Palgrave Handbook of Econometrics: Applied Econometrics

John DiNardo 101

(1980), for example, for a discussion of the case of “oxygen therapy” for infants
which, far from being harmless, caused blindness.^5
Concern about the ethics of a conventional randomized trial (RCT), where half
the patients are randomized into treatment and half to control, led the surgeons
who had developed the therapy to use a “randomized play-the-winner” statistical
method to evaluate the treatment. The purpose of this convoluted randomization
scheme was to evade

the ethical problem aris[ing] from the fact that during a “successful” random-

ized clinical trial (i.e., one that demonstrates a significant advantage to one

treatment) about half of the trial subjects will receive a treatment which, at the

end of the trial, will be known to be inferior. The recipients of the inferior treat-

ment are individuals whose own outcomes are, in some sense, sacrificed to the

greater good of knowing, with far more certainty than before the trial, the value,

lack of value, or actual harm of the treatments under investigation. (Paneth and

Wallenstein, 1985)

The randomization procedure is too elaborate to be described fully, but this gloss
should be sufficient.^6 The essence of their “modified randomized play-the-winner”
method is that “the chance of randomly assigning an infant to one treatment or
the other is influenced by the outcome of treatment of each patient in the study.
If one treatment is more successful, more patients are randomly assigned to that
treatment” (Bartlettet al.,1985).
Call ECMO “Treatment A” and conventional treatment “Treatment B.” Initially, a
group of biostatisticians prepared a sequence of blinded random treatment assign-
ments. When the outcome of a treatment was known, this information would be
sent to the biostatisticians, who would then create another sequence of blinded
random treatment assignments; the probability of being assigned to A or B, how-
ever, was now a function of the success or failure of the treatment. In their
study,

1. The first infant – with even odds – was randomly assigned to ECMO and
   survived.


2. The second infant – again with even odds – was randomly assigned to conven-
   tional treatment and died.


3. The third infant – with better-than-even odds in favor of being placed in ECMO
   as a result of the first two experiences – was randomized to ECMO and survived.


4. With now even higher odds the next infant was randomized to ECMO and
   survived.

This continued until there was a (pre-specified) total of 12 events. The result
of this unusual randomization was that only one child was randomized to the
conventional treatment and the 11 others received the ECMO treatment.
The outcome of this experiment was that the 11 infants randomized to ECMO
treatment survived; the one infant randomized to conventional treatment died.