Per-Protocol and As-Treated Estimates for the ScreenR2GDM Trial: Further Implications

Abstract

Objective:

In a recently published pragmatic clinical trial¹ we found significantly higher incidence of diagnosed GDM but no significant differences in primary perinatal outcomes with 1-step versus 2-step gestational diabetes (GDM) screening. The published inverse probability weighted intent-to-treat (IPW-ITT) analyses provide the most robust and unbiased test of potential differences between the two methods, as they adjust for lower adherence to the 1-step method. Our objective here is to newly report the results of per-protocol (PP) and as-treated (AT) analyses to further inform future study designs, meta-analyses, and clinical practice and address questions raised by our recently published trial.

Study Design:

Details of the ScreenR2GDM pragmatic randomized clinical trial (RCT) design, including IRB approval, have been published elsewhere.^{1, 2} The trial estimated the incidence and relative risk of GDM diagnosis and perinatal outcomes among 23,792 women randomized to 1-step vs 2-step GDM screening (2014–2017), with outcomes collected through delivery (2014–2018). To account for lower adherence to the assigned screening approach among those assigned to the fasting 1-step approach,² we compared the approaches using IPW-ITT analyses. Here, we report results using comparison groups that reflect adherence to protocol and actual test received. The PP analyses compare outcomes only of individuals who adhered to random assignment. The AT analyses ignore random assignment and compare outcomes based on the actual screening approach individuals received. For each analytic approach, we estimate incidence of GDM and relative risk, with 97.5% confidence intervals (CI), of each outcome for 1-step vs 2-step screening, adjusting for pre-specified covariates and factors related to nonadherence.

Results:

PP and AT results (Table 1) are largely consistent with the primary IPW-ITT results. However, compared to the IPW-ITT analyses, the PP and AT analyses find substantially higher GDM incidence for 1-step screening and lower GDM incidence for 2-step screening. This results in a RR of GDM of ~2.5 in both of these analyses compared to 1.9 in the IPW-ITT analyses. For other primary outcomes, there was very little difference between groups in any analysis, and point estimates were similar between PP and AT results and previously published IPW-ITT results. Although PP and AT confidence intervals suggest significantly lower risk of LGA with 1-step screening, it is important to note that the confidence intervals were not adjusted for multiple comparisons.

Conclusion:

Higher incidence of GDM by 1-step vs. 2-step screening was more pronounced when analyzed by PP and AT methods compared to previously published IPW-ITT analyses, while the risk of adverse perinatal outcomes was relatively unchanged. In the presence of disparate adherence to random assignment, PP and AT methods of group classification may be more useful than IPW-ITT analyses for estimating GDM incidence (and resulting impact on clinical resources) in settings where one screening approach is used exclusively.

These results also help address some of the questions raised by respected colleagues,^{3, 4} which are important to discuss in the context of our published IPW-ITT results.¹ Our trial was designed to evaluate whether there were any differences in outcomes with the 1-step vs. 2-step screening approaches, not to determine which approach is best. Using our own preliminary population data (including data on women without GDM) and published trial data among GDM women (see published online Supplementary Appendix [SA])^{1, 5} in sample size estimates, we designed the trial to be able to detect at least a 20% difference between screening methods in any outcome. We chose this 20% threshold with the assumption that a difference in outcomes of this magnitude would be necessary for clinical stakeholders to endorse doubling the rate of GDM diagnosis and treatment; we recognize some experts may disagree.

Questions have also arisen about why we chose 97.5% CI. We intentionally did not design our study to statistically account for testing multiple primary outcomes, as each outcome is important in its own right and thus it is important to be able to interpret each outcome individually. As a compromise, we present the more conservative 97.5% CI for each primary outcome. Had we done formal testing adjusting for multiplicity, there would be no significant difference in any primary outcomes in the IPW-ITT, PP, or AT analyses, except for the doubling of GDM diagnoses by the 1-step approach. Thus, our trial results, including these PP and AT results, provide robust and consistent evidence that 1-step testing more than doubles the GDM diagnosis compared to 2-step, with only a marginal potential difference in LGA and no difference in other primary outcomes.

Our pragmatic trial was designed to evaluate 1-step and 2-step GDM screening approaches in the entire real-world population that included all pregnancies regardless of risk or timing of GDM screening, for which we achieved an excellent overall population screening rate of 94%. In response to questions of whether including early screening impacted overall trial results,⁴ there were no differences between the groups in receiving another acceptable test (HbA1c or FPG) in early pregnancy and GDM treatment protocols were consistent regardless of timing of GDM diagnosis.² Moreover, rates of non-adherence to the 1-step approach were similar among those had screening only in early in pregnancy and those who had usual (24–28 weeks’ gestation) screening.² We also clarify here that the subgroup of 165 women with an isolated FPG≥95 mg/dl by the 2-step approach who may have been treated as GDM per clinical care protocols¹ were not classified as having GDM in any IPW-ITT, PP, or AT analyses. Furthermore, as reported in the SA, sensitivity analyses reclassifying them as GDM did not change overall results.¹

We agree with colleagues^{3, 4} that following our trial participants long-term will also be important. Future studies could also evaluate randomizing GDM treatment approaches. For the present, our pragmatic population trial¹ provides data that can be integrated with foundational findings from RCTs on GDM treatment and the HAPO study^{1, 5–7} to inform clinical guidelines. Both the PP and AT incidence rates and relative risks published here also provide useful estimates for determining effect sizes and power in future research studies and may inform meta-analyses and systematic evidence reviews.

Table.

Primary Outcomes, According to One-Step or Two-Step Screening for Gestational Diabetes Using Per-protocol and As-treated Analyses^a

	Per-protocolb			As-treatedc
	1-step N=7880	2-step N=10881	Relative Risk, Adjusted for Gestational Diabetes, Prespecified Covariates, and Nonadherenced (97.5% CI)	1-step N=8161	2-step N=14128	Relative Risk, Adjusted for Gestational Diabetes, Prespecified Covariates, and Nonadherenced (97.5% CI)
	no./total no. (%)			no./total no. (%)
Gestational diabetes	1597/7880 (20.3)	864/10881 (7.9)	2.51 (2.29 – 2.74)	1678/8161 (20.6)	1104/14128 (7.8)	2.57 (2.36 – 2.79)
Large for gestational age infants	614/7349 (8.4)	937/10130 (9.2)	0.89 (0.79 – 0.99)	644/7619 (8.5)	1235/13158 (9.4)	0.89 (0.80 – 0.99)
Perinatal composite outcomee	216/7481 (2.9)	284/10304 (2.8)	1.11 (0.90 – 1.36)	223/7755 (2.9)	378/13397 (2.8)	1.08 (0.89 – 1.31)
Gestational hypertension or preeclampsia	1031/7251 (14.2)	1380/9998 (13.8)	0.99 (0.91 – 1.08)	1069/7502 (14.2)	1754/12980 (13.5)	1.00 (0.92 – 1.08)
Primary cesarean section	1918/7803 (24.6)	2664/10770 (24.7)	0.97 (0.91 – 1.03)	1983/8081 (24.5)	3389/13978 (24.2)	f

^aThe denominators vary according to the ascertainment method and exclusion criteria for each outcome. The maternal outcome of gestational hypertension or preeclampsia excluded women with preexisting hypertension before pregnancy. Primary cesarean section excluded women who left the health plan before delivery. The perinatal composite outcome included pregnancies for which information was available in the maternal record (stillbirth and shoulder dystocia) or in newborn records that were matched to maternal records. Reasons for unmatched newborn records include adoption, deliveries within and outside the health plan in which the newborn was covered by other insurance, deliveries outside the health plan for which no reimbursement for newborn care was requested, and instances in which the mother left the health plan before delivery and no information was available for the newborn.

^bPer-protocol - compares pregnancies randomized to 1-step who received 1-step vs. those randomized to 2-step who received 2-step.

^cAs-treated – compares pregnancies based on actual test received (1-step or 2-step), regardless of randomized assignment. If there was more than one GDM diagnostic test performed during pregnancy, then pregnancies were categorized by the last GDM screening test received.

^dPrespecified covariates include race/ethnicity, prepregnancy obesity, and weight gain exceeding National Academy of Medicine weight gain guidelines.⁸ Factors related to nonadherence include maternal age, nulliparity, race/ethnicity, Medicaid insurance, previous gestational diabetes, preexisting hypertension, trial site, maternal obesity at first prenatal visit, provider type, and randomized group. The widths of confidence intervals have not been adjusted to account for multiplicity and cannot be used to infer treatment effects.

^eThe perinatal composite consisted of any of the following: stillbirth, neonatal death, shoulder dystocia, bone fracture, or any arm or hand nerve palsy related to birth injury.⁵

^fAs-treated analysis for primary cesarean section is stratified by GDM due to significant interaction term of GDM by As-treated group: RR was 0.89 (97.5% CI 0.78–1.02) for pregnancies with GDM and 0.99 (97.5% CI 0.94–1.06) for pregnancies without GDM.