Data analysis of the choice experiment

Two types of data were gathered during the field research in Shida Kartli, Georgia: i) socio-economic data and ii) data obtained with the choice experiment. This section depicts analysis methods of the choice experiment. Programmes used for the analysis were Excel, NLOGIT 3.0, SPSS 13.0, and Latent Gold Choice. Besides a section on the socio-economic situation in Shida-Kartli, another section of the questionnaire contained the choice experiment. In the choice experiment, each respondent received four choice cards offering two loan alternatives and a status quo alternative, for a total of twelve alternatives per respondent. Before completing the choice experiment, respondents had to choose between two types of loans: loans with joint liability and loans with individual liability. Only a small share of respondents (8 percent) chose loans with joint liability; therefore the choice experiment for this loan type was not analysed. The outcomes of the choice experiment were put in an Excel file in order to prepare these data for the first analysis with NLOGIT 3.0. The choice data set was then transferred to NLOGIT 3.0 for analysis. The first step was to examine the data with logit analysis. The following model, an indirect utility function, was used to analyse the data:

U(c1, c2) = bASC*ASC + bLOS*inlos + bINT*ininte + bCOL*incoll + bINS*ininst + bCOM*incomm + bLOD*inlod

U(c3) = bASC*ASC

U in the first model stands for the utility, which is produced by the two choice alternatives c1 and c2. The choice alternatives c1 and c2 stand for the two loan alternatives A and B depicted on each choice card. U is the dependent variable. Each choice alternative has six attributes written as independent variables and multiplied by their related influence parameters bASC and bLOS, etc. The influence parameters are the betas. U in the second model signifies the utility coming from choice alternative c3, the alternative-specific constant (ASC), which is multiplied by its influence parameter bASC. ASC is the status quo or ‘neither loan on the choice card’ alternative.

Train (2003 p. 24) defines the ASC as follows:

The alternative-specific constant for an alternative captures the average effect on utility of all factors that are not included in the model. Thus they serve a similar function to the constant in a regression model, which also captures the average effect of all unincluded factors.

The complete model is written

U(c1,c2) = bASC*ASC + bLOS*inlos + bINT*ininte + bCOL*incoll + bINS*ininst + bCOM*incomm + bLOD*inlod /U(c3) = bASC*ASC

The single attributes or independent variables are

- ASC: Alternative-specific constant

- Inlos: Loan with individual liability, loan size

- Ininte: Loan with individual liability, interests

- Incoll: Loan with individual liability, collateral

- Ininst: Loan with individual liability, instalments

- Incomm: Loan with individual liability, commission

- Inlod: Loan with individual liability, loan duration

Several types of logit analysis were executed with the model presented above. The first was multinomial logit analysis, which was conducted in NLOGIT 3.0. According to Hensher & Greene (2003), the multinomial logit model (MNL) should always be used as starting point for empirical investigation. First of all, it was necessary to determine whether independence from irrelevant alternatives (IIA assumption), a precondition for the MNL, exists in the given choice data set. The IIA assumption states that the ratio of the probabilities for any two alternatives stays independent if any or all of the remaining alternatives are removed or added. Independence from irrelevant alternatives exists if the result of model estimation with reduced alternatives does not deviate from the complete model. To test for violations of the IIA assumption, Hausman-McFadden (1984) tests were performed. They estimate first the complete model with all alternatives and next a restricted model with a smaller number of alternatives (Hensher et al. 2005 p. 519; Urban 1993 p. 133). As independence from irrelevant alternatives of two alternatives is assumed, the Hausman-McFadden test permits the simultaneous removal of more than one alternative in the restricted model (Hensher et al. 2005). While systematically comparing the complete model to the reduced model, the Hausman-McFadden test calculates whether the logit results will be influenced significantly by the model specification (complete or reduced model). To run the test, a null hypothesis is employed stating that there is no difference between the complete model and the reduced model. If the null hypothesis cannot be rejected, the IIA assumption holds. For hypothesis testing an asymptotically chi-square

distributed test statistic that can be verified with a significance test (Urban 1993) is used.

In addition to analysis of the preferences of all respondents in one group with the multinomial logit model, respondents’ preferences for loan attributes were grouped into classes. For this purpose, a latent class model (LCM) was estimated. The classes were calculated with Latent Gold Choice, a latent class modelling software. Four classes of preference types were distinguished.

After the multinomial logit analysis and the estimation of the latent class model, interactions between selected socio-economic and opinion variables and the seven CE attributes were calculated. The following variables were chosen:

1. Whether the respondent took a loan or not 2. How the borrowed amount was invested

3. Loan size of an individual loan according to respondent’s free statement 4. Whether respondent is familiar with financial systems or not

5. Respondent’s degree of certainty with regard to his/ her choice in the CE 6. Importance of loan size for respondent

7. Importance of implementation of a rural credit system 8. Likelihood of implementation of a rural credit system 9. Respondent’s gender

10. Respondent’s age

11. Respondent’s maximum level of education 12. Respondent’s main job

13. Respondent’s main income source

14. Kind of agriculture respondent engages in 15. Person in the household who owns the land 16. Area of agricultural land

17. Monthly household income

18. Person who decides on money use in the household 19. Expectation of income development

20. Importance of individual loan’s size for the respondent 21. Importance of individual loan’s interest for the respondent 22. Importance of individual loan’s collateral for the respondent

23. Importance of individual loan’s instalment frequency for the respondent

24. Importance of individual loan’s commission for the respondent 25. Importance of individual loan’s duration for the respondent

Each of these variables was interacted with all the loan attributes of an individual loan, as well as with the ‘neither loan’ alternative (ASC). The attributes were as follows:

1) Loan size, 2) interest, 3) collateral, 4) instalments, 5) commission, and 6) loan duration.

As only a small number of interactions were significant, six socio-economic key variables were re-coded into dummy variables to test whether the original code system, which included up to seven code numbers per variable, was responsible for the deficiency of significances. From the list above, the following socio-economic variables were re-coded:

11) respondent’s maximum level of education, 12) respondent’s main job, 13) respondent’s main income source, 14) kind of agriculture respondent engages in, 17) monthly household income, and 19) expectation of income development. The new codes were for variable 11) ‘university degree or other degree’, for variable 12)

‘farmer or off-farm economic activity’, for variable 13) ‘agriculture or off-farm income’, for variable 14) ‘fruits or vegetables’, for variable 17) ‘up to 200 lari or more than 200 lari’, and for variable 19) ‘increasing or falling expectation of income development’.

To measure respondents’ tendencies with respect to their willingness to pay for a loan, point elasticities for the attributes ‘interest’ and ‘commission’, which make up the cost of a loan, were calculated with regard to choice alternatives 1 and 2. Point elasticities are direct elasticities that measure the percentage change in the probability of choosing a particular alternative in the choice set with regard to a given percentage change in an attribute of that same alternative (Hensher et al. 2005). For example, if the interest in Alternative 1 increases by 1 percent, how much does the possibility of choosing Alternative 1 change in percent?