatment
program; a difference of 0.3 deaths per 100 patients per year, we are told)
and a failure to show a benefit for the primary outcome.
Looking at the crude numbers: 257 deaths in 5128 intensively treated versus
223 deaths in 5123 routinely treated, there is a higher mortality rate. From
the paper I cannot get person years, but assuming (and it is a vast
assumption, I appreciate) that there was equal follow up, an attributable
fraction for death due to exposure to intensive treatment is of the order of
20% albeit with wide CI (5-35%)!
I am no statistician and have to take much of this on faith. I know that if
the "signal" were for decreased death, many would argue that it supports
intervention. As things are, I am a little reserved yet; and am in neither
camp!
Best wishes
Terry
Professor Home's reply to Terence
The issue is statistical significance, and thus the possibility of
making both type 1 and type 2 statistical errors.
If you test a predetermined primary endpoint at a predetermined time
it is convention to take p<0.05 as statistical significance. A DSMB
in a study is then faced with a dilemma, because they test
recurrently, and they test mutliple outcomes.
As a result the p value for deaths in ACCORD does not represent
statistical significance by a long way (it would probably have to be
<0.005 for any confidence), and it cannot be said that 'deaths were
increased in the intensive group'. In safety monitoring this kind of
situation is quite common, because we are always testing multiple
endpoints of underpowered events, and one gets quite used for example
to passing for use drugs which say have an increased incidence of that
or the other cancer.
The dilemma facing the ACCORD DSMB was of course the possibility that
the signal was of a real effect, in which case continuing the trial
would kill more people in the intensive therapy group. Note that the
opposite course of action, if glucose control really reduces vascular
events, would be the opposite, and indeed many many more people will
die globally. In this situation what a DSMB does is look for
supporting evidence (eg does MI or stroke drift off in the same
adverse direction, are the increased deaths coherent as to cause?) -
it is a thankless task and has personal dangers in a litigatious world.
I look at the poor level of statistical significance. I look at the
failure of other vascular events to go in the same direction (the
opposite). I look at the lack of coherence of reasons for death or
increased death. I look at the Kaplan Meier curve (which overlaps for
two years, separates, then comes together again, before separating
again). And I conclude we are probably looking at play of chance.
What is clear is the statement 'intensive therapy caused increased
deaths in ACCORD' is wrong. Only the statement 'numerically more
deaths occurred in the intensive therapy group' is correct. These are
very different in meaning.
What should we do? And here we all come together. Pushing intensive
therapy (defined as multiple insulin injections + multiple oral agents
as in ACCORD) has downsides of other kinds. So if that is what is
necessary to achieve near-normoglycaemia (HbA1c <6.5%) then be
cautious. On the other hand if people can get to <6.5 % with
relatively simple approaches, as many now do, then go ahead in
suggesting these approaches.
Further comment welcome
Post a Comment
<< Home