Association of Long-Duration Breastfeeding and Dental Caries Estimated with Marginal Structural Models

Abstract

Purpose

Estimate the association between breastfeeding ≥24 months and severe early childhood caries (ECC).

Methods

Within a birth cohort (n=715) from low-income families in Porto Alegre, Brazil, the age 38-month prevalence of severe-ECC (≥4 affected tooth surfaces or ≥1 affected maxillary anterior teeth) was compared over breastfeeding duration categories using marginal structural models to account for time-dependent confounding by other feeding habits and child growth. Additional analyses assessed whether daily breastfeeding frequency modified the association of breastfeeding duration and severe-ECC. Multiple imputation and censoring weights were used to address incomplete covariate information and missing outcomes, respectively. Confidence intervals (CI) were estimated using bootstrap re-sampling.

Results

Breastfeeding ≥24 months was associated with the highest adjusted population-average severe-ECC prevalence (0.45, 95% CI: 0.36, 0.54) compared with breastfeeding <6 months (0.22, 95% CI: 0.15, 0.28), 6–11 months (0.38, 95% CI: 0.25, 0.53), or 12–23 months (0.39, 95% CI: 0.20, 0.56). High frequency breastfeeding enhanced the association between long-duration breastfeeding and caries (excess prevalence due to interaction: 0.13, 80% CI: −0.03, 0.30).

Conclusions

In this population, breastfeeding ≥24 months, particularly if frequent, was associated with severe-ECC. Dental health should be one consideration, among many, in evaluating health outcomes associated with breastfeeding ≥24 months.

Introduction

The World Health Organization (WHO) recommends continued breastfeeding up to age 2 years or beyond [1], and failure to breastfeed is associated with poor health consequences for both mother and child [2,3]. However, the nature of the relationship between dental caries and the age to which children are breastfed remains uncertain. Caries is among the most common diseases worldwide and often goes untreated, particularly in low-resource settings [4–6], with negative quality of life implications [7]. Some laboratory models suggest that human milk can cause caries [8,9], particularly in combination with added sugars [10], while some report no demineralization of tooth material by human milk alone [11]. The epidemiological literature [12] includes studies that support a positive association between long-duration breastfeeding and early childhood caries (ECC) [13–16] and others that do not [17,18].

Breastfeeding timing relative to other feeding habits complicates study of breastfeeding duration and ECC. Early breastfeeding cessation might accelerate the introduction of particular foods [19,20], and the foods consumed early in life likely influence caries development [21–23]. In turn, early-life food experiences might also influence the duration to which a breastfeeding child continues nursing [19]. Regression modeling is problematic in the presence of such time-dependent confounding, in which a variable (e.g. early-life food experiences) can be part of a causal pathway between an earlier aspect of exposure (e.g. early breastfeeding) and the outcome, while simultaneously operating as confounder with respect to a later aspect of exposure (e.g. continued breastfeeding). Marginal structural models (MSMs), in contrast, have been used to make causal inference from observational data in the presence of time-varying covariates [24–28]. Such techniques are particularly relevant for exposures, such as breastfeeding, that cannot be easily assigned as a randomized intervention.

We aimed to estimate the association between long-duration breastfeeding (≥24 months) and severe-ECC (S-ECC) in a birth cohort of urban, low-income Brazilian children. We hypothesized that long-duration breastfeeding is associated with greater caries occurrence. We secondarily hypothesized that the association between long-duration breastfeeding and S-ECC is stronger if daily breastfeeding episodes are more frequent.

Methods

Participants

We followed a birth cohort nested in a cluster-randomized trial in Porto Alegre, Brazil. The community water supply is optimally fluoridated [29], and 52 public healthcare centers provide primary medical services predominantly to low-income residents. A stratified random sample (n=20 health centers) was selected from 31 eligible clinics for participation in the original trial of healthcare worker training [30,31].

In 2008, 715 of 736 eligible pregnant women with appointments at participating clinics agreed to enroll their children in a cohort to track health outcomes (Figure 1). The trial had provided intervention clinics with healthcare worker training that promoted healthful infant complementary feeding for incorporation into maternal counseling. After 3 years, the intervention did not extend the total duration of breastfeeding (hazard ratio for breastfeeding cessation: 0.94, 95% confidence interval: 0.79, 1.11), although the mean duration of exclusive breastfeeding was increased [30]. S-ECC was not lowered significantly among children born to intervention group clinic attendees [31].

Figure 1. Flow of Participants.

Pregnant women were recruited from 20 municipal health centers in the city of Porto Alegre, Brazil and followed to a mean child age of 38 months.

Baseline variables

Trained fieldworkers collected baseline (during pregnancy) socio-demographic information via structured questionnaires. Data included maternal age, household size, maternal education (≤8 years), maternal smoking (current vs. never/former smoker), indoor bathroom (yes/no), city region (indicators for eight geo-administrative districts), parity (first child yes/no), maternal partner status (married or partnered vs. single, separated, or widowed), household income (≤1500 Brazilian reais monthly; approximately 900 US dollars in 2008), outside income source (e.g. government support), social class (Brazilian Association of Economic Research Institutes classification ≤C), low body mass index (BMI) (≤18, based on measured height and self-reported pre-pregnancy weight). Child sex and birth date were collected at age 5–9 months.

Time-varying behaviors and anthropometry

Infant growth and feeding habits were recorded at each of three home visits, corresponding to mean ages of 6 months (range: 5–9), 12 months (range: 11–15), and 38 months (range: 31–46). Infant length and child height were collected following standard protocol and converted to height/length-for-age Z-scores using WHO standards [32]. At each visit, mothers were asked whether they had ever breastfed and whether they were currently breastfeeding. Breastfeeding duration represented the age to which any breastfeeding continued, regardless of complementary feeding. Breastfeeding mothers were asked how frequently they nursed daily (0, 1, 2–3, or “many times,” separately for day and night). Mothers no longer breastfeeding were asked at what age (in months) breastfeeding ceased.

At the 6-month assessment, the number of feeding bottles consumed in the preceding day was recorded (later categorized 0, 1–3, ≥4). Sugar in the bottle corresponded to consuming ≥1 bottle containing any sweet additive: table sugar, powdered or liquid artificial chocolate, soft drinks, or powdered artificial juice. Questionnaires addressed use of commercially prepared infant formula and the age of introduction of 32 specific foods (e.g. fruits, beans, soft drinks, candies). At the 12-month assessment, the questionnaire posed whether 29 specific items were consumed in the previous month and the weekly consumption frequency of 5 complementary foods (fruits, vegetables, beans, meats, organ meats). Two feeding indices measured dietary patterns to account for foods consumed in combination and to increase the efficiency of the analysis [33]. The indices were created specifically for this analysis due to a lack of existing diet indices specific to cariogenic feeding behaviors in comparable populations. The first, referred to here as the food introduction index, was the count of low nutrient-density and/or presumably cariogenic foods introduced before age 6 months: added sugar, candy, chips, chocolate, chocolate milk, cookies, fruit-flavored drink, gelatin, honey, ice cream, soft drinks, and sweet biscuits. The second, termed here as the first-year feeding index, summed the food introduction index with the count of the following foods recorded at the 12-month assessment: added sugar in a drink, candy, cake, chips, chocolate, chocolate milk, cookies, creamed caramel, fruit-flavored drink, gelatin, honey, ice cream, other confection, soft drinks, and sweet biscuits.

At the 38-month assessment, data were collected regarding bottle use, height-for-age Z-scores, and tooth brushing with fluoride dentifrice. While these variables are likely associated with S-ECC, we did not consider them confounders, because our cut-point for defining the exposure (breastfeeding ≥24 months) temporally preceded these measures. However, we estimated separate models that included these variables as a sensitivity check.

Dental caries

Dental status was evaluated at 38 months following WHO protocol [34], with non-cavitated (white-spot) lesions also recorded. Assessments took place in participants' homes, aided by a lighted intraoral mirror. Teeth were brushed and dried with gauze. Severe-ECC was defined as ≥1 affected maxillary anterior teeth or ≥4 decayed, missing due to caries, or filled tooth surfaces (dmfs ≥4) [35]. Two dentist-examiners completed the evaluations following identical protocol (inter-rater unweighted kappa=0.75; intra-rater unweighted kappa=0.83 for both examiners).

Statistical methods

The proportion of children with S-ECC was compared across four breastfeeding duration categories: <6 months, 6–11 months, 12–23 months, and ≥24 months. Three marginal structural models were fit. The weights for estimating unadjusted models incorporated only clinic allocation status to account for the nested study design. Adjusted models additionally accounted for baseline socio-demographic variables: maternal age, education, parity, pre-pregnancy BMI, smoking status, social class, and child age and sex. Fully-adjusted models included those variables, as well as time-varying bottle use, feeding habits, and length-for-age Z-scores (see below).

In estimating MSMs, inverse probability weighting was used to generate a “pseudo-population” representative of a hypothetical population in which breastfeeding duration (A) had been allocated independently of confounding variables. Weights were assigned inversely to the predicted probability of observed exposure, given baseline characteristics (W) and longitudinally recorded variables (L_t), giving the greatest weight to observations with exposure and confounder combinations least represented in the sample, relative to what would have been observed under random exposure allocation. Figure 2 depicts the assumed relationships among variables as a directed acyclic graph. Exposure probabilities were estimated using Super Learner, a data-adaptive machine-learning tool [36].

Figure 2. Directed acyclic graph.

The graph depicts the assumed relationships between study variables. W₀ = baseline socio-demographic characteristics; A_t = breastfeeding at time t months; L_t = time-varying feeding behaviors and anthropometry at time t months; Y₃₈ = severe early childhood caries at 38-months. Time-dependent confounding occurs through L₁₂ variables, which are part of a directed path from A₆ to Y₃₈ but a back-door path from A₂₄ to Y₃₈.

To account for time-dependent confounding, weights were based on treatment models for three probabilities: the probability of breastfeeding at 6 months (Pr[A₆=1]); the probability of breastfeeding at 12 months, given breastfeeding at 6 months (Pr[A₁₂=1 | A₆=1]); and the probability of breastfeeding at 24 months, given breastfeeding at 12 months (Pr[A₂₄=1 | A₁₂=1]). Each treatment model was estimated while incorporating temporally appropriate putative confounders: in fully-adjusted models, the 6-month treatment model included clinic allocation status and baseline socio-demographic variables only; the 12-month treatment model included these variables and added the 6-month bottle use variables, formula use, food introduction index, and 6-month length-for-age Z-scores; the 24-month treatment model replaced the food introduction index and 6-month Z-scores with the first-year feeding index and 12-month Z-scores, respectively, and added complementary food frequency. To stabilize the weights, we multiplied by the marginal probability of the observed exposure category [25].

Equation 1 gives the stabilized treatment weights, where indicators (I) take a value of 1 when the exposure category was observed and 0 otherwise.

$$S{W}^{\tau }-\frac{\text{Pr}[A<6\mathit{\text{months}}]I[A<6\mathit{\text{months}}]}{\text{Pr}[A_6-1|W]}+\frac{\text{Pr}[A-6–11\mathit{\text{months}}]I[A-6–11\mathit{\text{months}}]}{\text{Pr}[A_6-1|W]\text{Pr}[A_{12}-1|A_6-1,W,L_6]}+\frac{\text{Pr}[A-12–23\mathit{\text{months}}]I[A-12–23\mathit{\text{months}}]}{\text{Pr}[A_6-1|W]\text{Pr}[A_{12}-1|A_6-1,W,L_6]\text{Pr}[A_{24}-0|A_{12}-1,W,L_6,L_{12}]}+\frac{\text{Pr}[A\ge 24\mathit{\text{months}}]I[A\ge 24\mathit{\text{months}}]}{\text{Pr}[A_6-1|W]\text{Pr}[A_{12}-1|A_6-0,W,L_6]\text{Pr}[A_{24}-1|A_{12}-1,W,L_6,L_{12}]}$$ (1)

For each model, missing variables and missing or incomplete breastfeeding histories were multiply imputed from probabilities estimated using Super Learner, corresponding to 2.6% of data among children with observed outcomes. Censoring weights, equal to the inverse probability of having an observed outcome, given exposure and covariates, up-weighted observations most resembling those with missing outcomes. The probability of an observed outcome (Pr[C=0]) was estimated via Super Learner using predictor variables clinic allocation status, maternal age, education, partner status, parity, smoking in pregnancy, and pre-pregnancy BMI, household income, indoor bathroom, number of inhabitants, outside income source, city region, and social class, and child breastfeeding duration, first-year feeding index, height-for-age Z-score, and sex. Stabilized censoring weight numerators were the product of the probability of having outcome data, given breastfeeding duration category, and a 1/0 indicator for having outcome data (equation 2).

$$S{W}^C=\frac{\text{Pr}[C=0|A]I[C=0]}{\text{Pr}[C=0|A,W,L_t]}$$ (2)

Final MSM weights were the product of stabilized treatment weights and stabilized censoring weights: SW = SW^T × SW^C. In the fully-adjusted model, non-zero stabilized weights had a mean of 0.997 (minimum: 0.28, maximum: 8.90).

For each model, point estimates (prevalence ratio, PR; prevalence difference, PD) were averaged over 200 multiple imputations. Percentile-based 95% confidence intervals (CI) were estimated as the 2.5 and 97.5 quantiles from 5000 bootstrap iterations to account for variance from sampling, imputation, and weighting. For comparison, an analogous complete-case regression analysis was completed using log-linear models and robust variance. Analyses were completed in R version 3.0.1 (http://www.r-project.org).

Secondarily, we examined whether frequent daytime breastfeeding (≥4 daily episodes) intensified the association of breastfeeding duration and S-ECC, restricting the analysis to children breastfed ≥6 months, the earliest age at which frequency data were collected. High frequency daytime breastfeeding and a long duration-high frequency interaction term were included as MSM covariates, and frequent breastfeeding was added to the treatment models. We defined the excess prevalence due to interaction (EPI) as a departure from additivity, following an example proposed for the relative risk [37]. If D and F represent the presence of long-duration and high frequency breastfeeding, respectively, and D̅ and F̅ the absence of these factors, then the EPI = PD(DF) – PD(DF̅) – PD(D̅F). As a sensitivity check, we also estimated models in which frequency strata were defined by high frequency breastfeeding in either the day or night, versus high frequency breastfeeding in neither. Nighttime breastfeeding was not used alone to define strata, because high frequency daytime breastfeeding was common (>50%) when nighttime high frequency breastfeeding was absent. Because tests for statistical interaction may have low power [38], 80% confidence intervals were provided.

Ethics

This study proposal received ethical approval from committees at the Federal University of Health Sciences of Porto Alegre (UFCSPA) and the University of California Berkeley and is in accordance with the Declaration of Helsinki. Informed consent was reached with mothers on behalf of their children. Children with caries or suspected anemia, under-nutrition, or overweight status were referred to their local health center.

Results

Table 1 demonstrates selected characteristics of the study population. The fraction of mothers interviewed at the 6-month assessment to report initiating any breastfeeding was 0.99 (627/633); the fraction who breastfed to 12 months was 0.47 (282/598). Exclusive breastfeeding continued to mean age 2.1 months. Nearly half the children were introduced to commercially prepared infant formula by 6 months (0.49, 309/632), but few children used formula at 12 months (0.03, 18/539). Bottle use and soft drink consumption were common, while inadequate length-for-age was rare (Table 1). S-ECC prevalence at 38 months was 0.34 (157/458); the prevalence of at least one affected tooth was 0.55 (250/458). Caries was most common among children breastfed for ≥24 months (Figure 3).

Table 1.

Characteristics of Participants

Characteristic			Number of observations1
Socio-demographic characteristics
	Maternal age at expected delivery date, mean (SD), years	26.0 (6.7)	715
	Mother has ≤ 8 y of formal education, n (%)	340 (47.6)	715
	Household income ≤ 3 times minimum salary2, n (%)	565 (81.9)	690
	Social class C or lower by ABIPEME index3, n (%)	569 (79.8)	713
	Self-identified maternal race white, n (%)	395 (55.2)	715
	Self-identified maternal race black, mixed, or other, n (%)	320 (44.8)	715
	Male child, n (%)	333 (52.4)	635
Anthropometry
	Length-for-age Z-score at age 5–9 months, mean (SD)	−0.13 (1.2)	631
	Length-for-age Z-score <−2 at age 5–9 months, n (%)	31 (4.9)	631
	Length-for-age Z-score at age 11–15 months, mean (SD)	−0.03 (0.9)	527
	Length-for-age Z-score <−2 at age 11–15 months, n (%)	4 (0.8)	527
Feeding Habits
	Introduced to soft drinks before age 6 months, n (%)	192 (30.3)	633
	Introduced to any sweets before age 6 months, n (%)	557 (90.8)	613
	Consumed soft drinks in prior month at age 11–15 months, n (%)	413 (76.6)	539
	Consumed vegetables ≥4 days per week at age 11–15 months, n (%)	341 (63.5)	537
	Ever initiated breastfeeding	627 (98.9)	633
	Duration exclusive breastfeeding, mean (SD), months	2.1 (1.6)	633
	Exclusive breastfeeding to age ≥4 months, n (%)	152 (24.0)	633
	Breastfeeding duration to age <6 months, n (%)	216 (34.1)	633
	Breastfeeding duration to age 6–11 months, n (%)	100 (16.7)	598
	Breastfeeding duration to age 12–23 months, n (%)	65 (12.1)	537
	Breastfeeding duration to age ≥24 months, n (%)	156 (29.1)	537
	Consuming sweet substances in bottle at age 5–9 months, n (%)	198 (32.3)	614
	Consuming sweet substances in bottle at age 2–3 years, n (%)	312 (68.4)	456
Dental Caries Experience at age 2–3 years
	Any affected tooth, n (%)	250 (54.6)	458
	S-ECC, n (%)	157 (34.3)	458
	dmfs (any decay), mean (SD)	3.2 (6.1)	458
	dmfs (cavitated decay only), mean (SD)	2.6 (5.9)	458

Abbreviations: SD, standard deviation; ABIPEME, Brazilian Association of Economic Research Institutes; S-ECC, severe early childhood caries; dmfs, decayed missing filled surfaces index

¹Number of observations differ for some variables due to missing data and/or losses to follow-up

²Monthly income of ≤1500 Brazilian reais; approximately 900 US dollars in 2008

³Socioeconomic classification scale based on material possessions and education, A = highest status, E = lowest status

Figure 3. Observed Prevalence of Severe Early Childhood Caries at Age 38 Months by Categories of Breastfeeding Duration.

The unadjusted (crude) prevalence of severe early childhood caries by categories of breastfeeding duration is shown for the 439 children with complete observed data for both breastfeeding duration and dental health.

The highest S-ECC prevalence was associated with breastfeeding ≥24 months in all marginal structural models (Table 2). As a sensitivity check, including 38-month variables (bottle use, fluoride toothpaste, height-for-age) in the fully adjusted model did not appreciably alter estimates; for example, the prevalence ratio comparing breastfeeding ≥24 months to <6mo changed to 2.11 (95% CI: 1.50, 3.30) from 2.10 (95% CI: 1.50, 3.25).

Table 2.

Unadjusted and Adjusted Associations of Breastfeeding Duration and Severe Early Childhood Caries in Preschoolers

	Marginal Prevalence1	95% CI	Prevalence Ratio	95% CI	Prevalence Difference	95% CI
Breastfeeding Duration
Unadjusted2 Model
<6 months (reference)	0.23	0.16, 0.30	1		0
6–11 months	0.38	0.27, 0.50	1.66	1.06, 2.57	0.15	0.02, 0.29
12–23 months	0.31	0.20, 0.43	1.35	0.81, 2.16	0.08	−0.05, 0.22
≥24 months	0.45	0.38, 0.53	1.98	1.44, 2.87	0.22	0.12, 0.33
Adjusted3 Model
<6 months (reference)	0.22	0.15, 0.28	1		0
6–11 months	0.39	0.27, 0.53	1.79	1.13, 2.80	0.17	0.03, 0.32
12–23 months	0.35	0.22, 0.49	1.59	0.97, 2.65	0.12	−0.01, 0.29
≥24 months	0.45	0.38, 0.53	2.06	1.51, 3.00	0.23	0.14, 0.34
Fully-Adjusted4 Model
<6 months (reference)	0.22	0.15, 0.28	1		0
6–11 months	0.38	0.25, 0.53	1.77	1.12, 2.85	0.17	0.03, 0.32
12–23 months	0.39	0.20, 0.56	1.82	0.85, 3.20	0.18	−0.03, 0.42
≥24 months	0.45	0.36, 0.54	2.10	1.50, 3.25	0.24	0.13, 0.36

Abbreviation: CI, confidence interval

¹Population-average prevalence of severe early childhood caries at given categories of breastfeeding duration, as estimated from marginal structural models

²Includes allocation status from nesting intervention study only

³Includes allocation status from nesting intervention study and maternal age, education, parity, pre-pregnancy BMI, smoking status, social class, and child age and sex

⁴Includes all adjusted model variables and time-varying bottle use variables, feeding habits, and height-for-age Z-scores

Compared to breastfeeding 6–23 months, breastfeeding ≥24 months was associated with elevated S-ECC prevalence, although not statistically significant (unadjusted PR: 1.31, 95% CI: 0.97, 1.79; adjusted PR: 1.22, 95% CI: 0.89, 1.66; fully-adjusted PR: 1.17, 95% CI: 0.85, 1.78). However, breastfeeding ≥24 months was more strongly associated with S-ECC when daytime breastfeeding was frequent (fully-adjusted PR: 1.38, 95% CI: 0.97, 2.16) (Table 3). The EPI was 0.13 (80% CI: −0.03, 0.30), suggesting positive interaction between frequent daytime nursing and breastfeeding ≥24 months. Results were similar defining high frequency based on frequent nursing in either the day or night versus neither day or night (fully-adjusted PR with frequent breastfeeding: 1.43, 95% CI: 1.01, 2.18); the EPI increased to 0.23 (80% CI: 0.03, 0.41).

Table 3.

Unadjusted and Adjusted Associations of Breastfeeding ≥24 Months and Severe Early Childhood Caries, Stratified by Frequency of Daytime Breastfeeding

	Marginal Prevalence1	95% CI	Prevalence Ratio	95% CI	EPI	80% CI
Breastfeeding Duration and Frequency
Unadjusted2 Model
Duration 6–23 months and low frequency	0.38	0.25, 0.51	1
Duration ≥24 months and low frequency	0.37	0.22, 0.52	0.97	0.53, 1.68
Duration 6–23 months and high frequency	0.31	0.22, 0.42	1
Duration ≥24 months and high frequency	0.48	0.39, 0.57	1.53	1.06, 2.30	0.18	−0.07, 0.43
Adjusted3 Model
Duration 6–23 months and low frequency	0.38	0.25, 0.51	1
Duration ≥24 months and low frequency	0.36	0.20, 0.53	0.94	0.51, 1.67
Duration 6–23 months and high frequency	0.33	0.23, 0.44	1
Duration ≥24 months and high frequency	0.47	0.38, 0.56	1.42	0.99, 2.12	0.16	−0.10, 0.41
Fully-Adjusted4 Model
Duration 6–23 months and low frequency	0.36	0.24, 0.49	1
Duration ≥24 months and low frequency	0.37	0.20, 0.55	1.01	0.52, 1.81
Duration 6–23 months and high frequency	0.35	0.23, 0.45	1
Duration ≥24 months and high frequency	0.48	0.38, 0.58	1.38	0.97, 2.16	0.13	−0.03, 0.30

Abbreviations: CI, confidence interval; EPI, excess prevalence due to interaction

¹Population-average prevalence of severe early childhood caries at given categories of breastfeeding duration, as estimated from marginal structural models

²Includes allocation status from nesting intervention study only

³Includes allocation status from nesting intervention study and maternal age, education, parity, pre-pregnancy BMI, smoking status, social class, and child age and sex

⁴Includes all adjusted model variables and time-varying bottle use variables, feeding habits, and height-for-age Z-scores

Complete-case regression analysis yielded modest differences in estimates (Table 4). However, findings were qualitatively consistent with MSM results, supporting a positive association between S-ECC and breastfeeding ≥24 months.

Table 4.

Unadjusted and Adjusted Associations from Regression Models of Breastfeeding Duration and Severe Early Childhood Caries

	Unadjusted Model n = 439		Adjusted Model n = 422		Fully Adjusted Model n = 338
Model Variables	Prevalence Ratio	95% CI	Prevalence Ratio	95% CI	Prevalence Ratio	95% CI
Breastfeeding <6 months (reference)	1		1		1
Breastfeeding 6–11 months	1.45	0.93, 2.23	1.52	0.99, 2.33	1.45	0.83, 2.53
Breastfeeding 12–23 months	1.28	0.80, 2.05	1.44	0.89, 2.32	1.39	0.73, 2.64
Breastfeeding ≥24 months	1.96	1.40, 2.73	2.04	1.45, 2.85	1.85	1.11, 3.08
Clinic allocation (intervention)	0.85	0.65, 1.09	0.89	0.69, 1.15	0.95	0.71, 1.27
Maternal age (years)			0.98	0.96, 1.01	0.98	0.96, 1.01
Maternal education (≤8 years)			1.24	0.93, 1.65	1.35	0.97, 1.89
Maternal smoking (current)			1.49	1.13, 1.95	1.12	0.81, 1.55
Parity (has previous child)			1.18	0.85, 1.63	1.22	0.87, 1.71
Social class (C or lower)			1.09	0.77, 1.54	1.06	0.74, 1.52
Pre-pregnancy BMI ≤18			1.43	1.06, 1.93	1.44	1.03, 2.01
Child age at dental assessment (years)			1.25	0.64, 2.42	0.99	0.46, 2.12
Child sex (male)			1.18	0.91, 1.53	1.34	1.01, 1.77
Length-for-age Zscore at 11–15 months (per SD)					1.05	0.90, 1.23
First-year feeding index (per unit)					1.05	1.01, 1.09
Daily bottles at 5–9 months (1–3)					0.62	0.38, 1.02
Daily bottles at 5–9 months (≥ 4)					0.84	0.47, 1.52
Added sugar in bottle at 5–9 months					1.46	0.94, 2.26
Ever formula fed					0.82	0.59, 1.14
Frequency of fruits at 11–15 months					0.95	0.90, 1.01
Frequency of vegetables at 11–15 months					1.05	0.98, 1.12
Frequency of beans at 11–15 months					1.02	0.94, 1.10
Frequency of meat at 11–15 months					1.04	0.97, 1.12
Frequency of organ meat at 11–15 months					1.17	0.98, 1.39

Abbreviation: CI, confidence interval; BMI, body mass index; SD, standard deviation

Discussion

In this population of low-income Brazilian families, we estimated an increase in severe early childhood caries prevalence with breastfeeding 24 months or beyond. While the overall health benefits of breastfeeding are considerable, this work adds evidence that, in some contexts, very extended and frequent breastfeeding might increase caries risk. In addition to exposing teeth to bacterially fermentable milk sugars, prolonged breastfeeding might enhance the fidelity with which caries-causing oral bacteria are transmitted from mothers [39].

Several investigations have reported positive associations between breastfeeding duration and caries when using cut-points exceeding 18 months to define the uppermost duration category [13–15,40,41]. Studies that reported no association between caries and breastfeeding generally used earlier cut-points to define long-duration breastfeeding (i.e. ≥8 or ≥13 months) and have featured populations in which breastfeeding to age 2 years is uncommon, such as in Germany [42], Italy [17], and the United States [18]. A large hospital-based breastfeeding promotion intervention in Belarus did not affect caries prevalence at age 6 years [43]; however, the study did not directly compare caries outcomes among children who were or were not breastfed for extended durations (e.g. ≥24 months), which was an uncommon behavior in that trial population.

Daily breastfeeding frequency was associated with S-ECC in a previous study of Brazilian preschoolers, in which breastfeeding frequency, but not duration ≥12 months, maintained statistical significance in multi-variable models [21]. A combined measure of breastfeeding duration and frequency was strongly associated with ECC in Mayanmar [44], and a measure of nighttime breastfeeding burden, which was based on frequency, was positively associated with ECC in Iran [45], although not statistically significant.

An important strength of this study was the longitudinal, prospective collection of feeding information. We use weighting estimators to respect the temporal sequence of exposure and covariate information, accounting for time-dependent confounding by early-life feeding habits [46]. These methods have not been broadly used in oral health epidemiology.

Causal interpretation of marginal structural model results depends on unverifiable assumptions, specifically, positivity, exchangeability, and correct specification of the treatment models used to generate the weights [25,28]. In this analysis, the distribution of the weights was not extreme, important socioeconomic and nutritional confounders were prospectively collected, and Super Learner estimation reduced the reliance on parametric model-building assumptions [36]. These strengths give credence to causal interpretations, however, as with any observational study, residual confounding cannot be ruled out. For instance, our main estimates do not adjust for earlier oral hygiene habits. However, adjustment for 38-month toothbrushing habits did not affect estimates, and we have no evidence that oral hygiene habits would be associated with breastfeeding duration. Also, while we adjust for the age of introduction and weekly frequency of selected foods, not every aspect of the diet was recorded. Conservatively, we consider our estimates to represent meaningful associations, for which definitive causal claims await confirmation from future studies.

Losses to follow-up were relatively high but not unusual among cohort studies in low-resource settings, where participants frequently change address and contact information. Inverse probability censoring weights can account for losses, and, in this study, weighted estimates were similar to those from complete-case regression. However, losses remain a limitation, as strategies to account for missing data introduce additional assumptions [47]. EPI estimates were imprecise, because a relatively small number of children breastfed both infrequently and to long durations limited the statistical power to confirm interactions. Finally, this study population, featuring a relatively high prevalence of breastfeeding and of caries, might not be representative of the breastfeeding-caries relationship in all historical, geographical, and socioeconomic contexts.

A critical question raised by our findings is why breastfeeding, a normative and otherwise health-promoting human behavior, would be associated with deleterious dental outcomes. One possibility is that the mechanism through which repeated, prolonged exposure to human milk could enhance caries progression might operate differently under the near universal availability of highly refined sugars in the modern diet versus historically. Laboratory studies demonstrating greater cariogenic potential of human milk with the addition of outside sugars support this hypothesis [10,11]. Future research is recommended to better define the relationship between particular breastfeeding practices and dental caries in the context of a high-sugar food supply. Our results in no way suggest that breastfeeding itself be discouraged but are congruent with guidelines from professional dental organizations, which recommend supporting a mother’s decision to breastfeed but avoiding ad libitum breastfeeding after tooth eruption [48].

Abbreviations

BMI

body mass index

CI

confidence interval

ECC

early childhood caries

EPI

excess prevalence due to interaction

MSM

marginal structural model

PD

prevalence difference

PR

prevalence ratio

S-ECC

severe early childhood caries