Spatio-temporal trends of the age-at-menarche percentiles among Portuguese women since 1920

Menarche, the term to denote the onset of menstruation, marks the beginning of a girl’s fertility. Later age at menarche (AaM) began to occur in the Modern Era, especially after the Industrial Revolution (1760–1840) possibly due to the deterioration of certain living conditions. Estimates for the second half of the 18th century point to an AaM between 15 and 16 years old [1]. This behaviour was found to undergo a reversal in the 20th century, with most studies showing a trend for earlier AaM in both developed and developing countries [2‐8]. This feature has been associated with life-style, environment, socio-economic status, access to health care and higher levels of literacy [9]. Additionally, several studies point to an association with the consumption of energy-rich food (i.e. with an increased body mass index) [10, 11]. The work of Cabanes et al. 2009 [12], which studied women born from 1925 to 1962, along with Sohn et al. 2017 [13], studying women born in the period 1941 – 1992, noted some rates of decline in AaM. The European Prospective Investigation into Cancer & Nutrition (EPIC) study found that mean AaM decreased among female participants born from 1912 to 1964 in nine European countries [14]. However, in the past few decades, this trend is disappearing, with some studies showing a cessation of the decline [1]. In France, Lalys et al. [15] studied a group of school girls born between 1979 and 1994, and in Netherlands Talma et al. [16] studied girls born before the year 2000. Both studies concluded that there are signs of slowing down or stabilization. Finer and Philbin (2014) [17] report a mean AaM among US girls born in 1993 of 12.3 years of age – similar to those born in 1980.

The Portuguese panorama is identical to those already described above. A work [18] that analyses a Portuguese community of students at the University of Coimbra born in the period 1972–1983 claims that the mean AaM for girls born in 1983 was 12.03 years and that the place of residence during childhood and adolescence had a significant effect on the mean AaM. Girls from rural areas had a later menarche when compared to those who spent their childhood/adolescence in urban areas. Despite being focused on a specific-group, this work offers an advantage in our study, in analysing a time-period immediately following that considered in our study \((1920-1973)\), thus becoming an interesting point of comparison.

Age at menarche plays a very important role in the research about breast cancer, hence having important individual and public health implications. Early puberty has been associated with an increased risk of this type of cancer, but also with obesity and diabetes [19]. Burgess et al. [20] suggested a protective effect of later AaM on breast cancer risk.

It is well known that AaM varies between and within populations and is influenced by many factors. It may be delayed under poor socio-economic and health conditions or accelerated by residence in an urban community [21]. The starting point for the present research was the idea to understand the spatio-temporal evolution of menarche in the central region of Portugal. To achieve this goal we analysed the dataset on the Breast Cancer Screening Program provided by the Portuguese Cancer League (Liga Portuguesa Contra o Cancro – LPCC) in the Central region of Portugal [22].

A typical regression analysis, in most contexts, focuses on explaining the expected value dependency of a response variable as a function of a set of explanatory variables. However in some situations, if we want to understand not only the behaviour of the average person, but also the behaviour of those belonging to the extremes of the population, we might consider percentile regression [23, 24] which allows us to go beyond mean regression, enabling building regression curves for the percentiles, instead of the mean of a response variable. Rigby et al. [25] argue that for data with more than 1000 observations, regression models beyond the mean should be the norm, not the exception.

The study of the AaM percentiles in the central region of Portugal was carried out considering the full dataset on the Breast Cancer Screening Program provided by the Portuguese Cancer League (LPCC) in the Central region of Portugal in the period \(1990-2019\) which has screened \(N=452\,348\) women born in the period \(1920-1973\). At the age of 45 (since 2010 the age is 50) all women in each of the 89 municipalities (see Fig. 3) were invited to undergo a free screening mammogram and every two years thereafter until the age of 69. These regions roughly represent 25% of the Portuguese population. Although we must add a caveat here, because this spatial information is regarding only the current place of residence, and no other spatial information such as place of birth or where women grew up as a child is known. More details about the screening program and the inclusion criteria are given elsewhere [22, 26, 27].

Age at menarche (AaM) registered in years at the first interview, was the response variable of interest. The average was 13.18 with a minimum of 8 and a maximum of 24. We could raise the question about the existence of bias in the reported ages, but many studies have demonstrated that this reported age is quite accurate, most likely due to the emotional significance of menarche for a young girl [28]. Figure 1 shows the histogram for the unconditional distribution of this variable with a modal class at 13 years old. Summarized in Table 1 are its descriptive percentiles by decade from 1920 to 1973, and Table 2 depicts the mean and the median by municipality and decade. Year of birth (Byear) and the demographic information given by the municipality of residence (Muni) were the independent variables.

A typical regression analysis has the advantage of being very well known to the users, although it is very likely that in many scenarios other properties of the response distribution (e.g. the variance) may also depend on covariates. Besides this, one may want to comprehend not only the behaviour of the average person, but the behaviour of those belonging to the population’s extremes. To account for these features, we will consider a statistical framework developed within the context of Generalized Additive Models for Location, Scale and Shape (GAMLSS) [29], also known as distributional regression models [30]. These models have several advantages, but for our study the principal characteristic is that it assumes a known parametric family of distributions for the response variable which allows an easy calculation of the conditional percentile curves of the AaM given birth year and municipality of residence. Furthermore, the covariate effects of interest can be of flexible forms (e.g. smoothing functions) and not restricted to the traditional, and perhaps unrealistic, linear effect. At the same time, it ensures that the adjusted percentile curves do not cross. A competing method is quantile regression [31], where the adjusted curves might cross, because it does not assume a distribution for the response variable and, therefore, can be considered in the realm of non-parametric methods [25]. For their part, other authors have been considering other approaches. Most use simple statistical methods like ANOVA [32] or linear regression [12]. Chumlea et al. [33] considered a probit analysis.

The conditional percentile values were thus obtained via the R software (v 4.1.0), namely using the main package gamlss (version 5.1-4) and two additional packages that provide a set of functions to fit models with spatial variables (gamlss.spatial and gamlss.add).

The chosen model is written below in terms of the R syntax based on the gamlss package:

Estimation of the percentile curves for a response variable is widely used in medicine for checking whether an individual has an abnormally low or high value of the response variable (given the covariates of interest), and hence whether she/he is potentially at risk. For example, let us think of a region where breast cancer risk is high and that girls living there tend to have an early menarche when compared to the rest of the country. And suppose that all other characteristics generally known for being linked to breast cancer are equal across that country. That specific region may be the target of differentiated public policies.

The dataset here analysed is based on the largest sample size reported to date in Portugal, although it does not include women born in the last three decades, meaning that the percentiles of these youngest women cannot be ascertained. Nevertheless, and to the best of our knowledge, this study is the first to analyse and produce percentile curves for the women residing in the Central Portugal since 1920 and to use a distributional regression approach.

Probably not all spatial effects are being correctly captured because the spatial information in the data set regards the place of residence at the time of the participation in the screening program, i.e. when the women are adults, as already mentioned in the data description. However, we assume that these women have always been living there. This is more likely to be true for the interior areas than for the areas closest to the coast given that municipalities in Central and Eastern Portugal have lower levels of wealth. Despite the above, we do believe that the spatial results are quite robust because of the dimension of the sample here analysed. That said, some works, e.g. [36] show that the place of residence, and even the altitude of the residence area [37] during childhood, are likely to be more effective in estimating regional differences for the AaM.

The probability distribution that we chose for describing the AaM, \(\text {BCCG}(\mu ,\sigma ,\nu )\), has several advantages for have being considered here. First its location parameter, \(\mu\), is interpreted as the median of the distribution, which when describing the percentiles is a benefit. Additionally, we are free to control the shape parameter, \(\nu\), representing the amount of skewness of the distribution, and from Fig. 1 we know that our distribution is positive skewed.

We resorted to a distributional regression model to estimate several spatio-temporal percentiles for the AaM in the central region of Portugal considering a long period of time and using a representative sample with a large number of observations. The results show that within the period \(1920-1973\), all the percentiles obtained for the AaM depict a decreasing trend. For a girl born in 1920, we would expect a median AaM of around 14 years. For a girl born in 1973, this value is about 1.5 years lower (Fig. 2). Interestingly, a pattern of AaM distribution by municipality emerges. Although all its percentiles are decreasing, the regions with the highest percentile values for AaM in 1920 are the same for 1970 (Figs. 3 and 4), with only a few exceptions.

In the middle of the 20th century the decreasing rate achieved its maximum of around 15.23 days per year (Table 3). This trend after 1973 seems to be slowing down or even stopping. We should note that the median’s rate of decrease is not replicated by the other percentiles. For instance, percentiles 90% and 95% decreased about 2.6 years in the span of 50 years. On the other hand, percentiles 5% and 10% decreased only about 0.8 years in the same range of 50 years. Early menarches (below percentile 5%) occur in the coastal (western) area of Portugal’s central region for all the decades considered and later menarches (above percentile 95%) occur in the central-north area.