Foraging complexity and the evolution of childhood

Our species’ long childhood is hypothesized to have evolved as a period for learning complex foraging skills. Researchers studying the development of foraging proficiency have focused on assessing this hypothesis, yet studies present inconsistent conclusions regarding the connection between foraging skill development and niche complexity. Here, we leverage published records of child and adolescent foragers from 28 societies to (i) quantify how skill-intensive different resources are and (ii) assess whether children’s proficiency increases more slowly for more skill-intensive resources. We find that foraging returns increase slowly for more skill-intensive, difficult-to-extract resources (tubers and game), consistent with peak productivity attained in adulthood. Foraging returns for easier-to-extract resources (fruit and fish/shellfish) increase rapidly during childhood, with adult levels of productivity reached by adolescence. Our findings support the view that long childhoods evolved as an extended period for learning to extract complex resources characteristic of the human foraging niche.


.1 Data sets and metadata
In figure S1 we summarize the literature selection process that produced our sample. Table S1 shows some metadata referring to the datasets included in the analysis. The data are a combination of datasets obtained directly from the literature, the first 11 rows in the table below, and original data organized by Koster et al. (13) in the cchunts package, marked by the code cchunts in the Data column. Specifically, Koster "searched for relevant studies on subsistence hunting in the anthropological and biological literature, subsequently contacting authors to invite them to contribute data. The contributors submitted data in a standardized format that included variables for the biomass acquired on terrestrial hunting trips, the ages of the hunters at the time of the hunt, the duration of the trip, the hunting weaponry carried by the hunters, and the presence of dogs or assistants" (13).
Following the inclusion of data from this source, we screened the data we extracted from published papers to remove repeated data sets. In particular, data relative to the Ache of Paraguay extracted from Walker et al. (16) have not been used because they are already present in the "Hill Kintigh" data set included in the cchunts package. Moreover, Tsimane data extracted from Gurven et al. (11) are a repetition of those included in the cchunts data ("Trumble Gurven"). Only the latter were used in the analysis.
Data collected by Bliege Bird and Bird among the Martu in Western Australia come from both a 2005 study on children's foraging (8) and from the dataset in the cchunts package (13). The 2005 paper reports data from individual of both sexes between 5 and 14 years old hunting goanna lizards in the rocky outcrop not far from the camp. These data were collected by the authors between 2000 and 2002. The cchunts data were collected between 2002 and 2010, are relative to individuals aged 7 to 79 and partially exclude female contributions ("This data set includes observations of female foragers when they were accompanied by men on trips, but not women on foraging trips that did not include male foragers"). The two data set are thus not fully overlapping, but there is the possibility that some data are present in both sets. In particular, 14 foraging returns collected in 2002 from individuals below 14 years old are present in the cchunts data set and could hence have been included in Bird and Bird (8). Looking in detail at these subsets, they do not appear to be repetitious (a 9 years old boy present in the cchunts dataset does not appear in the Bird and Bird (8) study, for example, and none of the younger individuals' returns reported here appear in cchunts). Table S1: Metadata for included datasets. These are relative to each source of foraging returns data included in the analysis. The first 11 datasets have been extracted from published papers, the remaining were part of the cchunts package. The Data column reports both the code referencing the dataset in our code, e.g. 'Bird 2005' and the reference to the relative study, in italics among parenthesis. As sample size we report the total number of observations for foraging returns, with the total number of foragers under 20 years included in our analysis in parentheses.

Data
Population Years

Excluded datasets
Several papers that passed the first rounds of selection (i.e. appeared to report original data on children foraging returns) were not included in the analysis for a variety of reasons.
Some were subsequently found not to report relevant data (20,67). Others did not include enough data on forager ages (12,(68)(69)(70). Kramer (71) does not include data relative to foraging, and Kramer (72) estimates returns from time allocation data, which is not consistent with the other data sets. Several papers reported data in formats that did not allow them to be extracted for analysis, such as failing to report errors around mean return per age class (5, 73), or including smoothed loess curves (74)(75)(76). Kawabe (42) reports ranges for individual returns (e.g. "more than 5 specimens killed"). Finally, a number of papers use the same data. In these cases, we included only the data sets for the latest or most informative paper. Bird and Bird (77), Blurton Jones (78), Pollom et al. (79) and Walker (80), for example, present data which are best extracted from other papers.

Pooled data
A small subset of the cchunts dataset features pooled data (380 out of a total sample size of 2297, representing 17%). For these data points, reported returns are for groups rather than individuals. Social foraging can be conceptualized similarly to technology (guns, dogs) which can increase individual returns. Indeed, including pooled returns has the potential to skew the results by inflating individual returns. To ensure that this is not the case, we first compared the distribution of pooled vs. non pooled data across ages. As can be shown in Figure S14 and S15 respectively, individual and pooled data within the whole data set and by society are evenly distributed across ages, and hence unlikely to bias our results. Moreover, we re-ran the analysis on a subset of data excluding pooled returns. Figure S13 shows that excluding pooled returns does not impact our findings.

Integrating individual-level data with study-level summary statistics
Our data included a mix of individual-level returns (e.g., a forager brought back y kilograms of fish) and summary statistics, such as the mean and standard deviation of returns for age classes (e.g., children of ages 5-10 collect tubers at a certain average rate). The challenge was to synthesize two distinct types of data: individual-level observations drawn from f (y|µ, σ) and group-level averages E[y|µ, σ]. Excluded as duplicated in the cchunts package: 2 Figure S1: Process of paper selection.
When returns were given as summary statistics (i.e., mean and standard error), we modelled them using a measurement error model: Where µ [obs] is the group-level mean, σμ is the standard error of that mean, and, following our generative model defined in the main text, E[y|µ, σ] = p(exp(log(µ) + σ 2 2 )). Finally, there was some variation in the number of measures available for individuals, and some studies have multiple measurements from the same forager, in which case we included a random effect on skill to account for non-independence of these data points. However, lack of longitudinal data meant that it was not possible to estimate individual differences in the life history parameters.

Priors
We employed regularizing priors for all parameters to reduce over-fitting and facilitate model convergence. Specifically, we assigned a N ormal(0, 1) to fixed effects (e.g., intercepts), an Exponential(1) for the random effect standard deviations and LKJ(2) for the correlations between random effects. Moreover, we fixed to zero the starting values for several parameters, as this can be helpful (and in some cases, necessary) for the model start sampling. However, after many iterations of warm up there is no dependence on starting values, unless the model is badly mis-specified or suffers from a multimodal posterior. We used standard MCMC diagnostics and found that all population parameters had Rhat < 1.01 and an effective sample size > 1000.

Dealing with uncertainty
Uncertainty in age: Forager age was not reported exactly in any study. Most frequently, authors reported an integer age for each child. In other cases an interval of possible ages was given (e.g., 4-7). We modelled age using a Gaussian measurement error model: age obs ∼ Normal(µ age , σā ge obs ) Uncertainty in sex : In cases where sex of the forager was not reported (or was given as a summary statistic), we average over sex differences in proportion to how often males and females appeared in a given study using Stan's log mix() function.               Table 1 Froehle 2018 table4

Additional figures
Hawkes 1995 Table 4 Makalita rates g.h Hawkes 1995 Figure S8: Fish data. Data points for all datasets reporting fish and shellfish data.
Hawkes 1995 Table 4 Figure S10: Game data. Data points for all datasets reporting game data.   Figure S11: Data from other resources. Data points for all datasets reporting data for other kinds of resources.

BlurtonJones 2002 Figure 5 tubers
Hawkes 1995 Table 4 ekwa rates g.h tubers Hawkes 1995 Table 4 Makalita rates g.h tubers  Figure S12: Tuber data. Data points for all datasets reporting tuber data. Figure S13: Returns by age per resource for non-pooled data only. Predicted change in foraging returns by resource type excluding pooled data in thecchunts package, with thick lines denoting the average posterior median and thin lines denoting the median for each unique study outcome for that resource type. All curves are scaled by their maximum value (predicted returns at age 20). Comparing this plot with in-text Figure 3b shows that there is no strong effect of the pooled data on our results. De n s i t y Figure S15: Data density at each age for pooled vs. non pooled data by population.