Background
Colombia is a lower-middle income country that faces the challenge of addressing health inequalities in a time of internal conflict and slow economic growth. Although Colombia fares well in regional comparisons, health inequalities in the country follow the same trends as for other countries in the region, where improvements in the average health status have been accompanied by greater relative health inequities [
1‐
3]. Despite substantial gains in sexual and reproductive health in the country, e.g. the total fertility rate fell from 3.6 children per woman in 1986 to 2.4 in 2005, average fertility rates mask important within country inequalities in fertility rates: urban (2.3) versus rural (3.8), highest (1.5) versus lowest level of education (4.0), richest (1.7) versus poorest quintile (5.2) [
4].
Disparities among regions and among people of different socioeconomic position (SEP) are particularly relevant for understanding determinants of uptake of family planning [
5,
6]. Effective contraception is a close determinant of fertility and as such can contribute to reducing the burden of reproductive ill health, child mortality and morbidity. Previous studies have consistently found higher contraceptive use among women with higher levels of education and other related dimensions, such as women's empowerment and autonomy, in low and middle income countries worldwide [
7‐
12].
Addressing socioeconomic inequalities in health constitutes one of the main challenges for public health worldwide [
13]. Increasing evidence of large and widening inequalities has stimulated international efforts to understand and monitor socioeconomic inequalities in various dimensions of health [
14]. In low and middle income countries, these efforts include the task of developing measures of SEP in populations where data on income and expenditure have limitations in terms of availability, reliability and applicability [
15‐
17].
Household wealth is an alternative measure of SEP widely used in low and middle income countries [
17], broadly defined by asset ownership and housing quality. Wealth represents a more permanent economic status at household level than income or expenditure, because it takes into account available resources and long-run economic status [
16]. The World Bank Household wealth index (HWI) of the Demographic Household Surveys (DHS) includes a broad set of assets: durable consumer goods, housing quality, water and sanitary facilities and other amenities [
16]. This composite index is valuable, but captures a set of publicly provided as well as private household assets which is important to distinguish with respect to public health interventions [
14]. An alternative to this limitation is a multidimensional approach, in which different dimensions of SEP are defined separately, providing a framework for attempting to disentangle causal mechanisms responsible for inequalities in health [
18,
19].
In research on household wealth in Latin America, asset approaches include a wider portfolio of items in comparison to literature on assets for high income countries [
18‐
20]. In the latter, the term asset is assigned to material items with a market value, whilst in the region the term refers to tangible and intangible resources [
20,
21]. Similar categories of assets have been commonly grouped into domains of capital such as Human capital (e.g. level of education), Physical capital (e.g. floor materials, durable consumer goods) and Public capital (e.g. electricity, sewage) [
18,
22,
23]. Studies in Peru, Brazil and Colombia found that access to public assets has different effects depending on women's level of education. This interaction indicates that Human capital and Public capital may complement or substitute for each other [
22,
24,
25].
This interplay between different kinds of social inequality is a growing topic of interest in research on social inequalities in health. For example, Sen [
26] proposes going beyond a unidimensional analysis where the focus is given to only one conventional measure of social stratification e.g. social class or gender, and instead study how these dimensions interact with each other. Identifying these interactions and which dimensions of socioeconomic position are stronger determinants of contraceptive use may better target effective policy interventions in family planning [
25].
In this framework, the aim of this study was to examine socioeconomic inequalities in women's contraceptive use through the construction of measures that capture distinct dimensions of SEP: material circumstances (Physical capital), publicly provided assets (Public capital) and psychosocial and cognitive aspects particularly related to women's level of education (Human capital), in the DHS for Colombia of 2005. The underlying hypotheses were: a) socioeconomic inequalities in contraceptive use associated with Human capital will be larger than those by Physical capital, Public capital and the HWI, consistent with gender and health empowerment perspectives [
7‐
12] and b) provision of Public capital compensates for low levels of Human capital (women's level of education) [
22,
24,
25].
Methods
Data
The DHS are nationally-representative household surveys that provide data for a wide range of monitoring and impact evaluation indicators in the areas of population, health, and nutrition for low and middle income countries [
27]. We used the Colombia 2005 DHS (version 51), which is the most recent survey conducted in the country (October 2004-June 2005). A total of 41,344 (92% response rate) women from 37,211 households were interviewed (88.4% response rate). Women of fertile age (15-49) were selected for the initial study sample (N = 38,143) (Table
1). Current-non use of contraception was estimated for women 'exposed to the risk of pregnancy' defined as fecund women (not pregnant, amenorrheic or menopausal) in union (married/cohabiting) and women not in union but sexually active in the interview month (N = 20,023) and never use of contraception was restricted to women ever sexually active (N = 32,783) (Tables S2 and S3 additional file
1). 16% of the interviewed women had missing data for the reported sexual activity and were therefore excluded from the analysis. Compared to those with recorded data, women with this missing variable were more likely to be aged 15-19 (65% vs. 10%), have achieved secondary level of education (68% vs. 48%) and be single (96% vs. 38%). The data have been described elsewhere in more detail [
27,
28].
Table 1
Distribution of women 15-49 years old for each measure of SEP by place of residence 2005 Colombian DHS
| |
% (No.)
|
% (No.)
|
% (No.)
|
Mean age
| Years (SD) | 30.3 (10.0) | 30.0 (10.1) | 30.2 (10.1) |
Marital status
| Single | 50.2 (14731) | 37.8 (3325) | 47.3 (18056) |
| Married/cohabiting | 49.8 (14606) | 62.2 (5481) | 52.7 (20087) |
Children ever born
| <2 | 72.2 (21193) | 58.5 (5153) | 69.1 (26346) |
| 3-4 | 21.3 (6247) | 24.6 (2169) | 22.1 (8416) |
| >5 | 6.5 (1897) | 16.9 (1484) | 8.9 (3381) |
Household wealth (HWI)
| Richest | 19.1 (5588) | 1.1 (97) | 14.9 (5685) |
| Fourth | 24.2 (7102) | 3.3 (286) | 19.4 (7388) |
| Middle | 26.9 (7882) | 8.2 (726) | 22.6 (8608) |
| Second | 23.3 (6826) | 27.1 (2392) | 24.2 (9218) |
| Poorest | 6.6 (1939) | 60.2 (5305) | 19.0 (7244) |
Physical capital
| Richest | 19.7 (5768) | 2.2 (189) | 15.6 (5957) |
| Fourth | 23.7 (6954) | 5.6 (492) | 19.5 (7446) |
| Middle | 24.0 (7031) | 14.2 (1250) | 21.7 (8281) |
| Second | 20.8 (6097) | 28.3 (2489) | 22.5 (8586) |
| Poorest | 11.9 (3487) | 49.9 (4386) | 20.6 (7873) |
Public capital
| Richest | a | a | a |
| Fourth | 31.0 (9088) | 0.6 (55) | 24.0 (9143) |
| Third | 32.2 (9455) | 8.1 (713) | 26.7 (10168) |
| Second | 34.1 (9991) | 20.6 (1816) | 31.0 (11807) |
| Poorest | 2.7 (803) | 70.7 (6222) | 18.4 (7025) |
Human capital
| University | 21.9 (6414) | 4.8 (415) | 17.9 (6832) |
| Secondary | 54.8 (16084) | 37.8 (3327) | 50.9 (19411) |
| Primary | 21.3 (6257) | 50.2 (4422) | 28.0 (10679) |
| None | 2.0 (582) | 7.3 (639) | 3.2 (1221) |
Contraceptive behaviour
The outcomes of interest in this study were 'current non-use of contraception' and 'never use of contraception'. Women were asked if they were currently using any method to delay or avoid getting pregnant at or about the time of the survey (yes/no). In addition, women were asked about their knowledge of various contraceptive methods, and those who reported knowledge of a particular method of contraception were asked if they had ever used that method (yes/no). Only modern methods of contraception were considered for the analysis (oral contraceptive, intra-uterine devices, injections, diaphragm, male/female condom, male/female sterilization, implants, foam/jelly and lactation amenorrhoea).
A multidimensional approach to SEP
The exposures of interest were four different dimensions of SEP: the HWI, Physical capital, Public capital and Human capital (Table
2). Physical capital is measured using durable consumer goods and indicators of housing quality. Public capital is defined as access to services supplied by the state or on its behalf such as electricity and piped water. Its components capture connectedness to the public infrastructure and organisation. There are no missing values in the Colombian DHS 2005 for any of the asset variables included to construct the HWI, Physical and Public capital measures. The information on these variables was collected using the standard household questionnaire of the DHS. Respondents were asked to respond yes or no for each item listed e.g. do you have a refrigerator yes/no. The Human capital dimension is defined as women's educational attainment. Women reported their highest achieved level of education based on four categories (none/primary/secondary/university).
Table 2
Asset categories and components of the Household wealth index and Physical, Public and Human capital
Physical capital | Housing characteristics | Floor materials | X |
| | Wall materials | X |
| | Toilet inside/outside household | X |
| Durable consumer goods | Shower | X |
| | Phone | X |
| | Radio | X |
| | TV | X |
| | Fridge | X |
| | Blender | X |
| | Stereo | X |
| | Washing machine | X |
| | DVD | X |
| | Computer | X |
| | Electric/gas range | X |
| | Electric/gas oven | X |
| | Microwave | X |
| | Vacuum or Floor polisher | X |
| | Hot water heater | X |
| | AC | X |
| | VCR | X |
| | Motorcycle or scooter | X |
| | Car or truck | X |
| | Fan | X |
| Dwelling type | Self contained | X |
| | Apartment | X |
| | Rents in someone's home | X |
| | Rents in other type of building | X |
| | Other type | X |
| | No. of members per sleeping room | X |
| | Domestic worker | X |
Public capital | Publicly provided services | Electricity | X |
| | Aqueduct | X |
| | Private toilet connected to sewer | X |
| | Shared toilet connected to sewer | X |
| | Access to natural gas | X |
| | Waste collected by the government | X |
Human capital | Women's level of education | None | |
| | Primary | |
| | Secondary | |
| | University | |
Explanatory factors
Basic demographic characteristics identified in the literature [
5,
11] that could mediate women's contraceptive use over the course of the reproductive life cycle include urban/rural place of residence, women's age in years, health insurance (no/yes), marital status categorised into two groups (single/married-cohabiting) and number of children ever born categorised into three groups (<2/3-4/>5) (Table
1).
Construction of the measures of SEP
Each measure of SEP was constructed by applying the following steps: selection of the indicators (Table
2), coding of variables, calculation of weights, construction of an index using these weights, and classification of households into SEP groups (quintiles). In the case of Human capital with only one component variable, only the first two steps apply.
The weights used to construct the indices were derived through principal component analysis (PCA) using Filmer and Pritchett's approach [
29]. Most of the assets variables collected in the Colombian DHS 2005 were categorical variables. To include them in the PCA qualitative categorical variables were re-coded into binary variables (no/yes). To examine the distribution for each index, a histogram with kernel-density estimates was generated (Figure S1, Figure S2, Figure S3 additional file
1). Sample weights were not used for the PCA, but were used when constructing SEP quintiles for the HWI, Physical capital and Public capital measures. The stability of the categorization into quintiles was assessed by comparing household classification by Physical assets and Public assets against the HWI. The percentage of households that remained in the same quintile and those that moved one, two or three quintiles were calculated through cross-tabulation. Additionally, spearman rank correlations between the HWI and Physical and Public capital quintiles were computed.
To assess the internal coherence of the HWI, Physical capital and Public capital measures, the mean value for each specific item was compared between quintiles. Internal coherence was defined as the agreement of the distribution of assets and services across quintiles [
17]. For assessing reliability, the household sample was split into random halves to run the PCA of each asset based measure on each and factor loadings were visually compared (data not shown here can be found in [
30]).
Data analysis
Inequalities in non-use of modern contraceptives were measured using the Relative Index of Inequality (RII) [
31]. The RII is a commonly used measure in the study of social inequalities in health, which takes into account the size and relative position of the wealth groups. For the construction of RII the categories of each measure of SEP were hierarchically organized from the richest to the poorest quintile and in the case of the Human capital, from the highest level of education to no education level. Each measure of SEP is converted to a continuous distribution between 0 (highest SEP) and 1(lowest SEP). The distribution is weighted according to the population in each SEP group by calculating the midpoint of the proportion in each category e.g. if 10% of the sample were in the highest social group and 15% were in the next category, those in the highest would be given a value of 0.05 (0.10/2) and those in the next group would receive a value of 0.175 (0.10+0.15/2). The Relative Index of Inequality (RII) is obtained by regressing each measure of SEP on a binary outcome in a logistic regression model [
31]. A large score on the RII implies large socioeconomic inequalities for the outcome under study. Likelihood ratio tests were used to test the difference in RII by place of residence (urban/rural).
Crude and adjusted RII with 95% confidence intervals were estimated. The simplest model was adjusted for age by entering age in years as a continuous variable, marital status (single/married-cohabiting) and number of children ever born (<2/3-4/>5); subsequently it was stratified by place of residence (urban/rural). The final model was mutually adjusted for Physical capital, Public capital, and Human capital to estimate the independent contributions of each measure of SEP. The HWI was not included in this model in view of collinearity, as the Physical and Public capital items are contained in the HWI.
Likelihood ratio tests were used to test for interactions between Human capital, the weighted distribution of women's level of education and Public capital (Low/High) separately for urban and rural women and adjusted for age in years as a continuous variable, marital status (single/married-cohabiting), and number of children ever born (<2/3-4/>5). To show the effect of each level of education odds ratios were calculated using three groups (University-Secondary, Primary and None) due to insufficient data especially in rural areas where there is a lower proportion of women with higher levels of education. Public capital quartiles were divided into two halves (low/high). We then added Physical capital and health insurance (no/yes) and assessed their effect in the interaction model and tested for trends. Health insurance had no effect and was omitted from the model. We calculated the percent change in the coefficients for the effect of University/High compared to no education in models with and without adjustment for physical capital. All statistical analyses were conducted using Stata version 10.0 (Stata Inc., TX, USA).
Results
77% of the respondents lived in urban areas and 23% in rural areas (Table
1). There were large rural-urban inequalities in Physical, Human and Public capital, with rural women being poorer, less educated, and with less access to Public capital compared to urban women. In addition, inequalities in these forms of capital were larger within rural areas than within urban areas (Table
1).
The extent to which households could be distributed into population quintiles differed between the measures of SEP. It was possible to differentiate five groups for the HWI and Physical capital, and four groups for Public capital (Table
1). Compared to HWI quintiles, agreement was strongest for Physical capital quintiles (Spearman rank correlation 0.92) with 75% in the same quintile, 24% moving one quintile and 1% moving two or three quintile groups. Agreement was weaker with Public capital quintiles (Spearman rank correlation 0.72) with 46% in the same quintile, 35% moving one quintile and 19% moving two or three quintiles.
There was evidence of internal coherence for all measures of SEP when comparing the mean value for each asset variable by quintiles. For example, for the HWI 41% of the poorest households, 82% in the middle households and, 90% of the richest households had access to piped water. In the case of durable assets, a refrigerator was owned by 17% of the poorest households, 78% of the middle households and, 99% of the richest households. The proportion of households with dirt walls, earth/mud floors, and connected to septic systems decreased for each richer quintile. Spearman rank correlations between the Physical, Public and Human measures ranged from 0.08 to 0.41 in urban areas and 0.19 to 0.32 in rural areas. When the sample was split into random halves, the PCA loadings and direction of the loadings were similar in each half for each asset-based measure of SEP.
Socioeconomic inequalities in contraceptive use
Table
3 shows that the adjusted prevalence of current non-use and never use of contraception was respectively higher for women in rural areas (28%, 17%) compared to those in urban areas (24%, 9%) (difference p-values <0.001). Reported non-use generally decreased from the poorest to the richest for each measure of SEP with some small deviations, e.g. current non-use by HWI among women in the middle quintile (18%) compared to women in the fourth richer quintile (24%) in rural areas.
Table 3
Adjusted prevalence of modern contraceptive use and RII (95% CI) by each measure of SEP
Current non-use of contraception among women in
union (married/cohabiting) and single sexually active
| 23.7 | 28.2 |
N = 15147
|
N = 4876
| <0.001 |
Household wealth (HWI) | Richest | 19.8 | 13.7 | | | |
| Fourth | 21.0 | 23.7 | | | |
| Middle | 23.7 | 17.5 | 2.84 (2.41-3.35) | 4.6 (3.06-6.79) | <0.001 |
| Second | 28.0 | 23.4 | | | |
| Poorest | 30.4 | 33.5 | | | |
Physical capital | Richest | 21.0 | 16.9 | | | |
| Fourth | 21.1 | 20.3 | | | |
| Middle | 22.5 | 19.6 | 2.64 (2.26-3.07) | 3.96 (2.82-5.55) | <0.001 |
| Second | 28.0 | 26.7 | | | |
| Poorest | 29.6 | 33.9 | | | |
Public capital | Richest | a | a | | | |
| Fourth | 22.3 | 17.7 | | | |
| Third | 22.3 | 21.2 | 1.84 (1.56-2.17) | 2.01 (1.35-2.97) | 0.09 |
| Second | 26.2 | 26.9 | | | |
| Poorest | 30.8 | 29.7 | | | |
Human capital | University | 23.0 | 20.8 | | | |
| Secondary | 22.9 | 24.1 | | | |
| Primary | 25.3 | 28.6 | 1.44 (1.24-1.67) | 1.64 (1.14-2.37) | <0.001 |
| None | 25.2 | 51.4 | | | |
Never use of contraception among ever-sexually active women
| 8.9 | 17.2 |
N = 25231
|
N = 7552
| <0.001 |
Household wealth (HWI) | Richest | 4.9 | 3.8 | | | |
| Fourth | 7.3 | 7.2 | | | |
| Middle | 9.0 | 8.0 | 7.14 (5.94-8.59) | 23.5 [14.6-38.1] | <0.001 |
| Second | 10.9 | 12.5 | | | |
| Poorest | 18.5 | 22.8 | | | |
Physical capital | Richest | 5.3 | 5.3 | | | |
| Fourth | 7.4 | 6.9 | | | |
| Middle | 8.6 | 9.0 | 5.81 (4.89-6.91) | 19.63 [13.13-29.35] | <0.001 |
| Second | 10.6 | 14.5 | | | |
| Poorest | 15.2 | 24.1 | | | |
Public capital | Richest | a | a | | | |
| Fourth | 8.3 | 7.3 | | | |
| Third | 7.3 | 9.1 | 2.47 (2.04-2.99) | 7.72 [4.93-12.10] | <0.001 |
| Second | 10.6 | 12.3 | | | |
| Poorest | 15.4 | 20.2 | | | |
Human capital | University | 6.3 | 9.1 | | | |
| Secondary | 8.3 | 12.0 | | | |
| Primary | 12.3 | 19.7 | 4.83 (4.04-5.78) | 7.98 (5.27-12.10) | <0.001 |
| None | 22.3 | 43.8 | | | |
Urban and rural areas have different levels of inequality in contraceptive behaviour. There was evidence of inequalities for both outcomes with all four measures of SEP among women in urban and rural areas. Inequalities were significantly larger for women living in rural areas compared to women in urban areas for each measure of SEP (difference p-values < 0.001), except by Public capital for current non-use of contraception (Table
3). In urban areas, inequalities were wider by HWI for current non-use of contraception (RII 2.84 95% CI 2.41-3.35) and never use of contraception (RII 7.14 95% CI 5.94-8.59). In rural areas, there were large inequalities in contraceptive behaviour by Physical, Public and Human capital but inequalities were wider by HWI.
A model mutually adjusted for Physical, Public and Human capital and age in years, marital status and children ever born (Table
4) suggests that Physical capital tended to be the stronger socioeconomic determinant of contraceptive behaviour in urban and rural areas. Public and Human capital showed substantial and statistically significant inequalities for never use of contraception among women in rural areas.
Table 4
RII (95% CI) for current non-use and never use of modern contraceptive methods
Current non-use among women in union (married/cohabiting) and single sexually active
| | | | |
|
(N = 15147)
| |
(N = 4876)
| |
Physical capital | 2.37 (1.99-2.83) | <0.001 | 3.70 (2.57-5.33) | <0.001 |
Public capital | 1.28 (1.07-1.53) | 0.01 | 1.25 (0.82-1.89) | 0.30 |
Human capital | 1.05 (0.90-1.24) | 0.51 | 1.04 (0.71-1.54) | 0.83 |
Never use of contraception among ever-sexually active women
| | | | |
|
(N = 25231)
| |
(N = 7552)
| |
Physical capital | 3.70 (3.02-4.52) | <0.001 | 11.23 [7.36-17.16] | <0.001 |
Public capital | 1.19 (0.96-1.48) | 0.10 | 2.83 (1.76-4.54) | <0.001 |
Human capital | 3.02 (2.50-3.65) | <0.001 | 3.41 (2.24-5.21) | <0.001 |
Table
5 shows results for the hypothesis that provision of Public capital compensates for low levels of Human capital. There was a strong association between education and contraceptive use in households with low and high levels of Public capital, such that women with lower education reported higher non-use of contraception.
Table 5
Effect of women's level of education on modern contraceptive use in households with low and high provision of Public capital
| |
Current non-use of contraception among women in union (married/cohabiting) and single sexually active
|
| | N | | | N | | |
Low Public capital
|
University/Secondary
| 3806 | 1 | | 1497 | 1 | |
|
Primary
| 1541 | 1.18 (1.02-1.36) | 0.03 | 2607 | 1.09 (0.93-1.28) | 0.28 |
|
None
| 160 | 1.66 (1.14-2.42) | 0.01 | 338 | 2.53 (1.93-3.32) | <0.001 |
|
Trend
| | | 0.17 | | | <0.001 |
High Public capital
|
University/Secondary
| 7423 | 1 | | 243 | 1 | |
|
Primary
| 2092 | 1.06 (0.94-1.21) | 0.36 | 183 | 0.84 (0.52-1.37) | 0.48 |
|
None
| 125 | 1.10 (0.68-1.76) | 0.71 | 8 | 1.16 (0.22-6.01) | 0.86 |
|
Trend
| | | 0.16 | | | 0.54 |
|
Interaction
| | | 0.91 | | | <0.001 |
| | |
(N = 25231)
| | |
(N = 7552)
| |
| |
Never use of contraception among ever sexually active women
|
| | N | | | N | | |
Low Public capital
|
University/Secondary
| 6464 | 1 | | 2451 | 1 | |
|
Primary
| 2587 | 1.81 (1.55-2.13) | <0.001 | 3841 | 1.51 (1.28-1.77) | <0.001 |
|
None
| 330 | 4.14 (2.98-5.77) | <0.001 | 582 | 5.82 (4.55-7.44) | <0.001 |
|
Trend
| | | <0.001 | | | <0.001 |
High Public capital
|
University/Secondary
| 12247 | 1 | | 393 | 1 | |
|
Primary
| 3376 | 1.66 (1.44-1.92) | <0.001 | 269 | 1.01 (0.56-1.79) | 0.99 |
|
None
| 227 | 2.13 (1.34-3.89) | 0.001 | 16 | 1.84 (0.39-8.63) | 0.44 |
|
Trend
| | | <0.001 | | | 0.46 |
|
Interaction
| | | 0.12 | | | 0.10 |
There was evidence of interaction between Human capital (level of education) and Public capital (Low/High) for current non-use of contraception in women living in rural areas in a model adjusted for age in years, marital status and children ever born. These interaction effect remained after adjusting for Physical capital (household wealth) (Table
5). Households with higher Public capital were wealthier in terms of Physical capital in urban and rural areas (urban areas: 0.79 SD wealthier on physical capital score; rural areas: 1.03 SD; unpaired t test p < 0.001 for both areas) Adjustment for Physical capital attenuated inequalities in current non-use of contraception according to level of education (15-62%). When self-reported health insurance cover was controlled for in the interaction model, unadjusted for Physical capital, the coefficients did not change. Results from sensitivity tests on all women of fertile age (15-49) showed a similar gradient (data not shown).
Conclusions
A multidimensional asset-based approach provides a framework for disentangling socioeconomic inequalities in contraceptive behaviour. Its application facilitates an intermediate level of analysis between a composite index (HWI) and multiple measures of SEP, by comparing the magnitude of health inequalities attributable to specific dimensions. We believe this approach could be a starting point to address questions about the relative importance of different dimensions of SEP on inequalities in health. For women living in urban and rural areas in Colombia we have shown that Physical capital identified important socioeconomic inequalities in current non-use and never use of modern contraceptive methods. Importantly, we provide some support for our interaction hypothesis that provision of public services compensates for women's low levels of education with respect to contraceptive behaviour. Our results suggest that women's education and household living conditions should be continued and strengthen in public health policy objectives for Colombia. If complemented with wider provision of public infrastructure in deprived urban and rural areas of the country, socioeconomic inequalities in modern contraceptive use may be reduced.
Competing interests
The authors declare that they have no competing interests.
Authors' contributions
CG and EB jointly developed the principal research idea for the analyses. CG obtained and analysed the data and reviewed the literature. CG and TH drafted the paper. EB supervised the analyses, and all authors read and approved the final manuscript.