Research That Matters (January 17 - 20, 2008)

Purpose The Beautiful Foundation conducts a biennial national survey on giving and volunteering behaviors of Korean citizens. Researchers have interests on analyzing the amounts of giving and the hours of volunteering with employing diverse multivariate analysis methods. However, a methodological question is asked: How much do data analysis methods matter? On the measurement and analysis of total formal giving and volunteering, there should be cautions in selecting more appropriate analysis methods with the consideration of selection bias. So this paper attempts to examine significance of predictors on the dependent variables with comparing the results by different data analysis methods.

Methods In analyzing the dependent variables, it should be remembered that the error terms in giving and volunteering do not obey the classical assumptions that justify the use of ordinary least squares. Donations and volunteering hours cannot be negative, so the error terms have a truncated distribution. In addition, giving and volunteering data appear to have a non-normal error structure. Under these circumstances, OLS is biased and inconsistent. Because there is no commonly accepted ideal remedy for these problems, this paper attempts to report results from four different approaches as follows: Tobit, Heckman two-stage, OLS on the full sample, and OLS on positive donors only (OLS on positive volunteers only). Tobit has the following advantages: (a) it eliminates selection bias and (b) it does not generate negative predicted donations. But it has the following disadvantages: (a) it is not robust to non-normal or heteroskedastic errors; and (b) it enforces proportionality between a variable's effect on the probability of giving (volunteering) and the sizes of the donations of those who give (of the volunteering hours of those who volunteer). Heckman two-stage has the following advantages: (a) it eliminates selection bias and (b) it separates a variable's effect on probability of giving from its effect on the amount given (of volunteering from its effect on the hours volunteered). But it has the following disadvantage; it is not robust to non-normality. OLS on donors only (volunteers only) has the following advantage: it is robust to non-normal errors and heteroskedasticity. But it has the following disadvantage: it suffers from selectivity bias. On the other hand, OLS on full sample has the following advantage: it is robust to non-normal errors and heteroskedasticity. But it has the following disadvantage: it suffers from truncation bias because it assumes a symmetrical distribution, including the possibility of negative giving and volunteering.

Results Employing these multivariate data analysis methods, this paper has the finding that there is a converging picture from the results of Tobit and Heckman second-stage. On giving amount, the following variables are significant: household income, volunteering, parental charity, and religious affiliation. On volunteering hours, the following variables are significant: age, parental volunteering, and religious affiliation.

Implication The result indicates that in predicting amount of giving and hours of volunteering, there should be more consideration on the issue of selection bias. On the result, this paper makes discussions on methodological issues and some directions for future study.