**5. Methods**

This study examined the ability of 12 independent variables to explain household migration decisions. Two variables are climatic (temperature and precipitation), while four involve household demographics (household head age, household literacy, males over 15, and females over 15). Six variables define household economic characteristics: availability of extension advice, land ownership, TLU, soil quality, total land area, and distance to market.

We analyzed spatial patterns of precipitation and temperature across the regions with *ArcGIS*. The georeferenced datasets were converted into *ArcGIS* shapefiles using the Arc 1960 projection for Tanzania [55]. Households were divided into migrant and non-migrant categories. A household with at least one member who migrated during the survey year (2012) was considered a migrant household. All households in the same sub-village were assigned the same locational coordinates to merge climate data with each household. Temperature and precipitation changes were identified and first standard deviation data were incorporated into the analysis [43]. These data were mapped to show spatial migration patterns for migrant and non-migrant households across the country's regions. Coefficients of variation (CoV), defined as σ/μ, for each sub-village's rainfall and temperatures were displayed and analyzed using choropleth and graduated symbol maps. Average maximum temperature and rainfall represented annual extreme conditions. Long-run average rainfall and temperature represent the mean of yearly average values. Anomalies were the difference between the annual average and the 30 year annual mean. Anomalies are also shown using choropleth and graduated symbol maps. Additionally, household exposure to climatic variables was measured by the between-years rainfall CoV, for the period 1983–2012. CoV provided several advantages. First, for a given level of standard deviation, the CoV changes as the mean changes showing a lower level of variability for sub-villages with higher levels of average rainfall or temperature. Second, CoV shows dispersion of temperature and rainfall values in relative terms, allowing comparison between two subvillages. We also computed sub-village average precipitation shortfall (i.e., the average of the annual totals' departures from the long-run average). The same procedure was followed for study period temperature.

Migrant and non-migrant household demographic characteristics were analyzed using descriptive and inferential statistics for variables such as age, sex, household size, female head of household, literacy status of household head, male and female workforces, and households' access to agricultural extension. Migrant and non-migrant

households; distance from household farm plots to market; total land holding size; and total number of Tropical Livestock Units (TLU) were analyzed with t-statistics.

Our control variables and migration are related as follows. Household educational level is expected to increase migration because better-educated individuals are more likely to have information about migration and job availability in urban areas than are less educated individuals [56]. Total land holding is expected to decrease migration as larger land size corresponds to increased labor requirements for the household. Conversely, larger family size is expected to increase migration since households may have an excess labor supply.

The relationship between household head age and migration is expected to be mixed. If the household head is middle-aged, he may send a family member to migrate since migration may not affect family labor demand. Alternatively, an elderly household head may not be able to perform demanding tasks and may not be able to afford to expend family members on migration.

As a wealth indicator, TLU is expected to have a mixed effect on migration. Poorer households (low TLU) may lack financial resources for migration [57]. However, low income families tend to engage in short distance migration, while better income families may choose long distance migration [58]. Furthermore, gender may be important as female members of the household may not migrate long distances.

A household holding title to land is more secure than a landless household. Increased tenure security may discourage migration. Likewise, households having good soil quality could prefer working their land instead of migration. Household plot distance to market, on the other hand, may positively influence migration since a household close to market may have better labor market information and therefore opt to migrate.

Besides descriptive statistics, t-tests, and spatial analyses, we employed a logistic regression analysis to explain migration. Our analysis was framed following [59] approach, presenting migration as a determinant of a set of explanatory variables. We extended this approach to feature rainfall and temperature as key migration determinants. The logit specification assumes the household chooses migration *m* if the utility derived from the choice of *m* is greater than the decision not to migrate. As the utility from migration/non migration is unobservable, it can be expressed as a function of observables in the latent variable model, given in Eq. (1):

$$\mathbf{M}\_{i}^{\ \ast} = a\mathbf{Z}\_{i} + \varepsilon\_{i}\operatorname{with}\mathbf{M}\_{i} = \begin{cases} \mathbf{1} & \text{if}\mathbf{M}\_{i}^{\ \ast} \\ \mathbf{0} & \text{otherwise} \end{cases} = \mathbf{0} \tag{1}$$

where *Mi* is a dummy variable for the choice of migration; *Mi* ¼ 1 if the household has chosen the decision to migrate, and *Mi* ¼ 0 otherwise. *α* is a vector of parameters to be estimated; *Zi* is a vector that represents household characteristics; and *εi* is the random error term.

Based on Eq. (1), the estimated relationship between migration and temperature and precipitation is given by:

$$\ln\left(m\_h\right) = a\_h + \beta \kappa\_h + \tau d\_h + \rho i\_h + \mu \varepsilon\_h + q j\_h + \varepsilon\_h \tag{2}$$

where *h* denotes household for migrant or non-migrant family. Migration is denoted by *mh*, *xh* represents the temperature for each household *h,* while *dh* represents the precipitation for household *h* over a 30 year period. The set of socio-economic characteristics are denoted by *ih* while *eh* and *j <sup>h</sup>*are the set of

*Spatial Analysis of Climate Driver Impacts on Sub-Saharan African Migration Patterns… DOI: http://dx.doi.org/10.5772/intechopen.106067*


#### **Table 1.**

*Key independent variable pearson correlation matrix.*

institutional and the physical capital variables respectively. The coefficients *β*, *τ*, *ρ*, *μ* and *φ* denote the respective vector of parameter estimates, and *ε<sup>h</sup>* indicates the error term.

The model's dependent variable is set to a 0–1 dummy variable, where 1 represents migrants and 0 represents the non-migrants household. Accordingly, the predicted values for the dependent variables will fall between 0 and 1 interval. These results will show the probability of households deciding to migrate.

As previously indicated, this study's first hypothesis states that temperatures and precipitation show different spatial patterns between migrant and non-migrant households across sub-villages in the region. Inferential t-test statistics compared the effects of temperature and precipitation as well as the other control variables on migrant and non-migrant households. The second hypothesis states temperature and precipitation have more influence on migrant than non-migrant households with differing impacts across space. This hypothesis was tested using regression analysis. Regression coefficients quantified the (positive or negative) impact of a unit change in that variable on the propensity to migrate.

Our control variable decisions were based on [60, 61]. Since climatic variables and control variables are exogenous (unaffected by patterns of migration), a logit equation can be estimated without endogeneity concerns. Additionally, we addressed multicollinearity concerns with an independent variable correlation analysis. None of the pairs of variables were highly correlated (**Table 1**).
