Study population

The dataset is part of the MoBa cohort, initiated by and maintained at the Norwegian Institute of Public Health [28]. In brief, MoBa is a nationwide pregnancy cohort, including more than 108,000 pregnancies during the period 1999 to 2009. Women were recruited by postal invitation in connection with the routine ultrasound examination offered to all pregnant women in Norway at around 17 gestational weeks. Overall, 38.5% of invited women have participated. Participants were asked to fill in questionnaires focused on general health status, lifestyle behavior and diet at gestational weeks 15 to 17 (Q1) and 30 (Q3). At gestational week 22, they completed a food frequency questionnaire (FFQ). All questionnaires are available on the Norwegian Institute of Public Health homepage [33]. This study used data from version 5 of the quality-assured data files made available for research in 2010. Pregnancy and birth records from the Medical Birth Registry of Norway (MBRN) are linked to the MoBa database [34]. Informed written consent was obtained from each participant. The Regional Committee for Medical Research and the Norwegian Data Inspectorate approved the study.

Of the 106,707 pregnancies included in MoBa version 5, 103,835 women gave birth to live-born singletons; 81,301 of these women had answered all three questionnaires. After exclusion of the following medical and pregnancy-related conditions 70,105 pregnancies remained in the study: diabetes mellitus, hypertension, autoimmune disease, inflammatory bowel disease, systemic lupus erythematosus, rheumatoid arthritis, scleroderma, other immune-compromised conditions, in vitro fertilization, pre-eclampsia, hypertension, gestational diabetes, placental abruption, placenta previa, cervical cerclage and serious fetal malformations. Women reporting improbable energy intake, that is, <4.5 MJ or >20 MJ, were excluded [35], leaving 69,045 pregnancies. If a woman participated with more than one pregnancy, only the first pregnancy was included, leaving 60,496. Finally, 59,123 women had complete data on pre-pregnancy weight and height.

Outcome

Gestational age in days was determined by second-trimester ultrasound in 98.3% of pregnancies and based on the last menstrual period in the remaining cases. The expected effect of caffeine on gestational length is minor; results are thus presented in hours instead of days. Spontaneous PTD was defined as birth after preterm labor or prelabor rupture of the membranes between 22⁺⁰ to 36⁺⁶ weeks, while controls delivered spontaneously between 39⁺⁰ to 40⁺⁶ weeks. A subanalysis was conducted for the subgroups early (22⁺⁰ to 33⁺⁶ weeks) and late (34⁺⁰ to 36⁺⁶ weeks) spontaneous PTD.

BW in grams was registered in the MBRN. As there is still no consensus on standard growth curves, data were analyzed according to three standards based on Northern European populations: ultrasound-based growth curves according to Marsal [36], population-based growth curves according to Skjaerven [37] and customized growth curves according to Gardosi [38].

The difference between BW and expected BW was calculated as a percentage of expected BW. This implicates that gestational length is taken into account by the definition of our outcome variable and analysis were not adjusted for gestational age. Percentage was used instead of the difference in g, as a slight weight difference matters much more if the expected BW for a preterm infant is very low compared with a normal-weight infant born at term. While the original outcome of the linear regression thus was a percentage of the expected BW, we chose to present the results in g for babies with an expected BW of 3,600 g, the rounded-out median BW in our study population (actual median: 3,620 g). SGA was defined according to the above-mentioned authors' respective definitions: less than-2 SD (Marsal) or <tenth percentile (Skjaerven, Gardosi). Standard deviation and percentiles used are based on the reference population in these publications, not on our study population, which is a highly selected subpopulation of the MoBa population, as described above.

Caffeine intake

In Q1 and Q3, women reported their coffee, tea and caffeinated soft drink consumption in cups or glasses/day. These data allowed following a participant's caffeine consumption from the three main caffeine sources from the time before pregnancy until gestational week 30. Caffeine intake was entered into the analysis in mg/day, adjusted to a woman's pre-pregnancy weight and recalculated as if every woman weighed 65 kg (median pre-pregnancy weight in the study population): 65 kg × caffeine intake/pre-pregnancy weight [44].

All analyses were based on the more detailed FFQ caffeine intake data, except when the association between caffeine intake and pregnancy outcomes at different timepoints was studied using Q1 and Q3 data.

Covariates

Information on maternal age at delivery and the baby's sex was available from the MBRN. Parity was based on both MoBa and MBRN data and categorized into number of previous pregnancies of ≥22⁺⁰ weeks' duration. Marital status was defined as either married/cohabiting or not. Self-reported pre-pregnancy height and weight were used to calculate pre-pregnancy body mass index (BMI), which was categorized according to the WHO classification as underweight (<18.5 kg/m²), normal weight (18.5 to 24.9 kg/m²), overweight (25 to 29.9 kg/m²) and obese (≥30 kg/m²). Maternal education was categorized as ≤12 years, 13 to 16 years or ≥17 years. History of previous PTD (22⁺⁰ to 36⁺⁶ weeks of gestation), as registered in the MBRN, was taken into account as a dichotomous variable. Women reported smoking habits during pregnancy in Q1 and were categorized as non-smokers, occasional or daily smokers. Passive smoking and use of other nicotine sources were considered to be dichotomous variables. Alcohol intake from different sources was self-reported in the FFQ (glasses/day, week or month) and calculated in g/day. Persistent nausea at the time of answering the FFQ was used as a dichotomous variable. Household income was categorized as follows: participant and her partner each earning <300,000 Norwegian Krones (NOK)/year, either participant or her partner earning ≥300,000 NOK/year or participant and her partner both earning ≥300,000 NOK/year. In MoBa, more than 99% of the participants are of Caucasian ethnicity; hence ethnicity is not a relevant confounder.

Statistics

All statistical analyses were performed using SPSS Statistics V.19.0 (SPSS, Chicago, IL, USA). Caffeine intake in relation to maternal characteristics was studied with the Kruskal-Wallis test. Associations between caffeine intake and gestational length and BW were studied with linear regression both in an unadjusted model and adjusted for the covariates mentioned above. We visually inspected residual plots to check if model assumptions were reasonably fulfilled. Odds ratios (OR) for caffeine intake and categorical outcome variables were estimated using logistic regression, both unadjusted and adjusted, as above. Statistical significance was assumed for a two-sided P value <0.05. Subanalyses were performed in the subgroups of non-smokers and non-coffee drinkers.

Gestational length and spontaneous PTD

A total of 49,102 women delivered spontaneously with a median gestational length of 282 days (IQR 276 to 287 days). Gestational length as well as its residuals were left skewed, but we preferred to use the original data scale for easier interpretation.

Caffeine intake from different sources (FFQ data)

Caffeine intake over time (Q1 and Q3 data)

At all timepoints studied, total and coffee caffeine intake was consistently associated with increased gestational length. The association was strongest for caffeine consumption reported at gestational week 17 (3 h/100 mg total caffeine/day, 95% CI 1 to 4 h; 4 h/100 mg coffee caffeine/day, 95% CI 2 to 5 h, both P <0.001), followed by that reported at gestational week 30 (2 h/100 mg total caffeine/day, 95% CI 0 to 3 h, P = 0.02; 3 h/100 mg coffee caffeine/day, 95% CI 1 to 5 h, P <0.001) and reported pre-pregnancy caffeine consumption (2 h/100 mg total caffeine/day, 95% CI 1 to 3 h; 2 h/100 mg coffee caffeine/day, 95% CI 1 to 3 h, both P <10^-6); all adjusted models.

Birth weight and SGA

The diagnosis of SGA varied considerably depending on the growth curve and SGA definition applied (Figure 3).

Caffeine intake from different sources (FFQ data)

When studied exclusively in non-smokers (n = 54,136), these associations remained significant, again with the exception of chocolate caffeine in the Gardosi model. However, the decrease in BW was somewhat lower: Marsal 18 g (95% CI 15 to 21 g) instead of 28 g, Skjaerven 15 g (95% CI 12 to 18 g) instead of 25 g, Gardosi 12 g (95% CI 9 to 15 g) instead of 21 g per 100 mg additional total caffeine/day (all significant with P <10^-15; adjusted models).

To test if there was a threshold effect, we performed the same logistic regression with sextiles of total caffeine intake (0 to 14.645, 14.646 to 32.093, 32.094 to 57.265, 57.266 to 96.029, 9603 to 163.806, >163.806 mg/day). In all three models the caffeine intake categories were associated with increasing odds ratios for SGA as compared to the lowest intake group (see Figure 4). According to FFQ data, 10.8% of all women exceeded the NNR recommendation of less than 200 mg/day caffeine intake during pregnancy and 3.3% also exceeded the WHO recommendation of less than 300 mg/day. If the odds for SGA were studied with reference to these recommendations, those 7.7% of women with a daily caffeine intake of 200 to 300 mg had significantly higher odds for SGA (1.27 to 1.62, depending on the SGA definition), in comparison with the lowest (0 to 50 mg/day) caffeine intake group. The odds of giving birth to a SGA infant were 1.62 to 1.66 in the 3.3% of women consuming >300 mg caffeine/day (Table 7).

Caffeine intake over time (Q1 and Q3 data)

Gestational length and spontaneous PTD

We found that total and coffee caffeine intake was associated with a highly significant increase of gestational length by 5 and 8 hours/100 mg respectively. The corresponding association with total caffeine intake was, conversely, 10 hours decreased gestational length, though only marginally significant, when coffee drinkers were excluded from the model. Additionally, we ruled out a threshold effect, as even the groups with the lowest coffee caffeine intake were significantly associated with gestational length. Our results do not support the hypothesis that caffeine intake influences the risk for spontaneous PTD. The only marginally significant finding was black tea caffeine being associated with higher PTD odds, in the relatively small subgroup of early spontaneous PTD. We therefore conclude that the association of total caffeine intake with gestational length is not related to caffeine but to coffee intake. This study was not designed to disclose the reason for this statistical association; one possible explanation is that there might be some other substance, present in coffee but not in the other caffeine sources, that influences gestational length. Human parturition is depending on a physiological inflammatory reaction leading to cervical ripening and increased uterus tonus [46]. Melanoidins that are generated from coffee bean components during the roasting process are, for example, known to have antimicrobial and anti-inflammatory effects [47] and thus might influence the timing of parturition. People in Scandinavia who do not drink coffee constitute a definite minority and those with a very low caffeine intake are probably a special group in many other ways. Drinking coffee, but not consuming other caffeine sources, might be associated with gestational length by some lifestyle factor.

The most important confounder on the behavioral level is smoking. Smokers are known to have higher coffee consumption [20]; furthermore, smoking is an established risk factor for spontaneous PTD. According to our results, however, coffee drinking and smoking have opposing effects on gestational length. The association with coffee intake remained significant after adjusting for smoking and excluding all smokers from the analysis, strongly suggesting an association between coffee consumption and gestational length that is independent of smoking behavior.

There was no major difference regarding the association for coffee consumption during different periods of pregnancy, suggesting either a rather continuous effect of coffee drinking on gestational length or confounding of coffee consumption with some other factor, as opposed to coffee consumption at a specific timepoint affecting some crucial step of pregnancy development.

In summary, we seem to have identified an association for coffee rather than caffeine, which must be kept in mind when discussing our findings in the context of earlier publications. To the best of our knowledge, this study is the first to separately study the associations between gestational length and caffeine from coffee, on the one hand, and caffeine from other sources, on the other. In comparison, most observational studies have not found any associations between caffeine or coffee and gestational length [48–53] and Maslova et al. found no significant association with overall PTD risk in their meta-analysis [31]. There may be several reasons for the fact that we found an association for coffee drinking, but not for caffeine per se. We used self-reported caffeine intake instead of measuring caffeine metabolites [49, 50]. Caffeine from several sources was assessed, rather than caffeine intake from a single source, usually coffee, as in many studies [30, 53–55]. For some populations using only caffeine from coffee would implicate that a major, if not the major, part of total caffeine intake was not considered at all, for example, in a UK study black tea contributed 62% to daily caffeine intake while coffee and cola drinks accounted for about 12% to 14% each [45]. PTD is a heterogeneous group of pregnancy outcomes with heterogeneous etiology [32]. In contrast to the studies mentioned above, we defined a clear spontaneous PTD phenotype by excluding all iatrogenic deliveries as well as all medical or obstetric complications. The etiologies of early and late spontaneous PTD differ, which has not been acknowledged in many other studies [20]. Mikkelsen et al. found an increased risk of early, but not late, overall PTD related to coffee intake of more than 2 cups/day in the Danish Birth Cohort [30], while Haugen et al. failed to confirm these results in an earlier version of MoBa [55]. In this study, in which only spontaneous PTD in uncomplicated pregnancies was investigated, black tea caffeine intake was significantly associated with increased risk for early spontaneous PTD while significant associations were not found for other caffeine sources, indicating that this association was not caused by caffeine. The association with black tea caffeine was only marginally significant and there was no significant association between black tea caffeine and gestational length so that these results should be interpreted with caution.

In addition to the above-mentioned paper by Mikkelsen et al., Klonoff-Cohen et al. reported decreased gestational length related to caffeine consumption of >50 mg/day, compared to 0 to 2 mg/day, in a sample of 39 pregnancies [29]. To our knowledge, this study of 49,102 pregnancies with spontaneous delivery is the biggest and most detailed observational study so far on the association between caffeine intake and gestational length, particularly spontaneous PTD. Our data do not support the hypothesis that caffeine intake or coffee consumption decrease gestational length or increase the risk for spontaneous PTD. Although we did find a significant association, a change in gestational length of several hours probably lacks clinical implications.

Birth weight and SGA

We found significant associations between caffeine intake and SGA and decreased BW. These associations were strengthened by concordant results for different caffeine sources, comparable overall findings regardless of the growth standard and SGA definition applied, remaining significance after adjustment, biological gradient, stability over period of pregnancy, consistency with other studies and biological plausibility. Caffeine is metabolized more slowly during pregnancy, crosses the placental barrier [7] and increases maternal levels of 3'5'-cyclic monophosphate and epinephrine [56], causing uteroplacental vasoconstriction and decreased intervillous placental blood flow, which could restrict fetal growth [48, 54]. Another hypothesis, postulated by Weathersbee et al., is that caffeine inhibits phosphodiesterase, leading to an increase in cellular cyclic adenosine monophosphate, which may interfere with fetal growth [57].

Smoking is a difficult confounder when studying effects of caffeine consumption [20]. Both smoking and caffeine intake are associated with lower BW. However, the association between caffeine intake and BW remained highly significant after adjustment for smoking and after analysis in the non-smoker subgroup, suggesting an independent association with caffeine consumption.

Caffeine intake reported at gestational week 17 was most strongly associated with BW. This could be explained by reverse causality, according to Lawson's hypothesis, that is, the placenta is comparatively smaller in pregnancies complicated by SGA than in healthy pregnancies, thus producing less hormones and evoking fewer pregnancy symptoms so that these women might maintain a higher caffeine intake [20, 58]. However, remaining significance after controlling for nausea and the finding of a significant association for pre-pregnancy caffeine intake as well contradicts reverse causality as only explanation.

While some earlier studies found no significant association between caffeine consumption and BW [48, 53, 54, 59], most publications are consistent with our findings in MoBa [49, 50, 60–63]. Especially the data from some of the largest observational studies published so far, are consistent with our findings, moreover with comparable effect size of a decrease in BW by 60 to 70 g for >200 mg/day [45] or 28 g for 100 mg/day caffeine consumption [50].

In this study population, more than 10% exceeded the NNR recommended maximum intake of 200 mg/day; this subgroup had 20% to 60% higher odds ratios for SGA. Although SGA babies are generally known to be at higher risk for both neonatal morbidity and mortality [24], this might not be true for babies born SGA due to maternal caffeine consumption. Hernandez-Diaz found that babies born SGA due to maternal smoking might have lower mortality than other SGA babies with more severe causes for being born SGA, such as congenital malformations [64]. However, as our results confirm earlier findings [45] that the increase in SGA risk can already be found in women following the current recommendations by Norwegian Authorities, further studies are needed to establish the impact of caffeine on neonatal morbidity and mortality. We could not find a threshold for the association of caffeine consumption and SGA risk. Until there is clarity if there is a causal association between caffeine intake and increased risk for SGA, women might be advised to reduce their caffeine consumption as much as possible during pregnancy.

Limitation and strengths

To the best of our knowledge, with its sample size of 59,123 pregnancies, this is the largest study performed so far on the association between caffeine intake and pregnancy outcome. The MoBa participation rate is 38.5%, and demographic comparison with the MBRN in 2002 showed that single women and women aged <25 are underrepresented in MoBa. Regarding SGA (4.6% in MoBa and 5.1% in the MBRN) and PTD (7.2% in MoBa and 7.7% in the MBRN), the differences are minor and the subgroup composition is similar to that in the total population, with spontaneous PTD accounting for 42% of all PTD [28]. Additionally, a recent study found no bias in eight selected exposure-outcome associations [65].

Due to the large study sample, there were 1,411 cases defined as spontaneous PTD and 852 (Marsal), 4,503 (Skjaerven) or 4,733 (Gardosi) cases of SGA in the study population. Estimation of gestational length by second-trimester ultrasound and clear definition of a spontaneous PTD phenotype are additional strengths of this study [20, 24, 32]. Different standard growth curves were applied and this study is one of the first caffeine effect studies using customized growth curves at all [20, 45]. The overall results for all three models indicate an adverse effect of caffeine consumption on SGA risk, strengthening the association found.

All dietary assessment methods have limitations, and so does the self-reported caffeine intake in this study. The mean caffeine concentrations in seven main categories of food and drinks were used for the exposure calculations, while large variations may actually exist within each category [41]. The caffeine contributed by soft drinks is likely to be underestimated as Coca Cola and Pepsi (including their diet versions) were the only soft drinks distinguished from other soft drinks in the FFQ. However, some other soft drinks also contain caffeine, for example, Urge, a citrus flavored soft drink produced by Coca Cola Norway. However, this soft drink comprised a rather small share of the market.

Relying exclusively on self-reported data without a biological marker to confirm the accuracy of estimated caffeine exposure is a weakness. For the present study we evaluated the agreement between caffeine intake estimated by the FFQ and a food diary. The estimated caffeine intake did not differ between the methods, and a high correlation was observed (r = 0.70, 95% CI 0.59 to 0.78). Furthermore, the MoBa FFQ has been extensively validated in a MoBa subpopulation using the four-day weighed food diary and several biomarkers as reference measures [42, 43]. The agreement between the FFQ and the food diary was particularly high for coffee and tea, which are the main sources of caffeine in this study. Coffee intake according to both the FFQ and the food diary correlated with serum β-carotene (0.31 and 0.36, respectively), which can be explained by interplay between antioxidants in coffee with β-carotene, as also reported by Svilaas et al.[66]. Likewise, tea intake according to both the FFQ and the food diary correlated with kaempferol, a flavonoid found in tea (r = 0.41 and 0.50 for the FFQ and food diary, respectively [42]). Similar, but slightly weaker correlations were observed for estimated caffeine contributed by coffee and tea. A Bland-Altman plot for the differences in caffeine intake between the FFQ and the food diary is available as Additional file 1.

There are further strengths related to the caffeine intake assessment: the prospective design ensured that women's responses were not influenced by their knowledge of pregnancy outcome. Caffeine intake assessment from different sources, as well as different coffee preparations being taken into account, are also clear strengths of this study [20]. As the FFQ covers the first four to five months of pregnancy, when many women change their dietary habits due to nausea, some women may have had difficulties reporting the frequency and amount of caffeine consumption for the whole period correctly. Coffee and black tea intake varies less and is easier to recall than intake of most other food groups. The associations with caffeine intake based on the FFQ were corroborated by caffeine intake estimates based on reported consumption of caffeine-containing drinks in the other two questionnaires (Q1 and Q3). As the relevant window of susceptibility for caffeine effects is not yet known [20], caffeine consumption assessment at different timepoints is a further strength of this study.

There is always a possibility of residual confounding in observational studies, but we reduced this possibility by controlling for a number of relevant factors, including history of PTD, nausea and smoking. Our smoking variable has been shown to be a valid marker for tobacco use when tested against plasma cotinine concentration [67].