A Life Course Perspective of Community (Non)Investment: Historical Financial Service Trajectories and Community Outcomes

Abstract

Drawing on techniques more commonly used to study changes within families over time, this paper highlights how holistic life course methods can help scholars conceptualize the dynamic nature of local built environments and measure impacts for families and communities. I use a novel dataset on the historical availability of banks, credit unions, and alternative financial services (AFS) between 2003 and 2015 to classify neighborhoods by their financial service trajectories using sequence and cluster analyses. I identify six distinct trajectories of financial service availability over the 13-year period; neighborhoods in these trajectories differ in terms of their socioeconomic and demographic characteristics. Descriptive multivariate analyses confirm that trajectories are linked to community outcomes at the end of the period; tracts exposed to AFS at some point over the 13 years are associated with higher predicted end-of-period poverty rates compared to both tracts that are only exposed to banks and credit unions and tracts that are chronic financial service deserts. Extensions of this approach to other aspects of the built environment are discussed.

Appendix 1

Comparison of Four, Five, and Six Clusters of Financial Service Trajectories in Five-Percent Sample, using Optimal Matching Sequence Analysis

The paper evaluates financial service trajectory characteristics and associations with poverty using six trajectory groups developed in a two-stage process using sequence and cluster analyses. Both sequence and cluster analyses involve researcher decisions about which number of clusters are most appropriate for the final set. To identify the best number of groups, I start with a 5% random sample of tracts (n = 3624), all of which have 13 years of data. In each year, the tract is categorized into one of the six categories of financial service presence (e.g., financial desert, heavy AFS presence, only banks). I use optimal matching sequence analysis to identify trajectories of financial service presence over time, wherein a full dissimilarity matrix of every pair of sequences is identified (Cornwell, 2015). Optimal matching identifies the lowest cost of making two sequences the same; for this analysis, I set the costs of individual insertions or deletions to make identical sequences to one. The resulting dissimilarity matrix is then used as the basis for cluster analysis that sorts sequences into distinct groups. Using results from the cluster analysis, I compare four, five and six trajectory clusters. Comparisons are presented below in Table 5 and Fig. 5.

Fig. 5

Comparisons of trajectory clusters, 5% sample. N = 3642 Census tracts; coverage includes a five-percent random sample of US tracts. Y-axes vary across panels within each figure to facilitate visual comparison across figures

Table 5 indicates that the trajectory of Non-Investment is stable across the three options. Figure 5 shows that these tracts are overwhelmingly financial deserts for the entire period, with a small share of tracts experiencing some financial service investment for a short period of time. Similarly, AFS-Only Investment tracts remain stable, and Fig. 5 confirms that these tracts consistently have AFS-Only Investment the duration across cluster options. Differences across four, five and six trajectory options emerge when considering other experiences of AFS and Bank investment. When using four clusters, one group emerges as Mixed Investment, making no distinction between tracts that have mostly AFS investment, with some Bank investment, and those with mostly Bank investment, with some AFS. Expanding to five clusters again fails to make a distinction between these qualitatively different experiences, but it does distinguish between tracts that have Bank-Only Investment for most but not all of the time and tracts that have uninterrupted Bank-Only Investment for the duration (see second panel of Fig. 5. Using six groups splits out mixed investment tracts into two groups, those that have an even Bank/AFS mix, and those with either heavy bank or heavy AFS investment during the period. This latter group is still conflating two qualitatively different investment types, at least in terms prior research showing diverging associations with community characteristics. However, splitting into seven or more groups would risk making the counts of tracts in one or more groups too small to meaningfully represent common experiences. Instead, I address this concern when moving to cluster analyses for the full sample by incorporating four additional indicators (see next section).

Table 5 Count of tracts in each trajectory, by K trajectory clusters (n = 3624 census tracts)

Comparison of Four, Five and Six Clusters of Financial Service Trajectories in Full Sample, Using K-Medians Cluster Analysis

The results from the 5% sample are useful as I move to the full sample to repeat the comparisons using cluster analysis only. In the full sample analysis, I include additional terms for entries into and exits from communities for both banks and AFS services as discussed in “Methods” section of the main paper. This helps to account for diverging motivations and timing of entries and exits, and ultimately results in better separation between AFS and Bank investment among tracts with “mixed investment”. The final cluster analysis incorporates all US Census tracts and uses the 13-year temporal sequence of financial service presence alongside the four indicators of financial service entries and exits. The three columns in Table 6, and the three panels in Fig. 6 can be compared to identify what happens when the number of trajectory groups changes (where k clusters = 4, 5, or 6). These figures are produced after conducting k-median partition cluster analysis, with the first k observations as the starting centers for analysis (as opposed to using a random seed each time; see https://www.stata.com/manuals/mvclusterkmeansandkmedians.pdf for details).

Fig. 6

Comparisons of trajectory clusters, full sample. N = 72,482 Census tracts; coverage includes all U.S. states. Panel A2.3 is the set of six trajectories presented in Fig. 1 in the main paper. Y-axes vary across panels within each figure to facilitate visual comparison across figures

Table 6 Count of tracts in each trajectory, by k trajectory clusters (n = 72,482 census tracts)

Using just four clusters of trajectories (first panel of Fig. 6) produces three groups that are similar to groups that appear in the final six trajectory set (third panel of Fig. 6)—Chronic Non-Investment, AFS-Only Investment, and Bank-Only Investment. However, the remaining category, Mixed Investment, does not distinguish between tracks that experience persistent heavy AFS investment and those that experience heavy/only Bank investment for the majority of the period. In other words, using four clusters results in a category that does not distinguish between the relative concentration of Banks and AFS for communities that experience both types of investment. Using five trajectory clusters (second panel of Fig. 6) improves on the four groups by largely splitting Mixed Investment into two groups, one that includes a set of communities experiencing sustained AFS investment for the duration of the period and another includes communities experiencing sustained Banking investment over the period.

Using six clusters (Fig. 6; see also Fig. 1 in main text) adds a final group—communities that see delayed banking investment, where the vast majority of tracts starting out as financial deserts and then witnessed banking investment in 2007 or later. Given prior literature noting negative associations between banking deserts and community economic characteristics (e.g., Hegerty, 2016), I expect that delayed banking investment may have a qualitatively different relationship with community characteristics, and thus I choose to keep six clusters instead of five. With six clusters, the smallest trajectory group comprises about 5% of US tracts. I elect to stop here to avoid splitting tracts into groups that are too small to meaningfully represent common trajectories during the analytic period.