In lieu of an abstract, here is a brief excerpt of the content:

appendix: chapter 4 data and principal component variables The surprisingly large amount of information available for Bolivia during the period 1987–2007 demands a strategy for choosing, from among 1,200+ variables, those that are most appropriate and most closely related to the underlying concepts I wish to test. In particular, a number of measures in which I am interested are present in my data set as multiple, ‹nely differentiated variables. I have data on, for example, 16 varieties of capacity-building exercises undertaken by municipalities and 13 different local actors who assisted in drafting municipal development plans. The challenge is to reduce such groups to at most one indicator each without loss of information. My empirical strategy is iterative and begins by ‹nding the best idiosyncratic model of public investment for each of the 10 sectors of interest. I ‹t the equation Gm = ζSm + ηZ + εm (A1) separately for central public investment (pre-1994) and local public investment (post-1994) where Gm is aggregate investment per capita in the public good subscripted by municipality, Sm is a scalar or vector of the existing stock of public goods of that type (variously de‹ned) at an initial period, and Z is a vector of socioeconomic, demographic, regional, political, institutional , administrative, and procedural variables that might affect investment decisions. The use of the Z term follows the literature on the demand for public goods exempli‹ed by Bergstrom and Goodman (1973) and Rubinfeld , Shapiro, and Roberts (1987) within the context of the available data. In particular, no income data is available at the municipal level in Bolivia, and so I substitute several alternative indicators of income and wealth, for example, type of cooking fuel and housing size, quality, and related characteristics . But I expand the scope of the Z vector considerably compared to 291 previous authors by including measures of the strength of local political forces as well as municipal institutional capacity. This innovation allows me to investigate the micropolitical basis of local government decision making, explored in detail by Faguet (2008b). No constraints across sectors are allowed on the particular variables admissible in Z. I use the Huber/White estimator of variance to produce consistent standard errors in the presence of non–identically distributed residuals . This produces 10 different models of public sector investment, one for each sector. Individually these models are quite satisfactory, with high R2 and few variables insigni‹cant. But because of large variation in the speci‹cation of the Z vector, comparison across sectors is problematic. In addition, on a theoretical level these models would seem to assert that public investment in different sectors happens according to different processes, in which different variables intervene. This is evidently unsatisfying. In a second iteration I reestimate equation (A1) holding the Z vector constant across all sectors. But I take advantage of the previous stage by using only those variables found signi‹cant there; in this sense the previous stage constitutes a method for reducing the 1,200+ indicators to a subset of 197. But a dimensionality problem persists even so. I then employ a method of forward and backward substitution and elimination in order to reduce this subset to 22 variables encompassing the 13 categories of Z, in speci‹cations of 23–30 variables overall. These models bene‹t from being readily comparable across sectors. The ratio of signi‹cant to insigni‹cant variables drops sharply compared to the ‹rst stage, however, and R2 values are somewhat lower. The insigni‹cance of the variables chosen is not entirely separable from the issue of comparability, however. In these results none of the variables is signi‹cant in most of the sectors, and many are signi‹cant in only two or three. How do we interpret a given variable across sectors, knowing that an alternative one from the same group would produce a different pattern of signi‹cance and insigni‹cance? For example, how do we interpret the insigni ‹cance of training & capacity-building variables in most models when we know from stage 1 that there is at least one alternative such variable that is signi‹cant in each sector? We evidently cannot assert for any sector that capacity-building does not matter; we must conclude that the comparability constraint forces us to omit from our models information that is important in explaining investment behavior. Indeed, given that there are 197 variables, many of them quite speci‹c, 292 appendix [18.221.187.121] Project MUSE (2024-04-16 10...

Share