Behavioural Attitude

);

(Equation 7)

If a random number ( ) between zero and one generated by the model is less than the resulting value, a score of 1 is therefore allocated to , migration is perceived by the agent to be within their means, and through Equation 1, an individual agent would develop an intention value towards that option. Otherwise a value of 0 is assigned to and, through the use of Equation 1, an individual agent’s intention to migrate to that destination will be zero. The higher an agent’s asset rate ( ), experience rate ( ) scores, the higher the value calculated and the greater the likelihood that an agent will return a value of 1. The agents’ perceptions of their behavioural control return a binary outcome in order to aid clarification of the migration decision. Rather than an agent thinking they ‘might’ have the capacity to migrate, their consideration of an option is defined as a yes/no decision formed on the basis of assets and experience. This enables clearer definition of an agent’s options and a greater ability of the model to quantify the ‘able’ population.

As illustrated by Equation 2, the model code necessary to fully implement the impact of upon behavioural intention requires the use of an additional parameter: . Each agent that scores a value of 1 for a migration option is placed in a virtual cohort with all other agents from their origin zone that have a score of 1 for the relevant option. The size of this cohort is then the value used in Equation 2 that is used to adjust values to reflect the likelihood of an individual within this new cohort migrating. As probability values have been calculated without any notion of perceived behavioural control, failing to incorporate this adjustment would mean that the original values would be applied to an artificially small population of agents resulting in inappropriately small modelled migration flows.

5.4 Behavioural Attitude

The behavioural attitude ( ) component of the decision to migrate is, for each agent, selected from a matrix of values stored within the model and represents the probability of an agent from origin location ( ), with rainfall conditions ( ) that year and with current age ( ) gender ( ) and marital status ( ) attributes, migrating to the destination option ( ) being considered. The origin location ranges from 1 (Ouagadougou) to 5 (Southwest) in the model while rainfall

conditions range from 1 (dry) to 3 (wet) through 2 (average). Although a more continuous set of rainfall values could have been developed through the calculation of statistically-derived functions between rainfall and the other model inputs, the tertiary (dry, average, wet) approach adopted by this research was selected in order to facilitate clearer analysis of model outputs.

Inevitably, the use of a tertiary classification will have introduced some margin of error due to the creation of discrete classifications defined by specific thresholds. However, the potential advantages of fuzzier thresholds that would have resulted from the use of a function was deemed inadequate to justify the lack of clarity in model output analysis that would result.

The probability value ( ) of an agent migrating to destination option ( ) is therefore calculated from the number of individuals within the EMIUB data with defined attributes , , and who are migrants ( ) from location , under the prevalent rainfall conditions in the period 1970-1999, divided by the population ( ) of that location with the same defined attributes at that point in time (Equation 8). The 1970-1999 probability values stored within the matrix are derived from analysis of the EMIUB dataset and represent the likelihood of an agent with the same characteristics undertaking migration as an adaptation strategy in the face of the existing conditions.

(Equation 8)

The probability values retrieved by an agent, as well as being defined on the basis of variables and , are further controlled according to the section of the EMIUB data from which they were calculated. The model can therefore be run using probability values derived from a total of four separate weight matrices calculated from combinations of two rainfall precursors and two data time periods. In rainfall terms probability values can represent the likelihood of an individual migrating in a year defined as dry, average or wet according to the July to September (JAS) rainfall conditions of that year ( ) or the current year and two preceding years ( ).

The purpose of the second rainfall weight matrix component ( ) is to accommodate the possibility that the migration response to a factor such as rainfall may be more of a slow onset decision to longer term trends than a rapid response to a short term change.

In time period terms, the two approaches used to calculate probability values are distinct as a result of the section of the EMIUB data analysed to produce the values stored in the relevant weight matrices. The EMIUB data provides reliable retrospective migration data from Burkina Faso for the thirty year period 1970-1999. By using the migration record for the period 1990-1999 ( ) to calculate for use in the AMARC model, the migration record from 1970-1989 can be used in a format where one third of the available data is used for model parameterisation and the remainder for cross-validation. Such an approach is often used in the field of artificial intelligence when training a neural network that is intended to abstract the high level of complexity inherent to most biological systems and focus on the core information processing structures. However, in accordance with the notion of data fitting (Schittkowski, 2003) and in order to calculate more accurate migration probability values from a longer time period, the entire EMIUB migration data period of thirty years (1970-1999) has also been used to calculate alternative probability values for the AMARC model ( . On the basis of Roberts and Pashler’s (2000) and Fum et al.’s (2007) comments on parameter fitting, the AMARC model can be strengthened by using the weight matrices that fit model parameters to the overall variability found in the full range of data, thereby avoiding

‘overfitting’ and permitting more representative future simulations in post-validation model applications. The four weight matrices available in the AMARC model are therefore (1 year rainfall and 10 year migration data analysis), (3 year rainfall and 10 year migration data analysis), (1 year rainfall and 30 year migration data analysis), and (3 year rainfall and 30 year migration data analysis).

Whichever combination of or , or the AMARC model probability values are derived from, those retrieved by each agent reflect the likelihood of an individual with age, gender and marital status characteristics , and migrating. However, as only those agents who perceive that they are able to complete the migration in question will actually do so, the values are adjusted to represent the increased likelihood of an agent within this reduced ‘able’

population migrating. Such agents are those that return a value of 1 towards the option in question and so consider themselves capable of migrating as a result of the behavioural control ( ) value calculated in Equation 6. The adjusted probability value represents the behavioural attitude ( ) of the agent and is calculated on the basis of Equation 9 using the probability value for the relevant population of the agent’s origin location ( ) and the population of agents that have scored 1 for their value towards that option ( ) in the current model cycle.

(Equation 9)

The probability values assigned to each option considered by each agent thus reflect the likelihood of an agent with the given characteristics migrating under the relevant circumstances.

Because the different combinations of characteristics used to govern the migration decisions of agents in the model result in twelve unique descriptions of agents that inhabit five different origin locations and respond to three different levels of rainfall, the matrix containing the probability values retrieved by agents is very large. In order to present an idea of the overall migration tendencies of agents in each of the five zones across all rainfall levels, Figure 5.3 displays the sum total of the probabilities of agents with characteristics and migrating to any available destination under all rainfall conditions. Although sum probabilities do not provide the full detail used by the ABM, the data displayed in Figure 5.4 provides an indication of the overall tendencies of agents with certain characteristics towards migration.

Figure 5.3: Sum probabilities of agents with given characteristics migrating from one of the five zones of origin to any other destination.

The sum probability of any class of agent migrating from one origin location to any other destination ranges from 0.0313 for 35+ single females originating in Southwest to 0.6901 for 15-20 married females originating in Sahel. As a result of the nature of the probability values used (the number of individuals with defined characteristics migrating from one origin location to an alternative destination in a particular year, divided by the number of individuals with those characteristics present in that origin location that year), they are dependent upon both the

-0.1000 0.0000 0.1000 0.2000 0.3000 0.4000 0.5000 0.6000 0.7000

Ouagadougou Bobo Dioulasso Sahel

Centre Southwest

Sum Migration Probability

number of migrants with the defined characteristics and the size of the population with those same characteristics. As identified in Chapter 4, those individual migrations recorded as being undertaken for the purposes of marriage and postings were removed from the analysis.

Migration probability values may therefore result from two features of the data. The first of these is the number of individuals with the given characteristics who migrated under the appropriate circumstances in the EMIUB data each year during the relevant EMIUB data analysis period. The second component is the population of individuals with the given characteristics that were present in the EMIUB data each year. A high migration probability value therefore indicates that a high proportion of the total population of agents with defined characteristics migrated but gives us no information on the actual size of that population.

Equally, a low probability value suggests only that a small number of the population of unknown size were seen to migrate in the data. As a result, it is important in the development of the model to ensure that the population of modelled agents display a similar demographic structure to those in the dataset from which the specific probability values were derived.

In addition to demographic change in the form of birth, ageing and death, due to the definition of classes of agents according to age, gender and marital status, functions controlling the rate of marriage within the model must accurately portray the rate of marriage witnessed in the data.

This is important because of the cultural norm referred to during focus group interviews in Burkina Faso that leads women to marry at a relatively young age and men to marry slightly later in life. As a result, a woman in her mid-thirties is more likely to be married than a man of the same age. The likelihood of there being an unmarried female of over 35 years of age in the EMIUB data is therefore low. If one were to be present in the data, should they migrate during the analysis period of 1970-1999, the probability of such an agent migrating in the model would increase significantly. To permit the probability values retrieved by agents in the model to have the appropriate impact upon migration, the population structure of the modelled agents must therefore reflect that of the real world population surveyed by the EMIUB in 2000. Figure 5.4 displays the 1970 population structure of individuals surveyed by the EMIUB in 2000.

Figure 5.4: 1970 EMIUB population of individuals in each zone that fall within each category of agent characteristics.

It is evident from Figure 5.4 that there are relatively few over 35 year old agents in the model at startup in 1970. This results from the retrospective nature of the EMIUB dataset (survey conducted in 2000, therefore restricting the number of individuals with records reaching 1970 to those aged 30 and over in 2000) and, in post validation runs of the AMARC model, is rectified by modelled demographic change creating a more representative population structure. In early versions of the model, required for as stringent validation as possible, the 1970 population of agents are followed to the year 1999 and their model migrations compared with those recorded in the EMIUB data. Using EMIUB data from the entire period 1970-1999 allows the full spectrum of agent probability values to be captured. If only 1970-1979 data were used, probability values relating to older agents may be developed from a relatively small sample.

Similarly, if the 1990-1999 data alone were used the same result would be felt on younger agents.

Of those agents aged between 15 and 20 years, very few are seen in Figure 5.4 to be married males. Similarly, if we were to extend the analysis to consider a date closer to 2000, very few of the agents aged 35 years and over, either male or female, would be single. This results from the cultural norm relating to marriage in Burkina Faso. The EMIUB data suggests that most individuals, male or female, are married by the age of 35. The greatest population classification represented in Figure 5.5 is that of married females between the ages of 21 and 35 years. This may again result from the cultural norm surrounding marriage in Burkina Faso. While there are a comparatively large number of married females aged between 21 and 35 years, there are a

0 20 40 60 80 100 120 140

Ouagadougou Bobo Dioulasso Sahel

Centre Southwest

Population

correspondingly small number of single females within the same age bracket. As a result, the total number of males and females within the age groups is comparable, with differing gender norms relating to marriage generating different population distributions within the agent classifications used in the model.

The EMIUB data provides the ABM developed by this thesis with the behavioural attitude values used in the migration decision-making process of each agent. Although these attitude values are affected by population dynamics such as those described above, the impact of these upon the actual running of the model is minimised by using the EMIUB data as a basis from which the agent population is constructed and changes over time. Demographic functions that affect behavioural attitude values retrieved by each classification of agent should therefore be reflected in the agent population as the model runs. Rules controlling marriage within the model are different for males and females in order to reflect the different ages at which males and females tend to marry in Burkina Faso.

The combination of agent characteristics apparently most likely to migrate according to the data presented in Figure 5.3 are married females aged between 15 and 20 years. As mentioned in Chapter 4, individuals surveyed by the EMIUB were asked to provide information on their motives for migration. Unfortunately, data on the motives of migrants was not comprehensive.

Nonetheless, in performing data analysis to derive migration probability values, those individuals who stated that their motive for migration was marriage (coded as 435 in the EMIUB) were removed, as were those who listed their motive as a work-related involuntary posting (coded as 105 in the EMIUB). Migration listed as being motivated by either marriage or a posting were removed from the data analysis as they are both considered to be largely independent of rainfall. Although the timing of the relocation of a female individual to a new location for the purpose of marriage may be affected by rainfall in terms of, for example, the ability of the household to invest in a dowry, it was considered by this research that the destination of the migrant is unlikely to be influenced by rainfall. Similarly, a work-related posting such as, for example, to fill a government position in a different department, is considered external to the influence of rainfall. Although the value of this decision may be debated it is important to note that the exclusion of individuals migrating for marriage and postings is consistently and transparently retained throughout this research.

Despite this measure to remove the influence of marriage migration from the EMIUB data

record, the apparent high proportion of married females aged between 15 and 20 years can likely be attributed, in part, to marriage migration as a result of the incomplete record of migrant motives. Field interviews performed in Burkina Faso in early 2009 provided verbal evidence that young married females are in fact very unlikely to migrate and would largely do so only when the family’s harvest yield was so low that family members could not be supported at home. Young females are however very likely to migrate for the purpose of marriage between the ages of 15 and 20 years. Despite removing those individuals that cited marriage as their motive for moving from the data analysis, the limited information on motives for migration results in numerous marriage migrants remaining in the data. Despite the fact that a large proportion of migrants in the EMIUB are young females probably migrating for marriage, they cannot be removed from the data on this assumption alone. The fact that the probability of migrating is considerably higher for married females aged 21-35 years than single females of the same age is however evidence of this phenomenon. As the EMIUB data analysis identified the number of migrants with given characteristics each year, those young females who migrated for marriage would have become married that year and therefore been represented in the data as married women. This phenomenon can be assumed as a result of the structure of the EMIUB data and resulting analysis. However, no action has been taken to rectify the anomaly due to the lack of concrete evidence and an unwillingness to ‘fix’ the data to better suit the world view as seen by the author. Given infinite time and resources, it would be possible to undertake such a fix and test its influence upon model outcomes. However, within the constraints of this research, such a step was deemed unnecessary.

Other than the potentially anomalous results for married females aged between 15 and 20 years, the migration probability values calculated for agents originating in each of the five zones show a good level of similarity in terms of the general migration trends for each classification of agent characteristic. The agent classes with generally the most consistently high trends towards migration are seen to be single males aged between 15 and 35 years. This aligns well with information on the migration of Burkinabé people collected during the field study period of this research. Focus group interviews conducted across Burkina Faso support the notion that the first and most willing to migrate are young males, particularly those who are unmarried. Overall, apart from the potentially anomalous probability results for married females aged between 15 and 20 years, female agents are, as a result of the behavioural attitude score they retrieve in the formation of their intention towards migration, less likely to migrate. This fits well with evidence collected during focus group interviews conducted in Burkina Faso where the general consensus appeared to be that women and old men only migrate away from their home village in years with a very bad harvest.

The behavioural attitude values retrieved by agents in the ABM reflect the agent’s age, gender, marital status, origin location, and the rainfall conditions affecting them in the year in question.

The five standard weight matrices used by the five sets of agents in the ABM (ppt.mp30) each have 36 behavioural attitude probability values that may be retrieved by an agent on the basis of

The five standard weight matrices used by the five sets of agents in the ABM (ppt.mp30) each have 36 behavioural attitude probability values that may be retrieved by an agent on the basis of