Life Satisfaction and Economic Position Relative to Neighbors: Perceptions Versus Reality

Abstract

I examine the relationship between economic position relative to local neighbors and life satisfaction in rural Bangladesh. In particular, I exploit a novel household level census of three villages that includes the geospatial coordinates of every household and a perception measure of economic position relative to neighbors. This allows for exploring the sensitivity of the aforementioned relationship to (1) different objective definitions of neighborhood, and (2) the type of positional measure used, objective or perceived. I find that a higher perceived position improves life satisfaction while objective position has no statistically significant effect. This difference stems from a very low correlation between the objective and perceived positional measures. The low correlation can be explained by individual specific heterogeneity in definitions of neighbor. However, observable socioeconomic and demographic variables cannot explain the difference.

Appendices

Appendix 1: Brief Explanation of Hierarchical Clustering with an Average Link

I choose the hierarchical clustering procedure since it is a popular method of clustering with a long history in the literature. It also provides some objectivity by not requiring me to make any subjective choice of initial seeding that some other clustering techniques necessitate (e.g. K-means). Assuming there are N distinct households, the hierarchical clustering technique involves N iterative steps with each step producing a unique way of dividing all the households into separate clusters. The first step defines each of the N household as a cluster. The second step calculates the distance between every pair of households (or clusters) and then identifies the two households with the shortest distance between them as a single cluster while the rest of the households continue being treated as their own cluster (leading to $(N-1)$-cluster neighbors). In the third step, the “average” distance between every cluster is calculated and the two clusters with the shortest average distance between them merged into one cluster.²⁴ This process is repeated until all households form a single cluster in the Nth step. Thus, the J-cluster neighbors refer to neighbors defined in the $(N-J+1)$th step and comprising of J clusters.²⁵

There are 1395 possible J-cluster neighbors (where $J = 1,2,\ldots ,1395$) associated with the hierarchical iterative process explained before in the data used for this paper. However, for higher values of J these are not very useful because there are too many single household clusters. Hence, I do not consider J-clusters where $J > 200$. Under the 200-cluster neighbors, there are 27 households that form singleton clusters while 43% of the households are located in clusters with less than 10 households. I also do not use 1-cluster because this makes HES and the ratio of HES to average neighbors’ HES perfectly collinear.²⁶

Appendix 2: Generalized Ordered Logits

Equation (3) summarizes the GOlogit associated with Eq. (1):

$$\begin{aligned} \begin{aligned} P(y_i > j) = \frac{\exp (\rho _{0j} + \rho _{1j} g_1(E_i) + \rho _{2j} g_2(R_i) + \sum \nolimits _{k=3}^K \rho _{kj} X_{ik})}{1 + \left[ \exp (\rho _{0j} + \rho _{1j} g_1(E_i) + \rho _{2j} g_2(R_i) + \sum \nolimits _{k=3}^K \rho _{kj} X_{ik})\right] } \quad \forall j = 1,2,\ldots ,M-1 \end{aligned} \end{aligned}$$

(3)

Note that the model parameters $\rho _0,\rho _1,\ldots ,\rho _K$ from Eq. (1) have an additional subscript of j in Eq. (3). This embodies the generalization that the parameters may vary across reported $y_i$ i.e. the “proportional-odds” assumption that is standard in ordered logits (Ologits) is relaxed. Indeed, relaxing the proportional-odds constraints (POCs) is equivalent to assuming the threshold cut-offs in ordered logits are heterogeneous and are weighted linear combinations of the independent variables. If the POCs are imposed for all variables, it becomes the standard Ologit. The interpretation of the regression coefficients and related odds ratio of GOlogits can be a bit tricky. Without loss of generality, for each regressor $\theta$, the GOlogit estimates $(M-1)$ different regression coefficients.²⁷

Generally, if the dataset has a large number of observations using a GOlogit without imposing any POCs (GOlogitF) is preferred to implementing the more parametrically parsimonious Ologit. Given the TS has a limited number of observations, running a GOlogitF entails estimation of additional parameters that may reduce efficiency unnecessarily if the true data generating process (DGP) imposes all of the POCs. However, if the true DGP is one where some of the POCs do not hold, then Ologits produce biased estimates.

A GOlogit which relaxes only those POCs that are violated (GOlogitP) is a practical compromise between an Ologit and a GOlogitF. Brant (1990) proposes using Wald tests to identify which POCs do not hold. I use the iterative autofit process developed by Williams (2007), an extension of the “Brant” test, to choose which POCs to relax when running the GOlogitPs. In particular, running the baseline regressions and the perception regressions with autofit reveal that only POCs associated with the village dummies and the perceived position variables are violated. Consequently, I relax the POCs for only the village dummies and the objective position measures when running the objective regressions.