210 likes | 368 Views
Biostat 215 Discussion #1. Thomas B. Newman, MD, MPH with thanks to Gabriel Escobar, MD; Michael Kuzniewicz, MD, MPH, Chuck, Eric and Steve). Outline. Background about jaundice and phototherapy Discussion/Review of some key topics Potential and counterfactual outcomes
E N D
Biostat 215 Discussion #1 Thomas B. Newman, MD, MPH with thanks to Gabriel Escobar, MD; Michael Kuzniewicz, MD, MPH, Chuck, Eric and Steve)
Outline • Background about jaundice and phototherapy • Discussion/Review of some key topics • Potential and counterfactual outcomes • Underlying assumptions • Causal model statements as if/then statements
Background for phototherapy dataset • Bilirubin: Yellow breakdown product of heme • Jaundice: Yellow color due to high bilirubin. • DAT: Direct Antiglobulin (“Coombs”) Test: measures maternal antibodies on infant red cells, a cause of jaundice • Phototherapy: Shining light on the babies skin -- helps lower bilirubin levels • Exchange transfusion: replace baby’s blood with donor blood – more drastic treatment to lower bilirubin levels • Kernicterus: Permanent brain damage (cerebral palsy, deafness) from very high bilirubin levels (and other illness)
Background -2 • 2/3 of newborns develop jaundice • 5-15% in Northern CA Kaiser hospitals are treated with phototherapy (PT) • Current treatment thresholds higher than used in previous trials • No randomized trials of PT as currently recommended
NNT paper title • Research Questions • For newborns with total serum bilirubin (TSB) levels close to those at which the AAP recommends phototherapy • What is the efficacy of phototherapy? • What is the number needed to treat (NNT)? • What factors affect the NNT?
Methods • Design: Historical cohort study using electronically available data • Setting: 12 Northern California Kaiser Permanente Medical Care Program hospitals, 1995-2004 • Subjects eligible if ≥ 2000 g, ≥ 35 weeks gestation and qualifying TSB level ± 3 mg/dL from AAP phototherapy threshold (PTT) (N=22,547)
Qualifying TSB levels and key confounder TSB - PTT = + 3 mg/dL TSB - PTT = -2 mg/dL
Qualifying TSB levels and key confounder In phototherapy.dta this variable is called qual_TSB and is coded as follows:
Methods, cont'd • Intervention: hospital phototherapy within 8 hours of the qualifying TSB (“phototherapy”) • Covariates: age, birth weight, gestational age, hospital of birth, sex, qual_TSB (6 categories), year of birth
Outcome variable • Crossing the AAP exchange transfusion threshold within 48 hours of the qualifying TSB = “over_thresh” • Rationale • Incorporates age and AAP risk group • If ET threshold crossed after 48 hours, initial decision not to do PT probably reasonable
Possible discussion questions -1 • What do we mean my potential outcomes and counterfactual outcomes? • Potential outcome estimation is deterministic, but the world isn’t. Is this a problem? • Why is potential outcome estimation like a missing data problem? What does it have to do with imputation?
Possible discussion questions -2 • What is the “positivity” assumption? (Also called the “experimental treatment assumption” = ETA). • What is the “randomization” assumption? • Why should estimates from causal models be considered if…then statements? • Other questions?
Splines -1 • Want to relate a continuous predictor to an outcome • Not reasonable to assume relationship is linear • What to do? • Example: relationship between birth weight and outcome (“over_thresh”) in PT dataset
Stata code *Biostat 215 Discussion #1 use "/Users/thomasnewman/Documents/TEACH/BIOSTAT/Biostat 215 Causal modeling MSMs DAGs etc/Stata and labs/phototherapy.dta" tab birth_wt logistic over_thresh birth_wt estat gof, group(10) tab * Logistic model fits poorly tabu bwcat over_thresh, row chi * Can use indicator variables logistic over_thresh i.bwcat * What will estat gof show? estat gof, group(10) tab * Indicator variables lead to discontinuity at cutoffs. * Demonstrate REGULAR splines mkspline bwsp 5 = birth_wt, di scatter bwsp1 birth_wt scatter bwsp2 birth_wt scatter bwsp3 birth_wt scatter bwsp4 birth_wt ggen junk = bwsp1+ bwsp2+ bwsp3+ bwsp4+ bwsp5 regress birth_wt junk logistic over bwsp* * Contrast with RESTRICTED CUBIC SPLINES mkspline bwsp = birth_wt, cubic scatter bwsp1 birth_wt scatter bwsp2 birth_wt scatter bwsp3 birth_wt scatter bwsp4 birth_wt logistic over bwsp* **Alternatively, include quadratic term sum birth_wt ggen bw2=(birth_wt-3.324)^2 logistic over birth_wt bw2 estat gof, group(10) tab