Page 110 - Contributed Paper Session (CPS) - Volume 5

CPS1113 Madhu Mazumdar et al.
patients’ chances of receiving a drain and is not directly associated with the
outcome, blood transfusion, except through its association with the treatment,
CWD. Patients undergoing a procedure performed by surgeons in the ≥80th
percentile of drain usage were labeled as ‘receiving a drain’; the rest were
labeled as ‘not receiving a drain’ (i.e. treatment allocation according to the IV
and not actual drain use; cut-off of 80% chosen to match observered drain
utilization).
Assessing the first assumption – that the IV is associated with the
likelihood of actually receiving a drain – is straightforward. The study showed
a significant correlation between the IV and treatment variable (r=0.78,
P<0.0001). The second assumption, i.e., the IV is independent of (unobserved)
confounders, cannot be assessed directly, but the relationship between the IV
and observed confounders can suggest whether the second assumption is
violated. This relationship can be calculated by comparing covariate
differences between IV-based treatment allocation groups (see Table 1 for a
simplified selection). Compared to the original cohort where groups were
compared based on actual drain use, characteristics of patients and hospitals
appear to be more balanced between groups created by the IV, as
demonstrated by standardized differences of <0.111,12 This is what would be
expected in an RCT.
To assess the validity of the third assumption (the IV is associated with
the outcome only through treatment), we need to explore whether other
treatments that may influence the outcome are given concomitantly with drain
use. For example, patients taking anticoagulants may be at increased risk of
bleeding and subsequent blood transfusion. Therefore, the third assumption
would be violated if the IV is associated with anticoagulant usage, a
confounder. In our example, the correlation between anticoagulant usage and
the percentage of drain utilization was not significant (r=0.0048, P=0.12),
suggesting that the third assumption holds.
After checking the validity of assumptions underlying the chosen IV, we
applied a 2SLS technique to estimate the effect of drains on blood transfusion
risk. In the first stage, each patient’s probability of receiving a drain was
predicted from the IV, adjusted for all observed confounders. Then the odds
of blood transfusion was calculated based on the estimated probability of
receiving a drain, again adjusted for all observed confounders except for the
IV. Patients receiving drains (allocation based on the IV) had significantly
higher odds of blood transfusion (OR: 1.61, 95% CI; 1.37-1.91). This finding
was consistent across all methods examined by Chan et al.

2. Discussion and Conclusion
Double-blind RCTs, the gold standard, are not always feasible.
Alternatively, administrative databases (e.g., electronic medical records, claims,

99 | I S I W S C 2 0 1 9