Causality, Impartiality and Evidence-Based Policy - Mechanism and Causality in Biology and Economics - page 202

treatment and control groups. Hence, any difference in outcome between groups
can be causally attributed to T in at least one K i relative to the causal structure CS
described by P. This is the conclusion ideal RCTs can clinch. However, according
to Cartwright, we need further assumptions still if we want to generalise this
conclusion to some target population

Θ

.
If we want to affirm, for instance, that T causes O in at least some members of

Θ

,

Cartwright ( 2007 , p. 17) argues, we need assumptions of this kind:
(a) At least one of the subpopulations (with its particular fixed arrangement of
'other' causal factors) in which T causes O in φ is a subpopulation of Θ .
(b) The causal structure and the probability measure are the same in that sub-
population of

Θ as it is in that subpopulation φ .

The warrant for these assumptions too is supposed to come from randomisation,
but we cannot judge whether a group of patients constitutes a random sample
without a previous idea of what factors are to be equally represented (Cartwright
2007 , p. 18). In a trial, we want to form, on the one hand, two treatment groups
that are balanced with respect to known relevant prognostic factors. On the other
hand, we want to avoid unknown confounders to affect the result. Randomisation
supposedly helps us in achieving both goals, but it is neither necessary nor
sufficient to that effect. By sheer chance, a random allocation may yield an
unbalanced distribution of the prognostic factors between the treatment groups
(these are called 'baseline imbalances'). This may bias the comparison between
treatments and invalidate the experimental results, and when imbalances occur,
trialists usually try to correct them (e.g. by repeating the randomisation).
Unknown confounders may differentially influence the outcome in one of the
groups even after a randomised allocation of treatments. Further randomisations
at each step in the administration of the treatment (e.g. which nurse should
administer the treatment today?) may avoid such interferences, but this is imprac-
ticable. Declaring such disturbances as negligible, as many experimenters do,
lacks any justification in the assumed statistical methodology (Urbach 1985 ;
Worrall 2007 ).
Both the correction of imbalanced allocations and the decision to randomise at
different stages of the trial beyond the allocation of treatments require extra-
statistical expert judgement. Against the ideal of mechanical objectivity, we need
an expert who can handle different sources of evidence other than the trial to justify
the acceptance of assumptions (a) and (b). More precisely, we need someone who
can certify that randomisation, the main warrant of (a) and (b), has indeed worked.
Without this judgement, subjective and intransparent as it may be, we cannot safely
generalise the conclusions of the trial to its target population, i.e. ascertain its
external validity. Expert judgements are naturally fallible too, but, according to
Cartwright ( 2007 , p. 19), to rely on mechanical methods without expertise and
watch out for failures is no satisfactory response.