4A-01 - Clarifying Contradictions: Transportability Using Doubly Debiased Machine Learning in 17P Trials and Preterm Birth Outcomes

Background: The Meis et al. trial identified a clinically significant benefit of 17-alpha-hydroxyprogesterone caproate (17P) on reducing the risk of recurrent preterm birth (PTB). However, the confirmatory PROLONG trial identified no effect of 17P on PTB. PROLONG had a different study population than Meis, despite efforts to recruit similar trial participants. Meis participants had a higher count of prior PTBs, higher pre-pregnancy BMI, fewer years of education, and were more likely to drink, smoke, or use substances during pregnancy. Approximately 59% of Meis participants identified as non-Hispanic Black, compared to fewer than 7% of PROLONG.

Objectives: We modified existing transportability methods to investigate whether the conflicting results from these trials could be explained by measured differences. We hypothesized that, conditional on measured effect-measure modifiers, the transported effect should be equivalent to the original effect observed in the trial population to which we transported (e.g. the effect of transporting Meis to PROLONG should equal the original effect observed in PROLONG).

Methods: We first estimated risk differences (RDs) for PTB comparing 17P to placebo in both trials, adjusting for measured baseline variables. We then transported the risk of PTB in each arm of Meis to estimate the causal RD in the PROLONG population and vice versa from PROLONG to the Meis population. We used double debiased machine learning for estimation, which combines a doubly robust estimator with an ensemble learner and cross fitting for estimating nuisance functions. We additionally evaluated if the results changed among individuals enrolled in the US and those with only one prior PTB. We obtained 95% CIs via nonparametric bootstrap (10,000 iterations).

Results: Comparing 17P to placebo, the adjusted RD was -17.5% (95% CI: -27.2%, -7.8%) in Meis and 1.1% (-3.3%, 5.2%) in PROLONG. Comparing 17P to placebo, the doubly debiased machine learning RD was -18.6% (-57.3%, 6.2%) when transporting Meis to PROLONG and 5.2% (-18.0%, 17.4%) when transporting PROLONG to Meis. To evaluate our hypothesis, we compared the Meis-to-PROLONG transported RD (-18.6%) with the RD in PROLONG (1.1%), and PROLONG-to-Meis transported RD (5.2%) with the RD in Meis (-17.5%). These results were consistent across subgroup analyses, other sensitivity analyses, and across risk ratios (modification on the multiplicative scale).

Conclusions: We were able to advance transportability methods by integrating state of the art machine learning techniques for estimation. This transportability approach did not reconcile the conflicting results between the two trials of 17P and SPTB. There may be several explanations for these results including unmeasured effect measure modification between the trials.