Small Works: Poverty and Economic Development in Southwestern China

Appendix

173 Appendix Methodology and Case Selection Faster growth is normally better for the poor than slower growth, and is not systematically offset by any change in distribution. But huge exceptions—and the possibility of clusters of countries where growth is much better for distribution, or much worse—mean that these findings are the beginning, not the end, of the inquiry. Residuals matter. —Robert Eastwood and Michael Lipton (2001) The cases I use in this volume—Guizhou and Yunnan—were selected according to expectations about the hypothesized relationship between economic growth and poverty using the most complete and reliable data available—that provided by the World Bank. In this appendix, I outline the case selection process and other methodological issues. I first introduce and execute the regression analysis used to choose the two deviant cases. Because this regression is based on two data points, and is thus highly sensitive to data errors, considering possible sources of bias in the data is especially imperative. To reduce my reliance on a single data source, I also explore here alternative data sources that increase our confidence in the validity of the puzzle. Deriving data from multiple qualitative and quantitative sources allows us to cross-check and verify each source. Because my research explores numerous variables as potential explanations for only two cases, I also address the question of whether the research design violates the degrees-of-freedom problem. Case Selection, Bias, and Regression Analysis What is the best way to choose exceptional cases for study? There are numerous ways to select cases, most of which do not fulfill this goal (see table 10). First, many social science studies select cases for practical reasons, including 174 Appendix previous expertise in the area, limited time, or limited financial support. Although this is understandable, cases selection using nonrigorous methods is often more biased than choosing cases more systematically. Second, there is random selection; this chooses typical cases, not atypical ones— the opposite of what we seek. Moreover, random selection is not appropriate in small-N studies because the risk of selecting a strange and nonuseful mix of cases increases as the number of cases decreases (Gerring 2007). Third, some scholars consult secondary research or panels of experts to select cases, an innovative method that increases the numbers of experts focused on finding relevant cases (e.g., Cimadamore et al. 2002). This method, however, risks following conventional wisdom and adopting the biases of others, thus potentially missing theoretically interesting cases that experts have overlooked. Fourth, cases can be selected because they have extreme values in their dependent variables, thereby indicating a high degree of success or failure. Selecting extreme cases of critical phenomena can be valuable in understanding the causes of rare but important events. But, if the extreme values occur with equally extreme values on potential explanatory factors, then conventional wisdom explains the phenomenon and nothing novel is uncovered. Fifth, cases can be selected using a typological design in which cases are divided into groups based on key criteria; cases are then selected because they fall into theoretically interesting categories (Bennett and George 2005), such as the studies of burden sharing during the first Persian Gulf War (Bennett, Lepgold, and Unger 1994). This method, although designed to select exceptional cases, often requires extensive knowledge of the relevant factors for each case, which is often unavailable or difficult to obtain. Moreover, by combining cases into categories, this method is most applicable when the data are in nominal or ordinal form. If reliable integer or continuous data exist, typological design may not be superior because it divides the independent variables into categories when continuous variables might be more accurate. In my research, I combined the three remaining case selection methods listed in table 10. First, I followed the most-similar-cases design pioneered by Adam Przeworski and Henry Teune (1970), Arend Lijphart (1971), and Alexander George (1979), selecting paired cases that were similar in as many exogenous factors as possible and, thus, controlling for them. This method chooses cases that differ primarily in the factors of interest and that are similar in many factors not of interest, increasing rigor. Second, I simultaneously focused on diverse cases. Although the background conditions for each case are similar, the variables of interest diverge. Kohli’s (2004) study of Korea, India, Brazil, and Nigeria also uses this method. In contrast with selecting cases with extreme values, selecting diverse cases allows a broad variance in both the independent and dependent variables. Neither of these...