Contributions of preeclampsia to preterm delivery: a hypothetical interventional cohort study using incremental propensity scores Authors: Wen Wei Loh, Cande V Ananth Published: 2026 ABSTRACT AbstractBackgroundPreeclampsia remains a leading cause of preterm delivery (PTD). Conventional average causal effect estimands may not always be perfectly suited for assessing preeclampsia’s contributions because they rely on uniform counterfactual scenarios, such as everyone suffering from preeclampsia. A complementary approach is the incremental propensity score intervention (IPSI), which can address a more action-oriented causal inquiry: How would the population average risk of PTD change from current observed levels if each individual’s odds of preeclampsia were reduced by a given amount?MethodsWe applied IPSI to two US population cohorts of singleton pregnancies, separated by two decades (1998–2002, 16.7 million; 2020–2023, 14.2 million), to evaluate the causal effects of gestational hypertension/preeclampsia (GHTN/PE) on PTD within each cohort. We developed and implemented a sensitivity analysis procedure using bias formulas to gauge the impact of potential unmeasured confounding.ResultsBetween 1998 and 2002, halving the individual odds of GHTN/PE would reduce the population average risk of PTD from observed [risk ratio (RR) = 0.969, 95% confidence interval (CI) 0.969–0.970]. Between 2020 and 2023, when GHTN/PE was more prevalent, halving the odds of preeclampsia would lead to an even greater reduction in average PTD risk (RR = 0.948, 95% CI 0.947–0.948). Sensitivity analyses demonstrated that these conclusions held under assumptions encoding relatively strong unmeasured confounding.ConclusionsThe IPSI approach offers more nuanced understanding and clinically meaningful interpretations of causal effects, such as by how much partially reducing the odds of GHTN/PE, while allowing it to remain non-zero, can lower the average risk of PTD.Keywords: incremental propensity score intervention (IPSI), causal estimand, gestational hypertension, preeclampsia, preterm delivery, sensitivity analysis, unmeasured confounding Using the incremental propensity score intervention (IPSI) approach, we seek to answer the causal query: How would the average risk of preterm delivery change from current observed levels if we merely reduced each individual’s odds of preeclampsia while still allowing it to remain non-zero? Using the incremental propensity score intervention (IPSI) approach, we seek to answer the causal query: How would the average risk of preterm delivery change from current observed levels if we merely reduced each individual’s odds of preeclampsia while still allowing it to remain non-zero? In each of two birth cohorts separated by two decades, we found that modest reductions in the odds of gestational hypertension/preeclampsia as recorded in US vital records could already lower the average risk of preterm delivery, even when allowing for relatively strong unmeasured confounding. In each of two birth cohorts separated by two decades, we found that modest reductions in the odds of gestational hypertension/preeclampsia as recorded in US vital records could already lower the average risk of preterm delivery, even when allowing for relatively strong unmeasured confounding. By applying IPSI, we gained unique insights into how practically achievable partial reductions in the odds of gestational hypertension/preeclampsia can lower the population average risk of preterm delivery relative to current observed levels in the US. By applying IPSI, we gained unique insights into how practically achievable partial reductions in the odds of gestational hypertension/preeclampsia can lower the population average risk of preterm delivery relative to current observed levels in the US. Introduction Preeclampsia is a leading cause of adverse perinatal outcomes, including preterm delivery (PTD) [1, 2] and perinatal mortality [3, 4]. Assessing preeclampsia’s effects often involves estimating average causal effects (ACEs) that compare counterfactual scenarios in which everyone experiences the same exposure level. While ACEs are essential for enhancing scientific understanding of causal relationships, they may not always be ideal for all causal queries [5]. First, relying solely on counterfactual scenarios precludes comparisons to observed, factual outcomes. Second, a practically feasible intervention that can achieve either scenario (e.g. eliminating or inducing preeclampsia across the entire population) may be less achievable. We implement a complementary approach that permits more nuanced assessments of preeclampsia’s contributions to PTD: the incremental propensity score intervention (IPSI [6]). IPSI frames causal queries as: How would the currently observed population average outcome change by reducing the propensity score a given amount (while remaining greater than zero)? IPSI integrates real-world heterogeneity in the individual odds of exposure into its definition and considers hypothetical interventions aimed at reducing—without eliminating or mandating—the exposure. IPSI has recently been introduced to investigate the effects of diet on preeclampsia [7] and of job loss on self-injury mortality [8]. We apply IPSI to evaluate preeclampsia’s causal effects on PTD using two US population cohorts of singleton pregnancies two decades apart. IPSI is ideally suited for this causal endeavor because preeclampsia is not experienced uniformly: real-world interventions reduce the risk of preeclampsia by varying degrees, but do not eliminate it. It is unrealistic to eliminate preeclampsia because a certain portion of preeclampsia cases may be due to developmental errors in placental implantation, such as uteroplacental dysfunction, that are difficult or impossible to prevent [9–11]. To assess IPSI using risk ratios (RRs), we describe a causal RR as the ratio of the average IPSI outcome to the observed outcome. This RR addresses the causal query: How would the population average risk of PTD differ from currently observed levels if each individual’s odds of preeclampsia were reduced? Finally, as with most causal analyses using observational data, conclusions using IPSI remain susceptible to confounding bias. Because no sensitivity analyses specific to the IPSI paradigm are currently available, we developed a procedure for IPSI using established bias formulas for the ACE [12] to gauge the impact of unmeasured confounding. Methods We conducted a secondary analysis of data from two US population cohorts of singleton pregnancies delivered at ≥20 weeks of gestation (16.7 million, 1998–2002; 14.2 million, 2020–2023). These data are derived from the US live birth and fetal death records assembled by the National Center for Health Statistics of the Centers for Disease Control and Prevention. No institutional review board approval was sought as these data are deidentified and publicly available. We contextualize the IPSI within each cohort spanning two decades because preeclampsia prevalence in the US increased between 1980 and 2010 [13], whereas PTD rates remained relatively stable due to improved obstetrical care. Our exposure (X) was a broad classification, as recorded in US vital records, that included diagnoses of either hypertension only (gestational hypertension, defined as systolic blood pressure ≥140 mmHg and/or diastolic blood pressure ≥90 mmHg, diagnosed at ≥20 weeks of gestation) or hypertension with proteinuria (preeclampsia) [14]. We henceforth refer to this composite simply as “GHTN/PE.” Validation studies showed limited sensitivity for gestational hypertension (33% in New York City; 66% in Vermont) but high specificity (>99%) [15]. PTD was the binary outcome Y (=1 if gestational age < 37 completed weeks). Let Y1 and Y0 denote the potential PTD outcome with or without GHTN/PE, respectively. We maintained causal consistency (assumption 1 of Kennedy [6]) so that Y=XY1+(1−X)Y0 almost surely. Baseline covariates, jointly denoted by C, were maternal age (<15, 15–19,…, 40–44, and ≥45 years), live-born parity (primiparous, parity 2, and parity ≥3), maternal education (<9, 9–12, 13–16, and ≥17 years of schooling), race/ethnicity (White, Black, or other in 1998–2002; Non-Hispanic White, Non-Hispanic Black, Hispanic, or other in 2020–23), tobacco smoking during pregnancy (nonsmoker or smoker), marital status (married or single; only in 1998–2002), chronic hypertension, pre-pregnancy body mass index (BMI; only in 2020–23), pregestational diabetes mellitus (only in 2020–23), and calendar year of birth. We intentionally avoided stratifying by or adjusting for causal intermediaries between GHTN/PE and PTD (e.g. gestational age) to maintain the causal effect’s definition and avoid overadjustment bias [16–18]. Incremental propensity score intervention We briefly introduce IPSI [6]. Let π(C)=Pr(X=1|C) denote an individual’s current propensity score for GHTN/PE. An IPSI is characterized by a shift in π(C) parameterized by δ [5, 6]. Specifically, let Qδ(C) denote the shifted propensity score. Then, δ describes the odds ratio between Qδ(C) and π(C), so that solving for Qδ(C) yields the functional form: The ensuing IPSI average outcome under (1) for a given δ is then: The causal quantity in (2) addresses the query: What is the average PTD risk after shifting each individual’s current propensity of GHTN/PE by δ through an intervention to Qδ(C)? The IPSI average outcome is assessed for a posited grid of values for δ encoding hypothetical interventions on GHTN/PE of different magnitudes, with δ=1 recovering the observed outcome E(Y). Relevant considerations for practically determining these values, whether a uniform value is necessary, and when the IPSI equals the ACE are addressed in the discussion. Details of IPSI, including estimation using the influence function and inference using confidence intervals (CIs) via the multiplier bootstrap, are explained elsewhere [5–7]. Substantive interest is often also in the relative change between the IPSI average outcome and the observed outcome (under δ=1). We consider the ratio of the average risk of PTD under a given δ<1 (representing lowered odds of GHTN/PE) to under δ=1 (at current levels), defined as RR(δ)=E(YQδ)E(Y). Its asymptotic variance can be estimated through an application of the delta method: The covariance of the average IPSI and observed outcomes, cov{E(YQδ),E(Y)}, can be obtained using the covariance between the respective individual influence functions. An asymptotic Normal-based 100(1−α)% CI would then be RR^(δ)±z1−α/2var^{RR(δ)}, where z1−α/2 denotes the 100(1−α/2) percentile of a standard normal distribution. The percentile can be adjusted to account for d pairwise comparisons of posited δ values by replacing α with α/d using a Bonferroni correction. We fitted a logistic regression model to estimate propensity scores, including all main effects and two-way interaction terms among the covariates. To predict potential outcomes, we fitted separate logistic regression models for each exposure subgroup, each including all main effects and two-way interaction terms among the covariates. This allowed for different exposure-specific coefficients and was equivalent to a single outcome model with all possible exposure-covariate(-covariate) interactions. Missing data No data were missing for either GHTN/PE or PTD in the 1998–2002 cohort. There were no missing data for PTD in the 2020–23 cohort; however, approximately 0.2% had missing GHTN/PE status. A few baseline covariates had missing values in each cohort. To account for partial missingness, we applied multiple imputations with chained equations for each variable predicted by all other variables, using predictive mean matching [19], assuming data were missing at random. Each imputation was used to complete the dataset before estimating the IPSI average outcomes, and the estimates from twenty imputations were combined using Rubin’s rule. We did not impute covariates entirely missing in each cohort (pre-pregnancy BMI and pregestational diabetes mellitus in 1998–2002; marital status in 2020–23). Sensitivity analysis to unmeasured confounding for IPSI Consistent estimation of IPSI causal effects in (2) assumes no unmeasured confounding (assumption 2 of Kennedy [6]): We propose a sensitivity analysis procedure specific to IPSI for putative violations using bias correction formulas for the ACE [12]. Details are provided in Box 1. Sensitivity analysis assumptions and procedure Let U denote an unmeasured (set of) confounder(s) of X and Y so that adjusting for C and U jointly suffices to eliminate confounding bias; i.e. Let γxcu=E(Y|X=x,C=c,U=u)E(Y|X=x,C=c,U=0) denote the conditional risk ratio for U=u relative to a reference level U=0. [29] The conditional average potential outcome E(Yx|C) can then be recovered from an estimator E^(Yx|C) that does not adjust for U using the closed-form expression for the bias factor rx(C) (on the risk ratio scale; Theorem 3 in VanderWeele and Arah [12]): Hence, adjusting for U when U and X are conditionally independent given C is unnecessary because Pr(U=u|X=x,C=c)=Pr(U=u|C=c) for all U=u so rx(c)=1 for all C=c. This general expression allows for binary or multi-valued U, comprising either a single variable or a set of variables. A crucial distinction between the average causal effect (ACE) estimand and incremental propensity score intervention (IPSI) average outcome in (2) is that, under no unmeasured confounding, the positivity assumption [25, 32] is sufficient for identifying the former but unnecessary for the latter. The positivity assumption contingent on no unmeasured confounding in (4) is: However, when there is unmeasured confounding so that (4) is violated, a sensitivity analysis using the bias factor approach then relies on the following version of positivity to ensure rx(C) is non-zero: Positivity in (8) is also required to ensure γxcu takes on finite values, i.e. The reference level U=0 in defining γxcu must thus satisfy both (8) and (9). We elaborate on the technical details and implications of (8) and (9) in the Supplementary Material. Hence, when assuming no unmeasured confounding, positivity (given observed covariates) is unnecessary for IPSI, whereas allowing for unmeasured confounding introduces positivity (given observed and unobserved covariates) as a requirement to use these bias factors for a sensitivity analysis. A sensitivity analysis procedure, therefore, proceeds by fixing values for the joint sensitivity parameters γxcu and Pr(U=u|X=x,C=c), then calculating the bias factors rx(C),x=0,1, for each individual with observed C. These bias factors are then applied to E^(Yx|C), the predicted potential outcomes using only C, to recover the adjusted potential outcomes accounting for U, e.g. E˜(Yx|C)=E^(Yx|C)/rx(C). The bias factors are also applied to π^(C), the propensity score estimates using only C, to recover the adjusted propensity score, e.g. π˜(C)=r1(C)π^(C). The ensuing IPSI average outcome in (2), having accounted for unmeasured U under the fixed values of the sensitivity parameters, can be recalculated using π˜(C) and E˜(Yx|C),x=0,1. However, there are two practical challenges specific to IPSI. First, it is computationally infeasible to recalculate IPSI estimates and CIs for all posited values of δ while holding the sensitivity parameters, whose cardinality grows exponentially with their granularity, fixed. Second, unlike other causal effect estimands [20, 21], it is practically impossible to visualize the IPSI average outcomes varying across δ and the sensitivity parameters simultaneously. Therefore, we simplify the assumptions following current methods for the ACE [12] to present sensitivity analyses for unmeasured confounding that systematically vary the sensitivity parameter values. Details are provided in the Supplementary Material. Results GHTN/PE prevalence increased between the two cohorts. Of the 16.7 million singleton births between 1998 and 2002, 4.1% had GHTN/PE. Of the 14.2 million singleton births between 2020 and 2023, 9.3% had GHTN/PE. The distributions of maternal characteristics by GHTN/PE for each cohort are shown in Table 1. PTD risk remained relatively stable between the two cohorts (8.6% in 1998–2002; 9.0% in 2020–23). We first calculated the ACE. The average counterfactual PTD risk assuming no one had GHTN/PE was 8.1% (95% CI 8.0–8.1) in both cohorts, but the counterfactual risk assuming everyone had GHTN/PE was 21.2% (95% CI 21.1–21.3) in 1998–2002 and 19.4% (95% CI 19.3–19.5) in 2020–23. Therefore, the ACE (RR) was 2.63 (95% CI 2.62–2.65) and 2.42 (95% CI 2.41–2.44), respectively. Table 1Distribution of maternal characteristics in relation to GHTN/PE for all US singleton births in 1998–2002 and 2020–23. 1998–2002 2020–23 Total births GHTN/PE No GHTN/PE Total births GHTN/PE No GHTN/PE Missing VariableNumber%Number%Number%Number%Number%Number%Number%Number of births16 700 491100.0690 761100.016 009 730100.014 167 656100.01 312 366100.012 833 167100.022 123100.0Maternal age (years) <1537 1190.222810.334 8380.271850.16840.164700.1310.1 15–191985 24211.994 14513.61 891 09711.8584 3554.155 2914.2527 8924.111725.3 20–244 267 23225.6178 71325.94 088 51925.52 523 06617.8237 33218.12 281 62717.8410718.6 25–294 496 14126.9184 21526.74 311 92626.93 948 82527.9362 08427.63 580 54827.9619328.0 30–343 779 89122.6141 75720.53 638 13422.74 272 11230.2379 77028.93 886 28230.3606027.4 35–391 773 76310.671 64610.41 702 11710.62 291 98016.2216 78516.52 071 61416.1358116.2 40–44346 1112.117 0562.5329 0552.1503 1623.655 3494.2446 9353.58784.0 ≥4514 9920.19480.114 0440.136 9710.350710.431 7990.21010.5Parity Primiparous5 517 00133.0328 90347.65 188 09832.45 602 76839.5665 27450.74 930 02738.4746733.8 24 878 47229.2170 16124.64 708 31129.44 409 35431.1335 12725.54 069 47831.7474921.5 ≥36 226 67437.3189 28627.46 037 38837.74 114 79629.0309 31323.63 799 29129.6619228.0 Missing78 3440.524110.375 9330.540 7380.326520.234 3710.3371516.8Maternal education (years) <9825 5154.925 3983.7800 1175.0447 4463.227 4232.1418 9373.310864.9 9–127 931 09347.5330 58447.97 600 50947.54 859 41034.3442 14433.74 409 06034.4820637.1 13–166 184 80937.0269 13539.05 915 67437.06 764 81047.7663 45050.66 093 57047.5779035.2 ≥171 747 20110.564 9309.41 682 27110.51 862 60013.1161 38212.31 699 25613.219628.9 Missing11 8730.17140.111 1590.1233 3901.617 9671.4212 3441.7307913.9Race/ethnicity White13 119 21078.6544 01278.812 575 19878.57 143 34750.4705 49153.86 427 53550.110 32146.7 Black2 772 88716.6122 87917.82 650 00816.61 981 80914.0219 13916.71 759 12213.7354816.0 Hispanic3 581 03825.3274 46320.93 301 74125.7483421.9 Others808 3944.823 8703.5784 5244.91 461 46210.3113 2738.61 344 76910.5342015.5Smoking during pregnancy Nonsmoker14 358 51586.0606 08287.713 752 43385.913 353 17094.31 230 93393.812 105 36494.316 87376.3 Smoker2 017 40812.172 19410.51 945 21412.2751 7155.375 8395.8673 7685.321089.5 Missing324 5681.912 4851.8312 0831.962 7710.455940.454 0350.4314214.2Marital status Married11 061 63366.2451 02665.310 610 60766.3 Single5 623 23533.7238 77134.65 384 46433.6 Missing15 6230.19640.114 6590.1Chronic hypertension Absent16 475 24198.7686 94699.415 788 29598.613 749 39697.01 309 97699.812 439 42096.9 Present139 4210.838150.6135 6060.8396 1372.823900.2393 7473.1 Missing85 8290.500.085 8290.522 1230.222 123100.0Prepregnancy BMI (kg/m 2) <18.5381 1322.717 4171.3362 9682.87473.4 [18.5,25)4 944 16334.9283 25121.64 653 80936.3710332.1 [25,30)4 020 96928.4346 60426.43 669 25328.6511223.1 [30,35)2 366 07016.7279 72321.32 083 63716.2271012.2 [35,40)1 227 2968.7182 47413.91 043 5528.112705.7 ≥40931 4326.6176 62613.5753 8735.99334.2 Missing296 5942.126 2712.0266 0752.1424819.2Pregestational diabetes mellitus Absent13 983 40198.71 281 07797.612 702 32499.0 Present162 1321.131 2892.4130 8431.0 Missing22 1230.222 123100.0Preterm delivery Term15 265 65791.4544 69378.914 720 96492.012 891 44691.01 062 11080.911 813 34692.115 99072.3 Preterm1 434 8348.6146 06821.11 288 7668.01 276 2109.0250 25619.11 019 8217.9613327.7 Distribution of maternal characteristics in relation to GHTN/PE for all US singleton births in 1998–2002 and 2020–23. We used IPSI to assess by how much the average PTD risk could be reduced by lowering the odds of GHTN/PE. Therefore, we focused on the values of δ≤1 and posited a uniformly spaced grid of δ=0.10,0.15,…,1.00 (i.e. odds of GHTN/PE reduced by 90%,85%,…,0%, respectively). The distribution of shifted propensity scores Qδ(C) for each δ value within each cohort is plotted in Fig. 1. An IPSI that halves the odds of GHTN/PE (δ=0.5) reduced the average GHTN/PE risk (4.1% to 2.1% in 1998–2002; 9.3% to 5.0% in 2020–23). Hence, even for individuals at low risk of GHTN/PE (with smaller propensity scores), it remains feasible to further lower their risk. The propensity scores were more variable in the later cohort because the fitted model had higher predictive accuracy (C-statistic = 52% in 1998–2002, versus 67% in 2020–23). The absolute mean difference in covariates (including product interaction terms), reweighted by inverse propensity scores, was no larger than 0.006 in 1998–2002 and 0.008 in 2020–23, indicating measured covariate balance. Distributions of propensity scores for GHTN/PE (%) under different IPSIs encoded by δ using US singleton births in 1998–2002 (green) and 2020–23 (red). The shift parameter δ encodes the odds ratio between the intended and naturally occurring odds of exposure. Each cross indicates the (marginal) average incremental propensity score under a given δ on the horizontal axis. Each vertical line represents the 0.5%- and 99.5%-tile of the propensity score distribution under a given δ. To improve visual clarity, the results for 2020–23 are shifted slightly to the right to avoid overlap. We can draw three conclusions from the IPSI results plotted in Fig. 2. First, as expected, when δ=1, the estimated average PTD equaled the observed sample mean (indicated by a horizontal broken line). Second, even modestly lowering the odds of GHTN/PE could lead to a commensurate decrease in the estimated average PTD risk. Between 1998 and 2002, halving the odds of GHTN/PE (δ=0.5) would lower the average PTD risk from the observed 8.59% (95% CI 8.57–8.67) to 8.33% (95% CI 8.31–8.35), with RR(0.5)=0.969 (95% CI 0.969–0.970). Between 2020 and 2023, halving the odds of GHTN/PE would lower the average PTD risk from the observed 9.01% (95% CI 8.98–9.03) to 8.54% (95% CI 8.51–8.56), with RR(0.5)= 0.948 (95% CI 0.947–0.948). The larger IPSI (RR) in the later cohort, using the same δ=0.5, was expected because GHTN/PE prevalence was higher while average PTD risk remained relatively stable; it did not, on its own, indicate a stronger causal effect of GHTN/PE on PTD. For the same reason, the ACE (RR) was weaker in the later cohort. Estimates of the IPSI average PTD (%) under different IPSIs encoded by δ using US singleton births in 1998–2002 (left; green) and 2020–23 (right; red). The shift parameter δ encodes the odds ratio between the intended and naturally occurring odds of exposure. Each circle corresponds to a point estimate, and each vertical line corresponds to the 95% CI, under a given value of δ on the horizontal axis. Each vertical dotted line corresponds to a value of δ based on external real-world evidence. The horizontal broken line corresponds to the average observed risk of PTD under δ=1. Third, in both cohorts, suppressing the odds of GHTN/PE by 90% (δ=0.1) would potentially lead to almost the same reduced average PTD risk of 8.11% (95% CI 8.09–8.13). However, as shown in Fig. 3, the relative reduction in the later cohort [RR(0.1) = 0.900 (95% CI 0.899–0.902)] was larger than in the earlier cohort [RR(0.1) = 0.944 (95% CI 0.943–0.945)]. Estimates of the risk ratio contrasting the IPSI average PTD under each posited value of δ relative to the observed risk (δ=1) using US singleton births in 1998–2002 (left; green) and 2020–23 (right; red). The shift parameter δ encodes the odds ratio between the intended and naturally occurring odds of exposure. Each circle corresponds to a point estimate, and each vertical line corresponds to the 95% CI, under a given value of δ on the horizontal axis. The horizontal broken line indicates no difference. BMI as a risk factor for preeclampsia and preterm delivery BMI influences the risks of preeclampsia. However, because BMI was recorded only for the 2020–23 cohort, we examined its impact by deliberately excluding BMI and repeating the analysis for that cohort. First, excluding BMI resulted in smaller estimated propensity scores on average only among those with GHTN/PE (see Supplementary Material for Supplementary Fig. S1). Second, the predictive accuracy of the outcome models remained approximately the same regardless of whether BMI was included. The C-statistics for the separate models predicting PTD among GHTN/PE and no GHTN/PE births were 0.50 and 0.52, respectively, with BMI; after excluding BMI, the C-statistics remained unchanged (0.49 and 0.53, respectively). Finally, IPSI estimates without BMI were almost identical to those with BMI; the gradient of the estimates as δ decreased was only very slightly steeper for smaller δ values when excluding BMI (see Supplementary Material for Supplementary Fig. S2). Subgroup analyses We estimated subgroup effects by stratifying on parity or maternal age. The results in Fig. 4 indicate that primiparous births have a higher average PTD risk than multiparous births. The results in Fig. 5 indicate that the average PTD risk was highest among individuals ≤19 years old in the earlier cohort, but highest among those ≥35 years old in the later cohort. Subgroup-specific estimates of the IPSI average PTD (%) under different IPSIs encoded by δ using US singleton births stratified by parity. The left (green) and right (red) columns display the 1998–2002 and 2020–23 cohorts, respectively. Explanations of how to interpret each panel are provided in the caption of Fig. 2. Subgroup-specific estimates of the IPSI average PTD (%) under different IPSIs encoded by δ using US singleton births stratified by maternal age group. The left (green) and right (red) columns display the 1998–2002 and 2020–23 cohorts, respectively. Explanations of how to interpret each panel are provided in the caption of Fig. 2. Unmeasured confounding bias Using our proposed procedure, we undertook a sensitivity analysis of unmeasured confounding. We considered four distinct data-driven possibilities for the sensitivity parameter values. The results, plotted in Supplementary Figs S3 and S4, suggested that, across all four scenarios, we could maintain the conclusion that lowering the odds of GHTN/PE, even by a modest amount, would reduce average PTD risk. Discussion Principal findings We sought to use IPSI to answer the causal query: How would the average PTD risk in the US change from current observed levels if we could partially reduce each individual’s odds of GHTN/PE? We demonstrated that halving the odds of GHTN/PE in 1998–2002 sufficed to lower PTD risk in the US population by RR(0.5) = 0.969. In a more contemporary cohort (2020–23) with almost twice the GHTN/PE prevalence of 1998–2002, while PTD risk remained relatively stable, applying the same hypothetical intervention would yield a larger reduction of RR(0.5) = 0.948. These counterfactual shifts in PTD from hypothetically intervening on GHTN/PE remained robust to certain violations of unmeasured confounding. We conducted cohort-specific analyses based on identification assumptions within each cohort. No cross-cohort comparisons were formally tested because discrepancies between cohorts (e.g. changes in ascertaining GHTN/PE, different covariate adjustment sets, or coding) can affect estimates and undermine comparability. Hence, differences between cohorts should be interpreted cautiously rather than definitively as evolving effects. Comparison of IPSI causal estimands with other causal effects We briefly compare the IPSI and ACE in Table 2. Importantly, while the ACE permits heterogeneous hypothetical interventions, they must produce the same exact endpoint of a uniform counterfactual scenario (e.g. eliminating or inducing GHTN/PE) across the entire population. This can lead to unrealistic promises about the causal effect because there may be no feasible intervention that can achieve either scenario [22, 23]. Hence, it is possible for a non-null ACE, whereas the IPSI does not achieve this because the shifts in individual δ are too small. We formally compare the IPSI with the (conditional) ACE, the average effect of treatment among the treated (ATT), and the average effect of treatment among the control/untreated (ATC/ATU) in the Supplementary Material. Table 2Comparison of the IPSI causal effect and the ACE.IPSI causal effectACEDefinition using potential outcomes E(YQδ)E(Y) E(Y1)E(Y0)What is the change in the population average outcome ……from the status quo if current observed exposure odds were shifted by δ?…if everyone vs. no one was counterfactually exposed?Expected change in the outcome brought about by ……a hypothetical partial reduction in exposure odds vis-à-vis observed levels.…comparing the hypothetical scenarios in which everyone vs. no one was counterfactually exposedMultiplying each individual’s current odds of exposure by a factor ……of δ, as defined in (1), that results in shifted propensity scores that vary across individuals.…that results in a shifted propensity score in (1) exactly equal to either zero or one uniformly for everyone, regardless of their current propensity score.Type of hypothetical interventionStochastic and dynamicDeterministic and staticCommentShift parameter δ need not be uniform; it can be individual-specific based on subject matter knowledgeCorresponds to comparing IPSI causal effects where δ=∞ vs. δ=0, respectively. Comparison of the IPSI causal effect and the ACE. In this paper, we considered IPSIs that shift the propensity score uniformly by δ for all individuals, but still allow the ensuing shifted propensity scores to vary across individuals. In practice, δ need not be identical for everyone. For example, one can postulate an IPSI that alters the odds of GHTN/PE differently depending on an individual’s preexisting PTD risk. Hence, researchers can leverage subject-matter knowledge to flexibly conceive IPSIs with fine-grained δ values. External real-world interventions have been shown to reduce preeclampsia risk Preeclampsia complicates about 7%–10% of all pregnancies, but it confers disproportionately increased risks of PTD and other adverse perinatal outcomes. For instance, preeclampsia remains one of the chief indications for clinician-indicated PTD [1, 2] and a strong risk factor for spontaneous PTD [24]. Substantial efforts through prophylactic antiplatelet clinical interventions have been undertaken to reduce preeclampsia incidence. A meta-analysis [25] of 10 studies (18 911 subjects) showed that low-dose aspirin (81 mg/day) could achieve a reduction of 30% in the general population (RR 0.70, 95% CI 0.52–0.95). Furthermore, different strategies have been developed to prevent preeclampsia, depending on whether it is early-onset, which has a placental origin, or late-onset, which is associated with metabolic syndrome [26]. Given this evidence, hypothetical interventions to reduce but not entirely eliminate the risk of preeclampsia seem practically achievable and clinically justified. Obesity and preeclampsia risk Pre-pregnancy overweight or obesity is a risk factor for preeclampsia [27]. Female obesity rates have increased from 33.1% in 2001–02 to 41.4% in 2017–18 in the US [28]. These findings prompted us to hypothesize that the increase in preeclampsia prevalence between the cohorts due to obesity may provide insights into whether a hypothetical preeclampsia intervention affects PTD. However, our fitted outcome models showed that BMI was a relatively weak predictor of PTD, alongside other covariates, to the extent that excluding BMI had a negligible impact on our results using IPSI. Limitations and future directions Our application of IPSI is motivated by how IPSI enables quantification of GHTN/PE’s contribution to PTD on a continuum, such as by how much modest reductions in the individual odds of GHTN/PE will decrease average PTD risk. However, another motivation for using IPSI is that, when no unmeasured confounding (4) holds, positivity [29, 30] is unnecessary for consistent estimation [5–7]. While IPSI avoids the need to empirically diagnose violations of positivity, this does not preclude researchers from assessing or diagnosing putative violations in practice. For example, we found little evidence of a glaring lack of overlap in the propensity score distributions between GHTN/PE and no GHTN/PE in either cohort (see Supplementary Material for Supplementary Fig. S1). We estimated propensity scores using logistic regression models with multiple interaction terms (the fitted model coefficients are in the Supplementary Material). Incorrect model specification can have a greater impact on IPSI estimates than on conventional doubly robust estimators of the ACE [7]. Machine learning algorithms can mitigate the risks of such biases [31]. IPSI estimators are based on efficient influence functions that enable these methods to model the propensity score and outcome [32]. Conclusion We argue that IPSI is better suited than ACE for assessing the nuanced causal contributions of preeclampsia in real-world, naturalistic settings with non-uniform exposure propensities. We developed and applied bias formulas for IPSI to gauge the impact of putative unmeasured confounding. IPSI is ideally suited to support the development of future interventions to reduce preeclampsia, while the methodology itself has far-reaching applications that could inform diverse public health priorities if widely applied to other obstetrical complications. Ethics approval No institutional review board approval was sought as the data are deidentified and publicly available. Supplementary Material Acknowledgements The authors thank the Editor, Dr Ali Khashan, and the four anonymous reviewers for their suggestions, which led to important improvements. The authors are grateful to Drs. Justin S. Brandt, Sarka Lisonkova, Ashley I. Naimi, Emily B. Rosenfeld, and Allen J. Wilcox for their thoughtful and insightful comments that greatly helped improve the manuscript. Contributor Information Wen Wei Loh, Department of Methodology and Statistics, Maastricht University, Maastricht, The Netherlands. Cande V Ananth, Division of Epidemiology and Biostatistics, Department of Obstetrics, Gynecology, and Reproductive Sciences, Rutgers Robert Wood Johnson Medical School, New Brunswick, NJ, United States; Cardiovascular Institute of New Jersey, Rutgers Robert Wood Johnson Medical School, New Brunswick, NJ, United States; Department of Medicine, Rutgers Robert Wood Johnson Medical School, New Brunswick, NJ, United States; Department of Biostatistics and Epidemiology, Rutgers School of Public Health, Piscataway, NJ, United States; Environmental and Occupational Health Sciences Institute (EOHSI), Rutgers Robert Wood Johnson Medical School, New Brunswick, NJ, United States. Author contributions Wen Wei Loh: Conceptualization, Methodology, Software, Formal analysis, Writing—Original Draft, Writing—Review & Editing, Funding acquisition, Visualization. Cande V. Ananth: Conceptualization, Data Curation, Writing—Original Draft, Writing—Review & Editing, Funding acquisition. Supplementary material Supplementary material is available at IJE online. Conflicts of interest None declared. Funding W.W.L. is partially supported by a grant from the Netherlands Organisation for Health Research and Development (ZonMw): 10430382410014. C.V.A. is partially supported by grants from the National Heart, Lung, and Blood Institute (R01-HL150065) and the National Institute of Environmental Health Sciences (R01-ES033190) of the National Institutes of Health. However, funding from these grants was not used for this project. Data availability The data are publicly available and derived from the United States live birth and fetal death records assembled by the National Center for Health Statistics of the Centers for Disease Control and Prevention. Use of artificial intelligence (AI) tools When preparing the revisions, AI-assisted tools (Grammarly and ChatGPT 5.2) were used to check spelling and grammar and to improve the readability of specific, selected statements. All text was reviewed, edited, and verified by the authors, who take full responsibility for the content of this final version. References Associated Data Supplementary Materials Data Availability Statement The data are publicly available and derived from the United States live birth and fetal death records assembled by the National Center for Health Statistics of the Centers for Disease Control and Prevention.