Details of the study design have been reported previously [17, 23]. In brief, a population-based cross-sectional study was performed in Shaanxi Province of Northwest China between August and November 2013. This area is normally divided into three regions: northern, southern, and central Shaanxi, with natural resources, culture, and lifestyle differing greatly among them. A total of 30,027 women who were pregnant during 2010–2013 were recruited using a stratified multistage random sampling method. The sampling process is as follows: twenty counties and ten districts were randomly sampled according to the proportion of rural to urban residents, population size, and fertility rate in Shaanxi, China; in each sampled county, six villages each from six townships were randomly selected; in each sampled district, six communities each from three streets were randomly selected; 30 and 60 eligible women were randomly selected from each selected village and community, respectively. Among the participants, 7750 women who were pregnant during 2012–2013 and had infants less than 12 months old were further interviewed to report their diets during pregnancy. We excluded 87 women with a multiple gestation, 65 women without offspring birth weight, and 288 women with an implausible total energy intake (less than 500 kcal/day or greater than 5000 kcal/day), leaving 7310 eligible participants for the final analysis. The flow diagram of sampling strategy with exclusion criteria is shown in Additional file 1: Fig. S1.

This study was conducted according to the guidelines laid down in the Declaration of Helsinki and all procedures were approved by the Xi’an Jiaotong University Health Science Center. Written informed consent was obtained from all participants.

Maternal dietary intake during the whole pregnancy was collected by a 107-item semi-quantitative food frequency questionnaire (FFQ) at 0–12 months (median: 3; 10–90th percentiles: 0–7) after delivery [17, 23]. Maternal dietary patterns and nutrient intakes tended to change little from early to late pregnancy [24]; thus, for large-scale epidemiological studies, especially for those with multiple dietary exposures and outcomes like the present study, diet assessment during the whole pregnancy at one time was reasonable, convenient, and economical [17, 23, 25]. The FFQ was established according to the previously validated FFQ designed for pregnant women in Shaanxi, China [26]. In the validation study, the Pearson correlation coefficient for protein between the FFQ and the average of six 24-h recalls was 0.61, with a range of 0.53 to 0.70 for other nutrients [26]. The frequency scales of five food items (animal oils, vegetable oils, salt, sugar, and sauce) were open-ended, and were listed as kilograms per month and the number of people regularly consuming them. The frequency of the other 102 food items was reported according to eight predefined categories ranging from never to two or more times per day, and their portion sizes were recorded according to food portion images [17, 23]. Daily intakes of total dietary protein, animal protein, plant protein, major dietary protein sources, and other nutrients were transformed using the China Food Composition Tables [27, 28]. Animal protein was derived from animal-based foods, including pork, beef, lamb, chicken and other poultry, eggs, dairy, fish, and seafoods. Plant protein was derived from plant-based foods, including cereals, legumes, nuts, vegetables, and fruits. The recommended percentages of energy from protein, fat, and carbohydrate in China were 10–20%, 20–30%, and 50–65%, respectively [29]. The recommended additional caloric intake per day during the first, second, and third trimesters of pregnancy in China was 0, 300 kcal, and 450 kcal, respectively [29].

Neonatal information including birth weight, gestational age, sex, and birth date was obtained by reviewing birth certificates. Birth certificates were finished by the medical staff once the neonates were born. Birth weight was measured with a baby scale with precision to the nearest 10 g. Gestational age at delivery was calculated according to the last menstrual period, and was confirmed by ultrasound scans. Medical records including physical examinations, clinical diagnosis, and medical history were referred to ascertain birth outcomes. The primary outcome of the present study was birth weight, and the secondary outcomes were LBW, SGA, and IUGR. LBW was defined as birth weight < 2500 g. SGA was defined as birth weight below 10th percentile of the gestational age-sex specific international reference for fetal growth [30]. IUGR was defined as birth weight below 3rd percentile of the gestational age-sex specific international reference for fetal growth [30].

The general information of the participants during pregnancy was collected face to face by well-trained interviewers using a standard questionnaire. The study information was classified as follows: (1) socio-demographic characteristics: geographic area (northern, southern, or central Shaanxi); residence (rural or urban); childbearing age (< 25 years, 25–29 years, or ≥ 30 years); maternal education (primary school or below, junior high school, or senior high school or above); maternal occupation (farmer or working outside); nulliparity (yes or no); (2) health-related characteristics: passive smoking (yes or no); alcohol drinking (yes or no); antenatal care visit frequency (< 6 or ≥ 6); folate/iron supplements use (yes or no); anemia (yes or no); medication use (yes or no). Passive smoking was defined as being exposed to other people's tobacco smoke for ≥ 15 min/day. Alcohol drinking included a wide range of alcoholic beverages (liquor, wine, and beer) consumed in pregnancy. Folate/iron supplements use was defined as taking dietary supplements containing folate or iron for more than 2 weeks. Anemia in pregnancy was diagnosed using the criteria of hemoglobin concentration < 110 g/L. Medication use was defined as taking any medication in pregnancy.

Because total energy intake is correlated with most nutrients, macronutrient intake was expressed as a percentage of total energy intake by the nutrient-density method and other nutrients were energy-adjusted by the residual method [31]. The study population characteristics according to quartiles of total dietary protein and animal protein intakes were described as percentages or means, with differences tested by χ2 test for categorical variables and analysis of variance for continuous variables. Household wealth index was established by principal component analysis according to the items reflecting family economic level (housing condition, vehicle type, income source, and type and number of household appliance), and this index was divided into thirds as an indicator for the poor, medium, and rich households [32]. To avoid multicollinearity of nutrients in regression analyses, we extracted the first component by principal component analysis according to the intakes of potential nutrients (vitamin A, thiamin, riboflavin, folate, vitamin C, vitamin E, calcium, zinc, and selenium) that explained 67.5% of the total variance, with the factor loadings of zinc, riboflavin, folate, thiamin, calcium, and selenium above 0.80.

Considering the stratified multistage random sampling design, multilevel models were applied to assess the associations of dietary protein intake with birth weight and the related adverse birth outcomes (LBW, SGA, and IUGR). After running the four-level empty models representing county (district)-township (community)-village(street)-individual, we observed nonsignificant within-group variations (all P > 0.05) and low intra-class correlations (all lower than 0.001) of the village (street) level; thus, simplified multilevel models with a random intercept at the county (district) and the township (community) levels were adopted. Multilevel linear regression models were used to estimate birth weight changes associated with different dietary protein sources during pregnancy, and multilevel logistic regression models were used to evaluate ORs (95% CIs) for adverse birth outcomes associated with different dietary protein sources during pregnancy. Protein intake values were computed as 3% energy units to assess the associations. The 3% energy unit was chosen because it was identical around 60 kcal energy and 15 g protein in the study population, which could be easily realized by increasing 75 g pork lean or 60 g chicken breast according to the China Food Composition Tables [27, 28]. Dietary protein intake was also categorized into quartiles to avoid the possible influence of extreme values. Based on previous studies [23, 33], models were adjusted for total energy intake, socio-demographic characteristics (including geographic area, residence, childbearing age, education, occupation, household wealth index, and parity), health-related characteristics (including passive smoking, alcohol drinking, antenatal care visit frequency, folate/iron supplements use, anemia, and medication use), and principal component score based on the nutrient intakes. For birth weight and LBW, models were additionally adjusted for offspring sex and gestational age. Animal protein and plant protein were mutually adjusted for one another. Further adjustment for other major dietary protein sources was performed in the analysis of specific major dietary protein source. To test for a linear trend, we used the median for each quartile of protein intake as a continuous variable. We further evaluated the interactions by introducing cross-product terms into regression models to assess whether the associations were modified by baseline characteristics including offspring sex, geographic area, residence, childbearing age, maternal education, maternal occupation, household wealth index, and parity.

To simulate the substitution of dietary protein for carbohydrate, we fitted isocaloric models [31] by simultaneously including the percentages of energy from fat and protein, total energy intake, and all other potential confounders. The effect estimate from this model can reflect the effect of increasing protein intake at the expense of carbohydrate while keeping calories constant. Similarly, to simulate the substitution of dietary protein for fat, we simultaneously included the percentages of energy from carbohydrate and protein, total energy intake, and all other potential confounders.

A two-tailed P < 0.05 was considered as statistically significant. All statistical analyses were performed using STATA software (version 12.0; StataCorp, College Station, Texas, USA).