Investigating the Shape and Strength of the Relationship Between Maternal Weight Gain and Gestational Age at Delivery in Twin and Singleton Pregnancies

Abstract

Both inadequate and excessive maternal weight gain are correlated with preterm delivery in singleton pregnancies, yet this relationship has not been adequately studied in twins. We investigated the relationship between time-varying maternal weight gain and gestational age at delivery in twin pregnancies and compared it with that in singletons delivered in the same study population. We used serial weight measurements abstracted from charts for twin and singleton pregnancies delivered during 1998–2013 in Pittsburgh, Pennsylvania. Our exposure was time-varying weight gain z score, calculated using gestational age–standardized and prepregnancy body mass index–stratified twin- and singleton-specific charts, and our outcome was gestational age at delivery. Our analyses used a flexible extension of the Cox proportional hazards model that allowed for nonlinear and time-dependent effects. We found a U-shaped relationship between weight gain z score and gestational age at delivery among twin pregnancies (lowest hazard of delivery observed at z score = 1.2), which we attributed to increased hazard of early preterm spontaneous delivery among pregnancies with low weight gain and increased hazard of late preterm delivery without labor among pregnancies with high weight gain. Our findings may be useful for updating provisional guidelines for maternal weight gain in twin pregnancies.

Abbreviations

BMI

body mass index

GA

gestational age

GWG

gestational weight gain

Maternal weight gain during pregnancy is a potentially modifiable characteristic (1) that is correlated with maternal and child health (2–4). Previous studies in singleton pregnancies indicate that both low and high weight gain are associated with increased risk of preterm birth (5–7). Twin pregnancies have increased in recent years (8, 9), and mothers experience more weight gain (10) and higher incidence of poor perinatal outcomes, including preterm birth (8, 11). Despite this, the contribution of weight gain to perinatal health outcomes among twin pregnancies, and its relationship to that in singletons, has not been adequately studied (12, 13).

Several studies have aimed to quantify the effect of weight gain on preterm birth in singletons; however, many have failed to account for the fact that longer gestational duration generally yields more total weight gain (14–16). Specifically, in pregnancies with a shorter gestational age at delivery, there is less opportunity to gain weight, creating a potentially spurious association between low weight gain and preterm birth. For this reason, many studies of cumulative measures of weight gain and preterm birth produce biased estimates and are therefore difficult to interpret (17, 18).

Studies that account for the connection between total weight gain and gestational duration describe a modest U-shaped relationship between weight gain and preterm birth among singleton pregnancies (5–7). A study among singleton pregnancies found that this relationship may be attenuated when using z-score charts compared with cumulative weight gain measures (i.e., total, rate, and adequacy relative to Institute of Medicine guidelines), likely because z-score charts are designed to characterize weight gain independent of gestational duration (6). A study of twin pregnancies within our study population described a slight U-shaped relationship between total weight gain z score and delivery before 32 weeks’ gestational age only among those with underweight, normal weight, overweight, and obesity class I (but not classes II or III) prepregnancy body mass index (19). Notably, this study was only able to examine total weight gain z score (i.e., lacked measurements of weight gain throughout pregnancy) and could not stratify by type of delivery.

Recently, researchers have recommended investigating preterm delivery, and more broadly perinatal outcomes that are related to gestational age at delivery, within a time-to-event framework (20, 21). Instead of examining preterm delivery as a binary outcome, time-to-event analyses capture the continuum of gestational duration—from extreme, severe, and moderate prematurity (11) to late preterm, early to late term, and post term—and thus account more precisely for gestational age at delivery (22). Additionally, time-to-event analyses easily incorporate time-varying exposures—namely, approximate weight gain per day or week of gestation—as well as competing risks, such as deliveries following spontaneous versus induced onset of labor. Using time-varying measures of both weight gain and gestational age at delivery may increase precision and relevance of the estimated associations (20).

We aimed to investigate the relationship between gestational age–specific weight gain and time to delivery in twin pregnancies and compare it with that in singletons. Since twins are less common than singletons, and are typically delivered at earlier gestational ages, we were especially interested in similarities in the shape of the relationship between weight gain and time to delivery, as well as its strength across gestational age. We hypothesized that twin and singleton pregnancies might differ regarding the range of weight gain associated with lowest risk of delivery (likely higher in twins) as well as the range of gestational ages at which the association is the strongest (likely earlier in twins).

METHODS

Study population

We analyzed data assembled for a cohort of diamniotic twin pregnancies and case-cohort of singleton pregnancies delivered during 1998–2013 at Magee-Women’s Hospital in Pittsburgh, Pennsylvania. Parent studies selected all diamniotic twin pregnancies delivered during 1998–2013 and a random sample of singleton pregnancies within 6 strata of prepregnancy body mass index (BMI; underweight, normal weight, overweight, obesity class I, obesity class II, obesity class III), as well as a random sample of singleton pregnancies with preterm delivery within the same prepregnancy BMI strata delivered during 1998–2011. Serial weights were measured and abstracted from clinical charts, while self-reported prepregnancy/delivery weights and maternal/pregnancy characteristics were obtained from the Magee Obstetric Maternal and Infant Database and supplemented with vital statistics information collected by the state of Pennsylvania. Details of both the twin cohort and the singleton case-cohort have been previously described (23, 24).

We excluded the second twin pregnancy among mothers that had 2 twin pregnancies during the study period, as well as all twin pregnancies with monochorionic placentation; this condition (in which there is only one placenta) increases the risk profile of these pregnancies (25), which may confer less importance to weight gain as a monitored characteristic of pregnancy. We additionally excluded any pregnancies with implausible maternal prepregnancy weight as identified by data managers using a conditional percentile method (details in Web Appendix 1, available at https://doi.org/10.1093/aje/kwad105) (26); implausible gestational age at delivery (defined as >43 weeks); gestational age at delivery (and/or fetal death among twin pregnancies) ≤20 weeks; missing maternal prepregnancy weight, height, delivery weight, or gestational age at delivery; no available serial weight measurements; presence of any weight gain values that could not be converted to gestational age–standardized weight gain z scores using twin- and singleton-specific charts (i.e., weight gain less than or equal to the constant added prior to log-transformation in the authors’ formulas) (23, 27, 28); or missing covariates of interest.

Exposure

Our primary exposure of interest was time-varying weight gain during pregnancy, which we calculated in kilograms as the difference between pregnancy weight (at a prenatal visit or at delivery) and the self-reported prepregnancy weight. For gestational ages at which weight was not measured (i.e., days between prenatal visits), we estimated weight gain using an interpolation method that we previously found to most accurately and precisely estimate weight gain between measurements (29). Specifically, we first pooled all pregnancies within pregnancy plurality and then used regression with restricted cubic splines for gestational age, knots at trimesters, and random intercepts and slopes for pregnancy to estimate weight gain between prenatal measurements.

We then converted natural log-transformed weight gain values to gestational age–standardized z scores using formulas derived from twin- and singleton-specific charts stratified by prepregnancy BMI (details in Web Appendix 2 and Web Table 1 for twins and Web Appendix 3 and Web Table 2 for singletons; natural log-transformation of weight gain values was required for conversion to z scores using the authors’ formulas) (23, 27, 28). We standardized weight gain to account for the inherent correlation between the actual amount of weight gain in kilograms and the gestational age by which it occurred. Additionally, a key assumption of our statistical model is that the relationship between exposure and outcome has a constant shape (i.e., lowest hazard of delivery associated with a similar value of exposure across gestational age), although the strength can vary over the time scale (30). We anticipated that time-varying actual weight gain in kilograms might violate this assumption; for example, a weight gain of 10 kg accumulated by the second trimester may have different implications for gestational age at delivery than a similar amount accumulated by the third trimester. Conversely, we expected that a gestational age–standardized weight gain z score at or near zero might consistently be associated with lowest risk for delivery across gestational age, with the exception of at or after term, where weight gain z score at or near zero may be associated with optimally timed delivery. Since twin- and singleton-specific charts were developed within prepregnancy BMI categories, we anticipated that effect measure modification by prepregnancy BMI, which has been observed in studies of weight gain and preterm birth, might be mitigated using this strategy. We applied z-score formulas developed for individuals with normal weights to those with underweight prepregnancy BMI, since formulas were not published for this prepregnancy BMI stratum.

Outcome

Our primary outcome was gestational age at delivery in days. In the study population, last menstrual period (used to calculate gestational age at delivery) was estimated per American College of Obstetricians and Gynecologists’ guidelines (31), which incorporated self-reported last menstrual period and ultrasound where available. We considered 20 weeks’ gestational age as time zero; this is generally the minimum gestational age at which pregnancies are systematically captured in administrative databases. First, we defined an event as any type of delivery. Second, in 3 separate analyses, we limited the event to one of: 1) delivery following spontaneous onset of labor, 2) delivery following induced labor, or 3) delivery without labor (i.e., cesarean delivery before labor). Specifically, we assumed ongoing pregnancies were at risk for any delivery, while occurrence of one type precluded risk for the other 2 types of delivery (i.e., competing events). In these models, we defined the event as the type of delivery of interest (for example, spontaneous onset of labor) and right-censored pregnancies at occurrence of the other types of delivery (for example, induced deliveries or deliveries without labor); we repeated this process for each type of delivery. We right-censored all twin models at 37 weeks and all singleton models at 39 weeks, since these are the gestational ages at which delivery starts to be considered as a desirable outcome of pregnancy (32, 33).

Statistical analysis

We applied a time-to-event framework to investigate the relationship between time-varying weight gain and gestational age at delivery. To avoid the restrictive assumptions of conventional Cox proportional hazards models (34), we employed a flexible extension (30) that allowed us to incorporate both the nonlinear and time-dependent effects of weight gain on the log-hazard of time to delivery. This model simultaneously estimated how the log-hazard changed across weight gain z score (nonlinear effect) and how the strength of this relationship varied across gestational age (time-dependent effect). In particular, we could discern whether the relationship between weight gain and log-hazard of delivery was U-shaped, as previously described in singletons (5–7), and whether it was more prominent at certain gestational ages. Functions for both nonlinear and time-dependent effects incorporated quadratic splines and 1 internal knot as recommended by developers of this method (30). We generated bootstrapped 95% confidence intervals, with 200 repetitions, using the 2.5th and 97.5th percentiles of the product of nonlinear and time-dependent functions.

We fitted the following model, where λ(GA|GWG_GA, C_i) is the hazard of delivery at a given gestational age (GA) conditional on current weight gain z score(GWG_GA) and covariates (C_i); λ₀(GA) is the baseline hazard of delivery among pregnancies with a weight gain z score of 0 at a given gestational age and referent values for all covariates; β(GA) is the estimated time-dependent function that describes how the strength of the relationship between current weight gain z score and log-hazard of delivery varied over gestational age; r(GWG_GA) is the estimated nonlinear function that describes how log-hazard of delivery at a given gestational age changed with changing current weight gain z score; and C_i is a vector of covariate values, while α_i is a vector of corresponding log-hazard ratios:

We adjusted all multivariate models for a predefined list of covariates (details in Web Appendix 4). Briefly, we selected prepregnancy characteristics that are commonly correlated with perinatal health (e.g., measures of maternal health and socioeconomic status) in part on the basis of data availability. Twin and singleton pregnancies were analyzed in separate statistical models, and sampling weights were incorporated in singleton models.

Sensitivity analysis

We fitted models with time-varying weight gain in kilograms (rather than z scores) as the exposure; models with or without nonlinear and/or time-dependent effects; models with different numbers of knots and/or degrees of splines; models incorporating adjustment and/or effect measure modification of prepregnancy BMI; models that did not right-censor gestational age at delivery at 37 weeks (for twins) and 39 weeks (for singletons); and singleton models that did not incorporate sampling weights (Web Table 3). We assessed whether these modifications substantially improved model fit by comparing the Akaike information criterion, which represents model deviance adjusted for number of parameters (35), and Bayesian information criterion (36), which represents model deviance adjusted for number of parameters and events. For both Akaike information criterion and Bayesian information criterion, a lower value indicated better model fit; differences of more than 4 are typically interpreted as meaningful, whereas differences of more than 10 provide strong indication of improved fit (37).

Analyses were performed with Stata (version 14.2; StataCorp LP, College Station, Texas) (38), R (version 3.5.1; R Foundation for Statistical Computing, Vienna, Austria) (39), and using the customized CoxFlex R program for estimating Cox models with nonlinear and time-dependent effects (see data availability statement for access to this program). Ethics approval was obtained from University of Pittsburgh and University of British Columbia for the parent studies and from McGill University Faculty of Medicine Institutional Review Board for the present study.

RESULTS

Sample selection

Our analytical cohort included 2,021 twin and 9,114 singleton pregnancies (Figures 1 and 2). Among twins, we excluded 243 pregnancies with missing or implausible prepregnancy weight, height, delivery weight, gestational age, and/or serial weight measurements; 9 pregnancies with gestational age at delivery and/or fetal death on or prior to 20 weeks’ gestational age; 1 pregnancy with at least 1 serial weight (either measured or imputed) on or after 20 weeks’ gestational age that could not be converted to z score using the twin-specific chart; and 35 pregnancies with missing covariates. Overall, we retained greater than 85% of the original dichorionic twin cohort.

Figure 1.

Sample selection among twin pregnancies delivered at Magee-Women’s Hospital, Pittsburgh, Pennsylvania, 1998–2013.

Figure 2.

Sample selection among singleton pregnancies delivered at Magee-Women’s Hospital, Pittsburgh, Pennsylvania, 1998–2011.

Sample characteristics

Twin pregnancies.

Among twin pregnancies, individuals in the lowest total weight gain z-score quartile were less frequently college graduates, married, and nulliparous; more frequently within non-Hispanic Black or other race categories, insured by Medicaid, and ever-smokers; and more likely to experience spontaneous delivery and stillbirth (Table 1). Conversely, individuals in the highest total gestational weight gain z-score quartile were more frequently nulliparous and classified as having preexisting hypertension and diabetes. Deliveries following spontaneous onset of labor were most frequent (56.9%), while deliveries without labor or following induced labor were less common (26.7% and 16.5%, respectively). Median total weight gain ranged from 9.5 kg in the lowest to 24.9 kg in the highest quartile, while median and interquartile range of gestational age at delivery appeared similar across quartiles.

Table 1.

Sample Characteristics by Total Gestational Weight Gain Z Score Among Twin Pregnancies Delivered at Magee-Women’s Hospital^a, Pittsburgh, Pennsylvania, 1998–2013

	Total Weight Gain Z-Score Quartile								All(n = 2021)
	Quartile 1(n = 506)		Quartile 2(n = 505)		Quartile 3(n = 505)		Quartile 4(n = 505)
Characteristic	No	%	No	%	No	%	No	%	No	%
Total weight gain z-score rangeb	−9.3; −0.8		−0.8; −0.2		−0.2; 0.4		0.4; 3.0		−9.3; 3.0
Maternal age, yearsc	30.4 (6.5)		30.6 (5.7)		30.8 (5.6)		30.2 (5.8)		30.5 (5.9)
Maternal education
Less than high school	39	7.7	24	4.8	21	4.2	28	5.5	112	5.5
High school/GED	117	23.1	102	20.2	83	16.4	108	21.4	410	20.3
Some college/associates	116	22.9	116	23.0	95	18.8	114	22.6	441	21.8
College graduate	234	46.2	263	52.1	306	60.6	255	50.5	1,058	52.4
Race/ethnicity
Non-Hispanic White	360	71.1	394	78.0	430	85.1	414	82.0	1,598	79.1
Non-Hispanic Black	117	23.1	91	18.0	59	11.7	79	15.6	346	17.1
Hispanicd	3	0.6	6	1.2	2	0.4	6	1.2	17	0.8
Otherc	26	5.1	14	2.8	14	2.8	6	1.2	60	3.0
Married
Yes	328	64.8	347	68.7	382	75.6	355	70.3	1,412	69.9
No	178	35.2	158	31.3	123	24.4	150	29.7	609	30.1
Insurance
Private/other	296	58.5	326	64.6	336	66.5	318	63.0	1,276	63.1
Medicaid/self-pay	210	41.5	179	35.4	169	33.5	187	37.0	745	36.9
Parity
Nulliparous	188	37.2	211	41.8	259	51.3	282	55.8	940	46.5
Primiparous	183	36.2	166	32.9	145	28.7	133	26.3	627	31.0
Multiparous	135	26.7	128	25.3	101	20.0	90	17.8	454	22.5
Prepregnancy BMI categorye
Underweight	15	3.0	21	4.2	21	4.2	7	1.4	64	3.2
Normal weight	244	48.2	225	44.6	232	45.9	259	51.3	960	47.5
Overweight	130	25.7	130	25.7	138	27.3	119	23.6	517	25.6
Obesity	117	23.1	129	25.5	114	22.6	120	23.8	480	23.8
Preexisting diabetes
Yes	14	2.8	9	1.8	14	2.8	20	4.0	57	2.8
No	492	97.2	496	98.2	491	97.2	485	96.0	1,964	97.2
Preexisting hypertension
Yes	29	5.7	32	6.3	21	4.2	31	6.1	113	5.6
No	477	94.3	473	93.7	484	95.8	474	93.9	1908	94.4
Preexisting polycystic ovarian syndrome
Yes	15	3.0	15	3.0	14	2.8	15	3.0	59	2.9
No	491	97.0	490	97.0	491	97.2	490	97.0	1,962	97.1
Ever smoker
Yes	90	17.8	70	13.9	55	10.9	70	13.9	285	14.1
No	416	82.2	435	86.1	450	89.1	435	86.1	1,736	85.9
Interpregnancy interval
<18 months	188	37.2	211	41.8	259	51.3	282	55.8	940	46.5
≥18 months	60	11.9	55	10.9	47	9.3	31	6.1	193	9.5
Nulliparous	130	25.7	130	25.7	91	18.0	92	18.2	443	21.9
Missing	128	25.3	109	21.6	108	21.4	100	19.8	445	22.0
Assisted reproductive technology
Yes	140	27.7	139	27.5	176	34.9	148	29.3	603	29.8
No	366	72.3	366	72.5	329	65.1	357	70.7	1,418	70.2
Number of unique weight measurementsf,g	12.0 (9.0; 14.0)		12.0 (10.0; 15.0)		12.0 (10.0; 14.0)		12.0 (10.0; 14.0)		12.0 (10.0; 14.0)
Gestational age at delivery, weeksf	36.4 (33.9; 38.0)		36.6 (34.0; 38.0)		36.3 (34.0; 37.7)		36.0 (34.0; 37.3)		36.3 (34.0; 37.9)
Ultrasound
Yes	506	100.0	504	99.8	505	100.0	505	100.0	2,020	100.0
No	0	0.0	1	0.2	0	0.0	0	0.0	1	0.0
Gestational age at first ultrasound, weeksf	12.3 (8.4; 18.0)		11.7 (8.0; 18.1)		11.9 (8.0; 18.2)		12.3 (7.9; 18.4)		12.0 (8.1; 18.3)
Missing	24	4.7	26	5.1	24	4.8	25	5.0	99	4.9
Type of delivery
Spontaneous	314	62.1	278	55.0	285	56.4	272	53.9	1,149	56.9
Induced	77	15.2	94	18.6	69	13.7	93	18.4	333	16.5
No labor	115	22.7	133	26.3	151	29.9	140	27.7	539	26.7
Delivery route (twin A)
Vaginal	248	49.0	234	46.3	219	43.4	207	41.0	908	44.9
Cesarean delivery	258	51.0	271	53.7	286	56.6	298	59.0	1,113	55.1
Delivery route (twin B)
Vaginal	218	43.1	218	43.2	197	39.0	186	36.8	819	40.5
Cesarean delivery	288	56.9	287	56.8	308	61.0	319	63.2	1,202	59.5
Stillbirth (twin A)
Yes	6	1.2	2	0.4	0	0.0	1	0.2	9	0.4
No	500	98.8	503	99.6	505	100.0	504	99.8	2,012	99.6
Stillbirth (twin B)
Yes	7	1.4	2	0.4	3	0.6	4	0.8	16	0.8
No	499	98.6	503	99.6	502	99.4	501	99.2	2,005	99.2
Total gestational weight gain, kgf	9.5 (5.9; 12.7)		15.4 (12.7; 17.7)		19.5 (16.8; 21.3)		24.9 (22.7; 28.1)		16.8 (12.2; 21.8)
Total gestational weight gain, z scoref	−1.3 (−1.7; −1.0)		−0.5 (−0.6; −0.3)		0.1 (−0.1; 0.2)		0.8 (0.6; 1.1)		−0.2 (−0.8; 0.4)

Abbreviations: BMI, body mass index; GED, General Educational Development.

^a Covariates are at the level of the mother/pregnancy unless otherwise stated.

^b Values are expressed as minimum; maximum.

^c Values are expressed as mean (standard error).

^d Categories combined in regression analyses due to paucity of pregnancies in the Hispanic category.

^e Weight (kg)/height (m)².

^f Values are expressed as median (25th percentile; 75th percentile).

^g Number of unique weight measurements includes the prepregnancy weight, prenatal weights, and the delivery weight for each participant.

Singleton pregnancies.

Among singleton pregnancies, spontaneous and induced deliveries were most frequent (44.2% and 44.3%, respectively), while deliveries without labor were less common (11.5%; Table 2). Median total weight gain ranged from 9.5 kg to 18.1 kg in the lowest and highest quartiles, respectively, while median gestational age at delivery was slightly lower in the first and fourth quartiles (38.1 weeks’ and 37.4 weeks’ gestational age, respectively) than in the second and third quartiles (38.5 weeks’ and 38.4 weeks’ gestational age, respectively).

Table 2.

Unweighted Sample Characteristics by Total Gestational Weight Gain Z Score Among Singleton Pregnancies Delivered at Magee-Women’s Hospital, Pittsburgh, Pennsylvania, 1998–2011

	Total Weight Gain Z-Score Quartile								All(n = 9,114)
	Quartile 1 (n = 2,284)		Quartile 2 (n = 2,273)		Quartile 3 (n = 2,279)		Quartile 4 (n = 2,278)
Characteristic	No	%	No	%	No	%	No	%	No	%
Total weight gain z-score rangea	−9.1; −0.3		−0.3; 0.5		0.5; 1.2		1.2; 3.9		−9.1; 3.9
Maternal age, yearsb	28.4 (6.2)		29.2 (6.0)		29.2 (5.9)		28.8 (6.0)		28.9 (6.0)
Maternal education
Less than high school	267	11.7	185	8.1	169	7.4	194	8.5	815	8.9
High school/GED	592	25.9	551	24.2	581	25.5	640	28.1	2,364	25.9
Some college/associates	546	23.9	560	24.6	657	28.8	667	29.3	2,430	26.7
College graduate	879	38.5	977	43.0	872	38.3	777	34.1	3,505	38.5
Race/ethnicity
Non-Hispanic White	1,661	72.7	1,718	75.6	1,699	74.6	1,649	72.4	6,727	73.8
Non-Hispanic Black	462	20.2	432	19.0	499	21.9	588	25.8	1,981	21.7
Hispanicc	25	1.1	22	1.0	21	0.9	18	0.8	86	0.9
Otherc	136	6.0	101	4.4	60	2.6	23	1.0	320	3.5
Married
Yes	1,308	57.3	1,398	61.5	1,350	59.2	1,289	56.6	5,345	58.6
No	976	42.7	875	38.5	929	40.8	989	43.4	3,769	41.4
Insurance
Private/other	1,206	52.8	1,336	58.8	1,306	57.3	1,215	53.3	5,063	55.6
Medicaid/self-pay	1,078	47.2	937	41.2	973	42.7	1,063	46.7	4,051	44.4
Parity
Nulliparous	918	40.2	999	44.0	968	42.5	1,123	49.3	4,008	44.0
Primiparous	819	35.9	744	32.7	782	34.3	669	29.4	3,014	33.1
Multiparous	547	23.9	530	23.3	529	23.2	486	21.3	2,092	23.0
Prepregnancy BMI categoryd
Underweight	642	28.1	409	18.0	178	7.8	107	4.7	1,336	14.7
Normal weight	858	37.6	544	23.9	277	12.2	173	7.6	1,852	20.3
Overweight	329	14.4	437	19.2	428	18.8	455	20.0	1,649	18.1
Obesity, class I	187	8.2	334	14.7	436	19.1	589	25.9	1,546	17.0
Obesity, class II	178	7.8	314	13.8	494	21.7	427	18.7	1,413	15.5
Obesity, class III	90	3.9	235	10.3	466	20.4	527	23.1	1,318	14.5
Preexisting diabetes
Yes	23	1.0	42	1.8	67	2.9	132	5.8	264	2.9
No	2,261	99.0	2,231	98.2	2,212	97.1	2,146	94.2	8,850	97.1
Preexisting hypertension
Yes	68	3.0	133	5.9	174	7.6	244	10.7	619	6.8
No	2,216	97.0	2,140	94.1	2,105	92.4	2,034	89.3	8,495	93.2
Ever smoker
Yes	558	24.4	429	18.9	390	17.1	406	17.8	1,783	19.6
No	1,726	75.6	1,844	81.1	1,889	82.9	1,872	82.2	7,331	80.4
Number of unique weight measurementse,f	12.0 (9.0; 14.0)		12.0 (10.0; 14.0)		12.0 (10.0; 14.0)		12.0 (10.0; 14.0)		12.0 (10.0; 14.0)
Gestational age at delivery, weekse	38.1 (37.1; 40.1)		38.5 (38.0; 40.3)		38.4 (37.7; 40.1)		37.4 (36.1; 38.4)		38.1 (37.0; 40.0)
Ultrasound
Yes	2,248	98.4	2,245	98.8	2,246	98.6	2,262	99.3	9,001	98.8
No	36	1.6	28	1.2	33	1.4	16	0.7	113	1.2
Gestational age at first ultrasound, weekse	15.9 (9.9;19.1)		16.0 (9.6;19.1)		14.4 (9.1;19.1)		13.6 (9.0;19.0)		15.0 (9.4;19.0)
Missing	176	7.7	182	8.0	156	6.8	157	6.9	671	7.4
Type of delivery
Spontaneous	1,083	47.4	1,004	44.2	915	40.1	1,028	45.1	4,030	44.2
Induced	1,036	45.4	1,048	46.1	1,032	45.3	919	40.3	4,035	44.3
No labor	165	7.2	221	9.7	332	14.6	331	14.5	1,049	11.5
Delivery route
Vaginal	1,815	79.5	1,671	73.5	1,507	66.1	1,449	63.6	6,442	70.7
Cesarean delivery	469	20.5	602	26.5	772	33.9	829	36.4	2,672	29.3
Stillbirth
Yes	7	0.3	7	0.3	8	0.4	12	0.5	34	0.4
No	2,277	99.7	2,266	99.7	2,271	99.6	2,266	99.5	9,080	99.6
Total gestational weight gain, kge	9.5 (4.5; 11.8)		13.2 (5.9; 15.9)		12.2 (8.2; 18.1)		18.1 (14.5; 23.1)		12.7 (8.2; 16.8)
Total gestational weight gain, z scoree	−0.8 (−1.3; −0.5)		0.1 (−0.1; 0.3)		0.8 (0.7; 1.0)		1.6 (1.4; 2.0)		0.5 (−0.3; 1.2)

Abbreviations: BMI, body mass index; GED, General Educational Development.

^a Values are expressed as minimum; maximum.

^b Values are expressed as mean (standard error).

^c Categories combined in regression analyses due to paucity of pregnancies in Hispanic category.

^d Weight (kg)/height (m)².

^e Values are expressed as median (25th percentile; 75th percentile).

^f Number of unique weight measurements includes the prepregnancy weight, prenatal weights, and the delivery weight for each participant.

Weight gain and time to delivery

Twin pregnancies.

Web Figure 1 displays relationships between time-varying weight gain z score and hazard of all types of delivery (row 1) and deliveries following spontaneous onset of labor, induced labor, and without labor (rows 2–4, respectively) at specific weeks of gestational age (28, 30, 32, 34, and 36 weeks) for twin pregnancies. In twins, when all types of delivery were combined, we observed a time-dependent relationship between weight gain and hazard of delivery. Specifically, hazard of delivery appeared to be increased among pregnancies with low weight gain at earlier gestational ages (28–34 weeks) and among pregnancies with high weight gain at later gestational ages (36 weeks; Web Figure 1, row 1), although small proportions of the confidence intervals for each relationship crossed the null value (hazard ratio of 1). From 21–35 weeks’ gestational age, the lowest hazard of any type of delivery was associated with a weight gain z score of 1.2 (or 1.2 standard deviations greater than natural log-transformed mean weight gain prescribed by twin charts), which corresponds to approximately 22 kg at 30 weeks or 26–27 kg at 35 weeks among individuals with prepregnancy BMI in the underweight, normal weight, or overweight categories, and 20 kg at 30 weeks or 25 kg at 35 weeks among individuals with prepregnancy BMI classified as obesity.

When only deliveries following spontaneous onset of labor were considered, and those following induced labor/without labor were censored at delivery, we observed an increased hazard of delivery at 28–36 weeks among individuals with low weight gain, which gradually attenuated as gestational age increased (Figure 3). Hazard of spontaneous delivery increased monotonically as weight gain z score decreased from 1.3; this corresponds to weight gain at 28–36 weeks of 21–28 kg among individuals with underweight/normal weight BMI, 21–30 kg among individuals with overweight BMI, and 20–28 kg among individuals with obese BMI, respectively.

Figure 3.

The relationship between time-varying weight gain z score and log-hazard of spontaneous delivery among twin pregnancies delivered at Magee-Women’s Hospital, Pittsburgh, Pennsylvania, 1998–2013. All panels are derived from the same regression model, which simultaneously incorporated nonlinear and time-dependent associations when modeling the relationship of weight gain z score and log-hazard of spontaneous delivery; each panel is a cross-section of that relationship at specified gestational age (GA): A) 28 weeks; B) 30 weeks; C) 32 weeks; D) 34 weeks; E) 36 weeks. CI, confidence interval; GWG, gestational weight gain; HR, hazard ratio.

When only deliveries following induction of labor were considered, we observed a slightly increased relative hazard of delivery with increased weight gain z score at 36 weeks’ gestational age, but large portions of these confidence intervals were on either side of the null value throughout ranges of weight gain z score (Web Figure 1, row 3). When only deliveries without labor were considered events, we observed a monotonically increased relative hazard of delivery with increased weight gain z score only at later gestational ages (34–36 weeks; Web Figure 1, row 4). The relationship between weight gain z score and hazard of delivery for induced deliveries from 28–32 weeks and deliveries without labor at 28 weeks could not be reliably estimated due to the paucity of events at these gestational ages in twin pregnancies.

Singleton pregnancies.

Similar to twins, low weight gain appeared associated with increased hazard of spontaneous delivery at 28–34 weeks’ gestational age among singleton pregnancies (Figure 4). In contrast to twins, higher weight gain z scores also appeared associated with increased hazard of late spontaneous preterm delivery (i.e., at 36 weeks’ gestational age). From 20–38 weeks, minimum hazard of spontaneous delivery was observed at a z score of 0.3 (or 0.3 standard deviations greater than the natural log-transformed mean weight gain prescribed by singleton charts), which corresponds to approximately 16–17 kg of weight gain at 36–38 weeks’ gestational age among individuals with normal weight prepregnancy BMI. Web Figure 2 displays relationships between time-varying weight gain z score and log-hazard of all types of delivery, spontaneous delivery, induced delivery, and no-labor delivery, among singletons.

Figure 4.

The relationship between time-varying gestational weight gain (GWG) z score and log-hazard of spontaneous delivery among singleton pregnancies delivered at Magee-Women’s Hospital, Pittsburgh, Pennsylvania, 1998–2011. All panels are derived from the same regression model, which simultaneously incorporated nonlinear and time-dependent associations when modeling the relationship of weight gain z score and log-hazard of spontaneous delivery; each panel is a cross-section of that relationship at specified gestational age (GA): A) 28 weeks; B) 30 weeks; C) 32 weeks; D) 34 weeks; E) 36 weeks; F) 38 weeks. CI, confidence interval; HR, hazard ratio.

Sensitivity analysis

Among twins, model fit marginally improved when nonlinear and time-dependent effects were not incorporated within models that considered all types of delivery as events (Web Table 4, models 5, 6, and 7). This may have resulted from simultaneously considering deliveries following spontaneous onset of labor (i.e., higher hazard at delivery only at lower weight gain z scores) and without labor (i.e., higher hazard of delivery only at higher weight gain z scores) as events within the same model instead of as competing events. Among twins, model fit did not improve when prepregnancy BMI effect measure modification was incorporated within models that examined weight gain z scores as the exposure of interest. Among both twins and singletons, model fit (as assessed by Akaike information criterion in twins and both Akaike information criterion and Bayesian information criterion in singletons) improved with the addition of covariates that may confound the relationship between weight gain and gestational age of delivery (Web Tables 4 and 5, model 17).

Among singletons, model fit improved when weight gain in kilograms (rather than z score) was investigated as the exposure (Web Table 5, model 4), and when more knots and splines were incorporated (Web Table 5, models 8, 9, and 10); however, inferences were similar with respect to the U-shaped exposure-outcome relationship across most common values of weight gain z score. Additionally, among singletons, model fit improved for models that incorporated BMI as a continuous/categorical variable and prepregnancy BMI effect measure modification of the exposure-outcome relationship (Web Table 5, models 12, 13, 14, and 15). Specifically, lower weight gain z score was associated with even higher hazard of delivery among individuals with underweight prepregnancy BMI, and higher weight gain z score was associated with even higher hazard of delivery among individuals with overweight prepregnancy BMI.

DISCUSSION

We found a strong association between lower maternal weight gain and preterm delivery and a modest association between higher maternal weight gain and preterm delivery in twin pregnancies, which mimics the U-shaped relationship commonly reported in singletons (5–7). Our findings suggest 2 separate phenomena : 1) increased hazard of early spontaneous onset preterm delivery among twin pregnancies with relatively low weight gain and; 2) increased hazard of late preterm delivery without labor among those with relatively high weight gain. This contrasted slightly with our findings among singletons, which suggested an increased hazard of spontaneous onset preterm delivery among pregnancies with both relatively low and high weight gain. Additionally, we found that strengths of these associations varied across gestational age. Thus, we find evidence for both nonlinear and time-dependent effects of weight gain z score on gestational age at delivery among twins.

Our findings are consistent with and expand upon prior research in singletons, and strengthen the evidence base for weight gain guidelines in twin pregnancies. A previous meta-analysis that examined the combined roles of prepregnancy BMI and total weight gain, quantified as a z score using an internal reference, among singleton pregnancies found that low and high weight gains across prepregnancy BMI strata were associated with 1.2- to 1.8-fold and 1.2- to 2.9-fold increased odds of preterm birth, respectively, relative to pregnancies with normal weight BMI and average weight gain (5). When separately examining spontaneous and induced deliveries, one study found an increased risk of spontaneous delivery among singleton pregnancies with low weight gain and increased risk of induced delivery among singleton pregnancies with high weight gain (2). Our findings support similar relationships among twin pregnancies while also considering types of delivery combined together as well as separately as competing events. Furthermore, we identified potential critical periods for effects of weight gain; specifically, relatively low weight gain confers higher hazard of spontaneous delivery at 28–34 weeks’ gestational age, while relatively high weight gain presents higher risk of delivery without labor only at 34–36 weeks’ gestational age. These findings may have important implications for the clinical monitoring of weight gain in both twin and singleton pregnancies. While weight gain is clinically monitored as a summary measure, we acknowledge that the extent to which modifying weight gain through diet and exercise affects gestational age at delivery is unclear (1, 40), especially among twin pregnancies.

Our results suggest that, among singleton pregnancies of individuals with normal weight, those that gain 16–17 kg by 36–38 weeks’ gestational age experience the lowest hazard of early delivery (i.e., before 39 weeks’ gestational age); this is within (but at the higher end of) the range of 11.5–16 kg total weight gain recommended by current Institute of Medicine guidelines (10). Conversely, our results suggest that, among twin pregnancies of individuals with normal weight, weight gain of approximately 28 kg at 36 weeks’ gestational age is associated with the lowest hazard of early delivery (i.e., before 37 weeks’ gestational age); this is higher than current provisional guidelines, which recommend 17–25 kg total weight gain for twin pregnancies of individuals with normal weight prepregnancy BMI (10). We acknowledge that further studies that include other pregnancy outcomes as well as accumulation of evidence are required before recommending weight gain above current provisional Institute of Medicine guidelines for twin pregnancies. We also found evidence for effect modification of this relationship by prepregnancy BMI, even when using prepregnancy BMI–stratified twin- and singleton-specific charts for z-score conversion. Previous research also describes an interdependent relationship between prepregnancy BMI, weight gain, and gestational age at delivery (5).

The nonlinear and time-dependent effects that we describe strengthen the rationale for studying both weight gain and gestational age at delivery as continuous measures. We are aware of only 2 studies that have investigated this relationship using a time-to-event framework, and only among singleton pregnancies (20, 41). One study examined time-varying weight gain linearly interpolated between measurements (20), while the other investigated area under the weight gain curve (termed “pound-days”), where weight gain was estimated between visits using restricted cubic splines with random intercepts and slopes (41). Both studies found that high weight gain was associated with lower hazard of earlier delivery (20, 41). The latter study additionally noted that earlier accumulated pound-days decreased hazard of delivery (41). Importantly, neither study considered potential nonlinear relationships between time-varying weight gain and time to delivery, despite the previously reported U-shaped relationship between total weight gain and preterm birth. Neither study investigated different types of delivery as competing events; studies either grouped all types of delivery (20) or focused only on spontaneous deliveries (41).

Strengths of our study include its use of time-varying, continuous exposure and continuous outcome definitions that maintained precision of both weight gain and gestational age at delivery, flexible modeling of both nonlinear and time-dependent effects of weight gain, and treatment of delivery following different types of labor onset as competing events. Instead of producing a single effect estimate as a summary measure of complex underlying relationships, our model allowed us to simultaneously examine the strength and shape of effects. We are unaware of any studies that have taken this approach, particularly among twin pregnancies. We analyzed a wealth of data from 2 large cohorts of twin and singleton pregnancies from the same study population, including prenatal weight measurements abstracted from medical charts. Our data sources and approach allowed us to compare results in twin and singleton pregnancies.

We also recognize some limitations of our study. Although our statistical approach is flexible in comparison with those used in other studies, the extended Cox model is still somewhat limited in that it assumes a single shape of effects across gestational ages (30). For example, if the true relationship between weight gain and relative hazard of delivery is linear at 28 weeks’ gestational age but J-shaped at 32 weeks’ gestational age, the model may inaccurately estimate the overall relationship as a mix of these 2 functions. We feel this is unlikely to be an issue, particularly since our findings mirror those found in previous studies in singletons. We additionally addressed this limitation by converting actual weight gain in kilograms to gestational age–specific z scores to improve comparability of results across gestational age. We applied z-score charts for persons with normal weight to those with underweight prepregnancy BMI, since z-score charts for this stratum were unavailable (23, 27, 28). More information is needed (from cohorts including persons with underweight prepregnancy BMI) to characterize weight gain and its associated risks in this stratum. Given the increase in parameters needed, we lacked power to fully incorporate effect modification by prepregnancy BMI. We calculated weight gain z scores using BMI-stratified charts, which may have reduced the need for such parameters, but sensitivity analyses suggested that this did not entirely account for effect modification by prepregnancy BMI. We cannot rule out that the relationships we observed might have differed if other z-score charts, including references and standards derived from other study populations, were instead used to account for gestational duration. Comparison of z-score charts with respect to weight gain trajectory and its relationship with pregnancy outcomes is recommended in future studies. Our approach also necessitated an estimate for weight gain per day of gestational age, which we estimated using available prepregnancy, delivery, and prenatal weights. Interpolating between measurements produces some error; however, the method we employed produced error that was nondifferential with respect to both gestational age and total number of measurements, which would likely bias results towards the null. Our ability to make separate inferences about the risk of each type of delivery is limited because we analyzed delivery following different types of labor onset as competing risks. Additionally, the generalizability of our findings may be limited both within the United States (given the relatively few pregnancies of Hispanic persons in our study population) and globally (given different average total weight gain observed in other countries (42)). Any substantial variation between study populations in characteristics that modify the effect of weight gain on gestational age at delivery may affect the generalizability of our results.

We conclude that both low and high maternal weight gain are associated with gestational duration among twin pregnancies. Additionally, we highlight that increased hazard of early spontaneous preterm delivery among pregnancies with low weight gain and increased hazard of late preterm delivery without labor jointly contribute to this relationship. Our results mirror the relationship found in singletons, while also strengthening the evidence base for weight gain guidelines in twin pregnancies.