Infancy weight gain, parental socioeconomic position, and childhood overweight and obesity: a Danish register-based cohort study

This register-based longitudinal cohort study links The Children’s Database with other nationwide administrative Danish registries. The administrative register The Children’s Database has collected data on children’s (0–17 years) health status since 2009. Data is compiled at regular preventive health assessments during infancy and childhood [40, 41]. From December 2011 it has been mandatory for Danish municipalities to report data collected by health visitors, including data on child height and weight [42]. Every Danish citizen is given a unique Civil Personal Register number at birth, which enables linkage of individual-levelled data across nationwide administrative registries [43, 44]. Individual-levelled data for the study population and their parents were obtained from The Medical Birth Registry [45], The Danish Civil Registration System [44], The Income Statistics Register [46], and The Population Education Register [47].

Health professionals measured weight and height during planned health examinations. Weight was measured using a validated scale and reported to the nearest 0.1 kg. Height was measured recumbent until two years of age and reported to the nearest 0.5 cm [40]. We calculated weight-for-age and body mass index z-scores describing infants anthropometric standard deviations in relation to the 2006 World Health Organization (WHO) Child Growth Standards using WHO Anthro 2011, version 3.2.2 [48]. Infant weight gain was defined as the change in weight-for-age z-scores between birth and 8 to 10 months. A change of 0.67 represents moving one percentile band on a growth chart [3, 4, 49], and the change in weight gain were thus categorised as: <− 0.67 standard deviations (SD) (slow weight gain), − 0.67 to 0.67 SD (mean weight gain), > 0.67 to 1.34 SD (rapid weight gain), and > 1.34 SD (very rapid weight gain). The timing of measurement was chosen as RIWG during the first, compared to the second, year of life has shown a stronger association with later risk of overweight and obesity [4] and because a preventive health assessments including infant measurement is offered around 8–10 months of age in Denmark. The measurement closest to 9 months of age were used if several measurements were available. Implausible values (<− 6 or > 5 SD) were excluded [50]. The study outcome was childhood overweight or obesity at 2 years of age, which was defined as having a body mass index z-score of more than 2 SD [51, 52]. The measurements closest to 24 months of age were applied if several measurements were available, and implausible values (<− 5 or > 5 SD) were excluded [50].

Maternal education and household income served as proxies for the socioeconomic position. Level of maternal education closest to the year before or after birth was identified through The Population Education Register. This information was categorised in accordance with the International Standard Classification of Education (ISCED) 2011 [53] and adjusted to a Danish education setting as done by Ullits and colleagues [54], thus consisting of the four categories: ISCED level 0–2 (up to and including lower secondary education), 4 (post-secondary non-tertiary education), 5–6 (short-cycle tertiary education, or Bachelor’s degree or equivalent) and 7–8 (Master’s/Doctoral degree or equivalent). Household income level was identified through The Income Statistics Register and defined as mean annual household income calculated from the two years prior to birth and divided into quartiles (low, low-middle, high-middle and high).

A recent review on the relationship between RIWG and COO informed the selection of relevant child and maternal covariates [4]; child sex, breastfeeding duration, birthweight, mode of delivery, maternal pre-pregnancy body mass index (BMI), maternal smoking status during pregnancy, gestational age at birth, parity, and gestational diabetes status. Data on infant sex and breastfeeding duration were identified from The Children’s Database. WHO recommend six months of exclusive breastfeeding [55], but many parents introduce complementary foods from four months of age [56]. Breastfeeding duration was thus categorised into four groups: 0–2 months, 2–4 months, 4–6 months, > 6 months, where the cut-off at two months was set to separate short- and medium-termed breastfeeding. Information on infant birthweight, mode of delivery, maternal pre-pregnancy BMI, smoking status during pregnancy, gestational age at birth, parity and occurrence of gestational diabetes were identified in The Medical Birth Registry. Birthweight and gestational age at birth were applied to identify infant size for gestational age. Thresholds reported by Kramer and colleagues [57] were applied to define three categories: Small- (SGA), appropriate- (AGA) and large-for-gestational age (LGA). Maternal pre-pregnancy BMI was grouped in accordance with WHO categories: < 18.5 (underweight), 18.5–24.9 (normal weight), 25.0–29.9 (overweight) 30.0–34.9 (obesity class I) and > 35.0 (obesity class II + III) [58]. Information on maternal smoking status during pregnancy was divided into three groups: Yes, Stopped during pregnancy, No. Gestational age at birth was included as the week of gestation, and parity was categorised as 1, 2 or > 2 completed pregnancies. Mode of delivery consisted of the categories vaginal birth and caesarean section. Finally, information on infant’s birth year, region of habitation at the time of the last anthropometric measurement, and the number of anthropometric measurements registered in The Children’s Database was included in all adjusted models in order to control for potential differences.

Between-group differences in covariate distribution across exposure categories (infancy weight gain, maternal education level and household income level) were tested using Chi-Square (X²) tests. Missing values were imputed by chained equations using the mice package in R [59]. The number of imputations was set to 33 and reflected the highest percentage of missing information on a single covariate, with 20 iterations in each chain [59]. All covariates in addition to categorical information on maternal age, birth length, and maternal ethnicity were used to inform the imputation. These additional covariates were left out of the main analyses due to the risk of over-adjustment. The administrative databases lack information on educational attainment and household income for some citizens, especially for immigrants that had their education before entering Denmark. Observations with missing data on educational attainment and income (n = 1123) were therefore excluded prior to imputation due to a potential violation of the missing not at random assumption [60]. Complete case analyses were carried out as sensitivity analyses. Furthermore, the study population and the population excluded due to missing anthropometric data were compared on sex, size for gestational age at birth, maternal education, household income, weight gain category, and COO.

Unadjusted, intermediate partly adjusted, and fully adjusted logistic regression model were used to estimate odds ratios (OR) and corresponding 95% confidence intervals (CI) of COO risks. Two intermediate models comprised of either maternal education or income level, while two fully adjusted models additionally included all covariates. This approach was chosen, as we wanted to examine the potential confounding effect of SEP alone if SEP was not found to be an effect modifier in the relationship between weight gain category and COO. Multiplicative interaction between weight gain category and SEP were tested in the fully adjusted models. Models including all covariates were stratified for SEP to further study between-group differences. Further search for high-risk groups in special need for preventive attention was conducted via tests for additive interaction between weight gain category and SEP. This was evaluated through the calculation of relative excess risk due to interaction (RERI) using the method involving categorical exposures, as described by VanderWeele and Knol [61]. RERI was calculated based on 2 × 2 tables. Maternal education and household income were dichotomised in order to ease interpretations of the results. Dichotomisation involved collapsing the two lowest and two highest categories. Comparisons between the categories of mean weight gain and rapid or very rapid weight gain were carried out. This enabled calculation of RERI for four subgroups. High level of SEP and mean weight gain category were coded as reference, as these exposure groups produced the lowest outcome risk when considered jointly [62]. These analyses were conducted on complete case data, as the method described was not developed for application on imputed data. Data management and calculation of RERI was conducted using the SAS Statistical Software package for Windows, version 9.4 (SAS Institute, Cary, NC, USA), and all further statistical analyses were conducted by using the R statistical software, version 3.5.0 [63].

This study was approved by The Danish Data Protection Agency (ref: 2008-58-0028, internal ref.: 2017–67). In Denmark, it is not necessary to obtain ethical approval for the use of administrative register information in research, as long as the project does not involve human biological material [64].

In total, 196,376 infants born between December 2011 and May 2015 were registered at least once with their height or weight in The Children’s Database. Figure 1 illustrates how the study population of 19,894 children was arrived at. Their background information stratified by weight gain category is shown in Table 1.

Of the study population, 3800 (19.1%) were classified with rapid and 2998 (15.1%) with very rapid weight gain. The distribution of all covariates except maternal pre-pregnancy BMI and occurrence of gestational diabetes differed statistical significantly across weight gain categories. Compared to infants with mean weight gain, a higher proportion of infants with rapid or very rapid weight gain were born SGA, before 39 weeks of gestation, by caesarean section, by primiparous mothers or by mothers that smoked during pregnancy, and were breastfed for less than 2 months (Table 1). The prevalence of rapid and very rapid weight gain increased with decreasing level of maternal education and household income. A social gradient was evident for rapid and very rapid weight gain when measured by either the level of maternal education or household income. Covariate distribution appeared mostly similar in both maternal education and household income (Additional files 1 and 2). The lowest prevalences of most risk factors were seen in the highest positioned groups.

A total of 1497 (7.5%) individuals were categorised with COO at follow-up at two years of age (22–26 months). The proportion increased with increasing weight gain category. 932 (15.9%) of the 5866 infants with rapid or very rapid weight gain were classified with COO. In the unadjusted model, the risk of COO was increased in infants with rapid (OR 2.18, [95%CI 1.90 to 2.51]) and very rapid weight gain (OR 4.10, [95%CI 3.59 to 4.68]) when compared to infants with mean weight gain (Table 2). These estimates remained unchanged after adjusting for SEP in the intermediate models, but they increased considerably after adjusting for covariates (Adjusted model 2a and 2b, Table 2). The risk of COO for each specific infancy weight gain category were relatively similar across levels of maternal education and household income in the stratified analyses, and confidence intervals for each specific infancy weight gain category overlapped when compared across SEP categories (Table 3). Correspondingly, no signs of interactions on a multiplicative scale were detected (maternal education: p = 0.89, household income: p = 0.24).

The values of RERI were generally positive, indicating super-additive effect modification when comparing the combined effect of low SEP and rapid or very rapid weight gain on COO risk (Table 4). Only the RERI for household income and rapid weight gain were negative (RERI: -0.19 [95%CI − 1.19 to 0.81]), but the combination of household income level and very rapid weight gain produced the largest positive RERI (RERI: 1.07 [95%CI − 0.97 to 3.11]). Although indicating super-additive effect modification in most combinations, none of these reached statistical significance.

Of the study population, 13,157 (66.1%) had complete information on all covariates. Frequencies of missing data were generally low except 32.7% lacking information on breastfeeding duration. Main results from the complete case analyses did not differ notably from results based on imputed data (Additional files 3 and 4). The distribution of weight gain category and COO in infants excluded due to missing anthropometric information did not differ from the distribution in the study population (Additional file 5). Significant differences were observed in the distribution of infant sex, size for gestational age at birth, maternal education and household income when comparing infants with and without missing data on exposure and outcome.

Our results suggest that the relationship between rapid and very rapid infancy weight gain and COO are consistent across socioeconomic positions, implying that SEP does not modify this risk-relationship. There were weak signs of additive interaction in some combinations of SEP and rapid or very rapid weight gain on COO risk, suggesting that the public health value of preventing rapid or very rapid weight gain as a part of an early-life COO prevention strategy could be greater if carried out in groups of low compared to high SEP groups. However, these results did not reach significance. Thus, there were no support for our hypotheses proposing that the association between rapid infant weight gain and risk of developing COO were stronger in infants from parents with low than with high SEP. Thus, our comprehensive register-based cohort study confirms the results of the previous, smaller studies within this area [38, 39], and our findings suggest that activities involving prevention of RIWG and promotion of healthy infant weight gain as a part of an early-life COO prevention strategy can be relevant for all socioeconomic groups.

As in previous studies [10, 11], our results show a social gradient in infancy weight gain, where the prevalence of rapid and very rapid weight gain is higher in those with lower levels of SEP. The gradient was clearest in infants with very rapid weight gain, and higher accumulation of risk factors in infants with this weight gain pattern as well as in infants from low SEP could help to explain the social gradient. Similarly, the risk of COO increased with decreasing SEP in the unadjusted analyses. Much of this social gradient disappeared after adjusting for covariates, which indicate that a social distribution of the pre- and postnatal risk factors accounted for in this study could explain these social differences to some extent. However, risk estimates were larger for household income than for maternal eduction levels, and only the estimate for the group with the lowest income level remained significant after adjusting for covariates. This may suggest that processes linking income to health can be of special importance for early life COO risk [65]. Better access to quality food, education, health and leisure services may then be central mechanisms [65, 66].

The prevalence of COO risk factors like being born LGA, having high maternal pre-pregnancy BMI and maternal smoking during pregnancy were unevenly distributed across weight gain categories in the study population. We found that the association between infancy weight gain and COO more than doubled for both rapid and very rapid weight gain after adjusting for covariates, which suggests that other factors confounded this relationship. Our study focus was limited to considering effect modification between infancy weight gain and SEP on COO risk, and no further analyses were carried out to identify potentially responsible covariate(s). Lamb et al. [36] found that the relationship between size for gestational age and higher childhood BMI was strengthened after adjusting for infancy weight gain during the first year of life, and they explained this by suggesting that high birthweight infants still have increased risk of overweight and obesity in later childhood despite having slower weight gain trajectories during the first year of life. Notably, the risk of COO also increased considerably for LGA infants in the fully adjusted models, and this might give some support to the hypothesis proposing that the risk of COO may already commence during pregnancy through pre-programming processes that promote high birthweight [67, 68].

In the study population, a total of 15.1% experienced very rapid weight gain, and this weight gain category was associated with a seven-fold increase in COO risk when compared to infants with mean weight gain. This is consistent with the findings of a meta-analysis by Druet et al. [69], where very rapid weight gain during the first year of life, which is the equivalent to moving up two or more percentiles on a growth chart, was a particularly important predictor for later COO development. The implications of very rapid weight gain on COO risk can also be traced longer-termed, as Johnson et al. [70] observed that infants with very rapid weight gain between 0 and 3 years of age had higher mean BMI and COO risks at the ages of 11 and 14 years. This increase was higher than the increase observed in children with rapid weight gain only.

RIWG has commonly been studied as a dichotomization using a cut-off score of > 0.67 SD change in weight, as suggested in the seminal paper by Ong in 2000 [49]. As opposed to previous studies [38, 39], the large sample in the current study enabled a more detailed elaboration and division between rapid and very rapid weight gain in relation to population characteristics and COO risk. Our results showing significant risk differences between weight gain categories suggest that application of a simpler weight gain categorisation could result in a notable information loss. A more nuanced categorization may improve interpretation of risk estimates, as the group of reference simply consist of infants with mean weight gain rather than a combination of both slow and mean weight gain.

Another strength is having up-to-date information from national administrative databases on both infants and parents, thus enabling comprehensive confounder adjustment in multivariable regression models on contemporary data. The choice of covariates was informed by literature, but it is possible that our models only partly capture the true relationship between infancy weight gain, socioeconomic status and childhood obesity, as this relationship is fairly complex [71]. We could have conducted a more comprehensive conceptual interpretation of the relationship in e.g. a directed acyclic graph (DAG), but the traditional approach was chosen as this relationship is complicated and would possibly not be captured in a DAG.

Using data not collected for the purpose of this study can also have introduced some limitations. General practitioners are, in contrast to the health visitors, not legally obliged to register their data to the database, so local differences in data registration over time may have occurred. Information on the number of registrations for each child, birth year and administrative region were therefore included in the adjusted models in order to reduce the magnitude of this potential selection bias.

Furthermore, a high proportion of missing data was observed for the duration of breastfeeding. The task of collecting and reporting these data to The Children’s Database is generally new to health visitors, and the lack of everyday routines could have compromised data quality and completeness. However, we have no reason to believe that the data quality is compromised differently across exposure or outcome categories and possible bias would thus lead towards the null and not explain the large OR detected in our study. Furthermore, multiple imputations were made on missing values and results from complete case analyses support the robustness of our main results. We assumed that missingness of breastfeeding data were related to regional differences in registrations or birth year rather than to the length of breastfeeding itself, and multiple imputation are a more efficient method associated with less bias than complete case analyses when data are missing at random [72]. The validity of the additive interaction analyses can be questioned as these were based on complete case data, but the consistency in analyses made on both imputed and complete case data support the validity of these results.

A large proportion of infants were not registered with sufficient data in order to define exposure and outcome, why our results could be affected by selection bias. However, no significant difference was observed in the distribution of exposure or outcome when comparing populations with and without missing data, which suggest that potential effect of such bias on the results may be small. Similarly, we do not believe that differences in covariate distribution in these populations have had any significant impact on the direction of the results, as the distribution differences were small and reached significance presumably due to high power.

Finally, the generalisability of our findings is restricted to populations born at term with a birthweight of more than 2500 g. Postnatal growth of infants experiencing in-utero growth restrictions may differ from other infants due to physiological mechanisms [5], why our study population did not comprise low birthweight infants. Low birthweight is more common in groups of low SEP [23], and exclusion of low birthweight infants could thus have led to greater consistency in the relationship between birthweight, weight gain and COO across different SEP groups. It is possible that our results would appear differently if we had included low birthweight infants, and this should be assessed in future studies.

Weak signs of additive interaction in some combinations of SEP and rapid or very rapid weight gain on COO risk were identified, but the results as a whole does show any signs of COO risk increase in any SEP groups after rapid or very rapid weight gain. Thus, there does not currently exist suffiecient evidence to suggest that promotion of healthy infant weight gain as a part of an early-life COO prevention strategy should be targeted certain socioeconomic groups, but future research should pursue the signs of additive interaction identified in our study. Anyhow, the social gradient in the prevalence of rapid and very rapid weight gain highlights that such prevention should make use of approaches that are effective in populations of low SEP. Access to resources such as money, knowledge, and power affects individuals’ ability to achieve behaviour change, and a lack of access to these resources can make individual focused behaviour change interventions less successful in populations of low SEP compared to those of high SEP [13, 73]. Conversely, interventions that integrate an intend to make changes at more upstream levels like the community, institutional, or policy level may be better strategic choices, as such interventions are more successful in reducing social inequalities in health [74, 75].