«By Nathan B. Goodale A dissertation submitted in partial fulfillment of the requirements for the degree of DOCTOR OF PHILOSOPHY WASHINGTON STATE ...»
Health and fertility are major factors in demographic change. Based upon data from the American Southwest, North America in general, the Near East and Europe, Bocquet-Appel and colleagues (Bocquet-Appel 2002; Bocquet-Appel et al.
2008 Bocquet-Appel and Naji 2006; Guerrero et al. 2009; Kohler et al. 2008; Kuijt 2008a) have defined the NDT as being characterized by two phases. In the first phase there is a dramatic increase in fertility and thus in population size since death rates are approximately constant, whereas in the second phase, there is an increase in mortality due to increased disease parasitic infection (Bocquet-Appel 2002; Bocquet-Appel et al. 2008). This model not only suggests that health and fertility were factors in the NDT, but indicates the archaeological data in these regions (and in every other NDT) can be defined by a time sequence where first came an increase in fertility and second a decrease in health with increasing death rates.
The NDT as a two-phase phenomenon is demonstrated by Bocquet-Appel and colleagues to be evident in several regions of the world; however, its accuracy has not been explored in its entirety in the Near East. This is because the data is available to test increases in fertility, but not the second phase of deteriorations in health (Guerrero et al. 2009). In addition, although suggesting increased fertility as the primary motivator for population growth, Bocquet-Appel (2002:647) provides an explanation that there may be some link between the increasing intake of better
Quantifying Paleodemographic Change Bocquet-Appel (2002:637) defines the NDT as a “substantial increase in human numbers.” While accurate, this definition is limited in that it does not incorporate enough detail for detecting the NDT in the archaeological record. It is important that the NDT be defined more broadly to include the processes by which populations increase, in turn, providing the quantitative assessment of past demographic transitions.
It is essential to expand on the processes that occur within a demographic transition. In turn, this will allow a quantifiable measure of detecting demographic change in the archaeological record. To accomplish this, researchers need to be able to utilize the most basic factors witnessed universally in demographic change that are also variables commonly reported in the archaeological literature. These include factors of time measurement, differential use of space, and the scale of archaeological remains that people leave behind. These variables modeled over long periods may aid in our understanding of changes in population growth rates.
Here, I propose a definition for a demographic transition as any significant departure in population growth rates over a defined period of time brought about through changes in birth, death, or migration. In this way, measures can be made regarding differences in population growth or decline between two defined time
population sizes in a region based on archaeological evidence, we may contribute further understanding of prehistoric paleodemographic transitions.
Detecting the NDT: Archaeological Attempts and Viable Data Bocquet-Appel (2002:637) argues that a Neolithic demographic transition may be inferred from three lines of evidence: 1) genetic markers of migration (Ammerman and Cavalli-Sforza 1984), 2) paleoanthropological markers in cemeteries (Bocquet-Appel 2002), and 3) interpretive models based on examining the densities of archaeological sites through time (similar to Hassan 1981 and references therein), each of which is discussed in detail below. All of these models have been used to some extent to suggest a demographic transition associated with the origins of food production. However, to date there has been very little use of population estimates based on multiple variables from archaeological assemblages to specifically detect a large-scale NDT.
Genetic Markers and Migration Until recently, the most frequently used technique to detect an NDT is through genetic markers of migration and expansion of the spread of food production from the Near East to Europe (Ammerman and Cavalli-Sforza 1971, 1984; Cavalli-Sforza and Cavalli-Sforza 1995; Simoni et al. 2000). Ammerman and Cavalli-Sforza developed a demic diffusion model focused on detecting genetic clines generated by a wave of
farmers outcompeted foragers through population growth and territorial expansion.
This model views population growth reaching carrying capacity at some defined point in time after the initial diffusion of the technology/idea (for instance the spread of domesticated plants and the techniques to efficiently use them). Population growth rate reaches equilibrium and then returns to zero, as specified by the logistic growth curve. Bocquet-Appel (2002:638), following Fix (1996), argues the model is weak because there is an ambiguous interpretation in that “The demographic signal itself the change representing the transition - is not…directly observable in the data presentation” (Bocquet-Appel 2002:638). In other words, a substantial increase in human numbers is not directly evident in the in the archaeological record.
Paleoanthropological Markers in Cemetery Data The second component of research supporting a world-wide NDT is age estimates based on osteological data that indicate an increase in the number of individuals between the ages of 5-19 in relationship to the entire population (BocquetAppel 2002; Bocquet-Appel and Naji 2006; Bocquet-Appel et al. 2008; Guerrero et al. 2009; Kohler et al. 2008). These studies utilize cemetery data and loess fitting statistical procedures to recognize the NDT in Europe, the Near East and in North America.
The logic behind Bocquet-Appel and colleagues reasoning for examining the 5-19 age bracket follows that the youngest age individuals (those under the age of 5)
susceptibility to mortality. These remains are also the least likely to be presearved in many instances. Measuring the number of individuals from the ages of 5-19 should show larger numbers in growing populations (both in the living population, and as sampled by death from it) relative to the entire population 5 years of age or older. We would expect the opposite in stable or declining populations. Their model is interesting in that they are able to detect both the timing and the rate of an NDT in multiple regions and have suggested universal characteristics of the worldwide NDT.
While universal, the NDT occurs more slowly in centers of independent invention than in areas where it spread due to the movement of people or exchange of ideas (Guerrero et al. 2009; Kuijt 2009).
Methodological Means to Detect the NDT There are several critical aspects to the analyses Bocquet-Appel and colleagues (2002; Bocquet-Appel and Naji 2006; Bocquet-Appel et al. 2008) conduct focusing on the results obtained through the loess smoothing procedure (see below) as well as how to apply the results in terms of our understanding of the long process of the transition to agriculture.
Originally developed by Cleveland (1979), LOWESS or more commonly referred to as loess fitting, is a nonparametric regression technique commonly used by political strategists examining voter patterns in relationship to amount of education, race, economic standing and so forth (Cleveland et al. 1993; Jacoby 2000). The
underlying structure within a dataset (usually depicted as a scatter plot with a line drawn as a goodness of fit in the overall trends of the data). The technique is also relatively robust against extreme outliers unlike least squares regression, its parametric counterpart.
Nevertheless, there are potential problems with the loess procedure that are inherent in its implementation. As a non-model-based method, there is no strict definition of goodness of fit and therefore requires the analyst to make several arbitrary decisions about the fitting parameters and what line is a best fit for representing the major structure within the data (Jacoby 2000:608) (although Cleveland and Grosse have outlined procedures to get statistical verification of the best fit line, it has only been utilized in Kohler et al. 2008). Instead, in most instances goodness of fit is determined through eyeballing by the analyst until the trends are close to what is expected. In this instance, eyeballing refers to the analyst, rather than a statistical software package to make the final decision of what is the best fit is (this is not true for Kohler et al. 2008 following Cleveland and Grosse 1991). Second, as demonstrated by Kohler et al. (2008), loess fitting should always be used in tandem with parametric fitting procedures or ones where statistical verification can be achieved as to the best fit line (also noted by Jacoby 2000:608).
The loess method can be thought of as a sliding window, and the size of the window predicates how much of the data is analyzed at a time and therefore how the best fit line is drawn. The data are scaled in some format such as years before
The first is α, which controls the size of the window as it slides from left to right over the data. The function of the window is to set a limit of how much data are incorporated into drawing the best fit line at each shift. In all of Bocquet-Appel’s analyses, the α values are consistently provided and range from α =.3 -.6. The second parameter important in loess fitting is λ, which relates to the type of polynomial whether 1st, 2nd, 3rd etc. order that is being fitted to the structure of the data, a parameter not specified in Bocquet-Appel’s analyses. Finally, the weighting algorithm specifies how the data are weighted with respect to the sample size of each data point and can often include where the data point is in relationship to the center of the window. The specific weighting algorithm used (or if one was used) in BocquetAppel’s analysis is also absent from his writings although it is provided by Kohler et al. (2008).
We can illustrate potential problems of this method by conducting an analysis of the data presented in Bocquet-Appel (2002: Table 1). The data represents skeletal populations from Mesolithic and Neolithic cemeteries from across Europe. In Figures 3.1-3.2, the loess fitting procedure is run with a series of α values from.1-.9 and λ values of 1 and 2 with a tricube weighting algorithm (to weight sample size). One can readily see that as the α value increases the less detailed the best fit line is drawn, or in other words, the line becomes smoother. Examining λ values of 1 and 2, we can see that the lines do not appear visually distinct (Figures 3.1 and 3.2). It should be mentioned that I could never exactly duplicate the line drawn by Bocquet-Appel
Appel’s (because the method he used is not mentioned).
Loess fits to the data, with λ = 1. The α parameter is indicated by the individual captions. Data from Bocquet-Appel 2002:640-641, Table 1.
Other questions arise on close examination of the methods used with paleoanthropological data from cemeteries with regard to relating the data to some aspect of temporal reality. The definition of dt=0 (supposedly, dt = is the temporal point of the demographic transition) as the local introduction of agriculture (BocquetAppel 2002) seems overly simplistic as well as problematic. For example, in the case
when people started cultivating plants, storing plants, or when the foods had evolved into their fully domestic form? In the Near East, this was a very long process; on the order of several thousand years. It is not clear, then, how dt = 0 is determined since it is not clearly defined in any of the publications utilizing this method (Bocquet-Appel 2002; Bocquet-Appel and Naji 2006; Guerrero 2008).
Loess fits to the data, with λ = 2. The α parameter is indicated by the individual captions. Data from Bocquet-Appel 2002:640-641, Table 1.
of the analyses presented by Bocquet-Appel and his colleagues, the data points representing hunter-gatherer cemeteries are extremely few in number, and small in sample size, since hunter-gatherer cemeteries are very rare. Could this cause problems with the way in which the best fit line is drawn pre-dt = 0 in relationship to post-dt = 0? We can investigate this by simulating random data and applying the loess procedure to them. In Figures 3.3 and 3.4, when there are equal amounts of data on each side of dt=0, variation in the data is lost by α =.5. This demonstrates that α values less than.5 may be subject to random variations in the data. This is interesting because most of the α values used by Bocquet-Appel and colleagues are between.3 and.6 suggesting that these are probably well fit lines that are not subject to random variations in the data but reflective of overall patterns within the data.
I also make this point because, in several of the analyses presented thus far, some proportions of 5P15 (the number of individuals between 5 and 19) samples of hunter-gather datasets are on a par with those of agricultural populations (for example Bocquet-Appel and Naji 2006: Figure 3 and Guerrero et al. 2008: Figure 2), making it likely that as more data is utilized, the line may be drawn significantly different.
38 Figure 3.3. Loess fits to randomly generated data, with λ = 1. The α parameter is indicated by the individual captions. Note how the best fit line behaves the smaller the window size and the size of the dataset dramatically influences how the line is drawn.
While I have these critiques, I do not suggest that these analyses are ultimately futile, but an important aspect to a greater understanding of the NDT. Instead, what I suggest is that these potential issues open the door for the need to employ new and different techniques as well as different data to examine the timing, rate, and characteristics of the NDT.
39 Figure 3.4. Loess fits to randomly generated data, with λ = 2. The α parameter is indicated by the individual captions. Note how the best fit line behaves the smaller the window size and the size of the dataset dramatically influences how the line is drawn and the line is less curved with smaller window sizes than when λ = 1.