EFFICACY OF INPATIENT PHOTOTHERAPY IN JAUNDICED NEWBORNS: COMPARISON OF MULTIPLE LOGISTIC REGRESSION, PROPENSITY SCORE, AND INSTRUMENTAL VARIABLE ESTIMATES

Objective:  There are no randomized trials of phototherapy for neonatal jaundice as it is currently recommended by the American Academy of Pediatrics (AAP).  We estimated the efficacy of phototherapy in an observational study using three methods of controlling for confounding: multiple logistic regression, propensity scores, and use of an instrumental variable.

 Methods:  Newborns ≥ 2000 g and ≥ 35 weeks gestation born in Northern California Kaiser hospitals 1995-2004 were included if they had a "qualifying" total serum bilirubin (TSB) level within 3 mg/dL of the  2004 AAP phototherapy threshold (N=22,547) and were excluded if they had a positive Direct Antiglobulin Test (N=1,816).  The intervention was inpatient phototherapy within 8 hours of the qualifying TSB.  The outcome was exceeding the AAP exchange transfusion threshold within 48 hours.  The instrumental variable was the newborn's hospital of birth, previously shown to be a strong independent predictor of phototherapy. To test robustness of the models, we compared efficacy estimates with and without including important confounders.

 Results:  128 infants (0.62%) developed the outcome.  Estimates of the efficacy of inpatient phototherapy within 8 hours were adjusted odds ratio (OR)=0.15; 95% CI (0.07, 0.31) for logistic regression; OR=0.16 (0.08, 0.33) for a logistic adjusting only for the propensity to receive phototherapy; OR=0.21 (0.12, 0.36) for analysis stratified by phototherapy propensity score quintiles, and OR=0.033 (0.01-0.10) for the instrumental variable analysis.  Omitting the strongest confounder (the difference between the qualifying TSB and the AAP phototherapy threshold) increased the OR by a factor of 1.9 for the logistic, 2.0 for propensity and 2.2 for the instrumental variable model.  Omitting all additional covariates except birth hospital further increased the OR from the logistic model by a factor of 1.8, and the OR in the propensity model by 1.3, but had no effect on the instrumental variable model.

 Conclusions:  Inpatient phototherapy is highly effective for neonatal jaundice. In this study, the instrumental variable model was not immune to omitted confounding variables, but appeared more resistant than the other models.  Use of multiple different techniques to control for confounding can suggest robustness or fragility of the results.