Table 4.18: Variables for the second step of the nested m-logit model for irrigation decisions
Variable Definition Data Source
Dependent Variable
Hirr method Irrigation method (dam, pump or bucket irri-gation)
Interview Independent Variables
Hcash rainy Cash income from the rainy season (in
Cedis)
Interview and Calculation Hdry years Number of years the farmer is involved in
ir-rigation farming
Interview
Hpump Dummy variable indicating whether the
household owns a motor pump
Interview
Hdist dams Minimum distance to dams (in m) Map-based
Calculation
Results of second step of irrigation m-logit model
This model, which simulates the choice among the three irrigation alternatives, has a rela-tively high predictive power (Table 4.21), with a Nagelkerke R Square of 0.940, although the variables show fairly good significance levels (Table 4.19). Among the three irrigation alter-natives, all cases of dam irrigation and bucket irrigation are correctly predicted, with about 76.2 % of correct predictions for the pump irrigation method. In total, 94.1 % are correctly predicted (see Table 4.20).
The results of the m-logit regression are not fully consistent with the theory of the influence of the selected variables as outlined above. In fact, cash income positively influ-ences the choice of the more costly pump irrigation, but the pump dummy variable and the number of years the farmer is involved in irrigation farming hardly show any influence in the choice among these two riverine irrigation methods.
Table 4.19: Second step of the nested irrigation decision model: parameter estimates
Irrigation Std.
Method Variables B Error Wald df Sig. Exp(B)
Motor pump Intercept - 484.079 3025.502 0.026 1 0.873
Hcash rainy(log) 27.112 203.890 0.018 1 0.894 6E+011
Hdist dams - 0.130 0.000 470405 1 0.000 0.139
Hdry years - 8.647 62.060 0.019 1 0.889 0.000
Hpump - 413.307 6970.703 0.004 1 0.953 3.18E-180 Bucket Intercept - 484.079 3025.502 0.026 1 0.873
Hcash rainy(log) 26.614 203.890 0.017 1 0.895 4E+011
Hdist dams 0.131 0.000 . 1 . 1.139
Hdry years - 8.457 62.060 0.019 1 0.892 0.000
Hpump - 437.689 0.000 . 1 . 8.2E-191
The reference category is irrigation Table 4.20: Second step of the nested
irrigation decision model: correct predictions
Predicted
Percent Observed Dam Pump Bucket Correct
Dam 25 0 0 100 %
Pump 0 16 5 76.2 %
Bucket 0 0 39 100 %
Perc. 29.4 % 18.8 % 51.8 % 94.1 %
Table 4.21: Second step of the nested irrigation decision model:
statistics
Model Fitting Pseudo
Information R Square
Cox
Chi- and Nagel- Mc
Square df Sig. Snell kerke Fadden 149.595 8 0.000 0.828 0.940 0.828
derive such livelihood groups, the livelihood framework for selecting livelihood indicators was applied, followed by the application of PCA and k-CA. Based on the identified liveli-hood indicators, the PCA revealed seven core factors that differentiate liveliliveli-hood typologies of farming households in the study area, namely land, labor, livestock, and income factors, two factors representing the preference for groundnut and compound farming, and the depen-dency ratio.
Based on these seven extracted components, classification using k-CA resulted in three livelihood typologies of households: the ’middle class’ (household type 1), the ’poor farmers’ (households type 2), the ’rich farmers’ (household type 3). Further land-use analyses for each household type revealed differences in patterns of land-use choice. As such, the
cultivation of cash crops had a higher proportion among the rich and middle class farmers, whereas the poor farmers had a tendency to focus on subsistence crops. Moreover, there was an imbalance of irrigation practices among the identified livelihood groups, i.e. the percentage of irrigation farmers in general and pump farmers in particluar increased with the level of living/livelihood standard.
After the derivation of livelihood groups, sub-models for land-use choice were pre-sented and calibrated, whereby the range of explanatory variables and the choice of model were justified, and the results presented. These sub-models include the choice between rainy-season and dry-rainy-season land-use types, the decision to do irrigation farming, and the choice of irrigation method. All decision models were developed on the basis of m-logit regres-sion, apart from the choice among dry-season land-use types, as no meaningful variable set could be identified to explain choices among land-use types in this season. The preference coefficients for the m-logit model for rainy-season land-use choice were determined for each livelihood group separately, since the results of a descriptive comparison of land-use pref-erences among livelihood groups suggested the relevance of such a differentiation. These differences in land-use choice are reflected by the differences in the direction, magnitude and significance of the preference coefficients, which clearly show considerable heterogeneities in local land-use choice behavior. In general, households of all groups choose land-use types based on the considerations of a range of household characteristics, natural conditions and particular policy factors.
With respect to the modeling of irrigation-related decisions, a group-wise approach was considered to be unreliable due to the relatively small sample size of irrigation farmers, which did not allow any further splitting. Instead, the preference coefficients were computed for the total population, which turned out to be the more robust approach. These irrigation-related decisions were modeled as a nested m-logit model, which included the decision to do irrigation as a first step, and as a second step, the choice of irrigation method. Both environ-mental and household characteristics as well as policy factors were included as explanatory variables within this nested model to reflect the socio-economic as well as the environmental conditions necessary for the engagement in irrigation.
The results and structure of these land-use choice models were integrated into GH-LUDAS within the Decision Module The preference coefficients were used to compute the
land-use choice probabilities/utilities, whereby each land-use option during model run is se-lected by an agent with its respective probability, thus allowing bounded rational decision-making behavior.
5 ECOLOGICAL DYNAMICS OF HETEROGENEOUS LANDSCAPE AGENTS