Using data collected from neighborhood audits and GIS-derived variables, we sought to confirm neighborhood constructs developed by others. The CFA analysis of the items that composed the a priori constructs physical incivilities, territoriality and social spaces from Caughy et al , and safety, aesthetics, destinations, and functionality from Pikora et al , indicated that the items composing a priori constructs did not hold in this geographic area with the exception of physical incivilities. Therefore we moved to EFA, where a four-factor solution was derived that included the following constructs: arterial or thoroughfare, walkable neighborhood, physical incivilities, and decoration. These constructs performed well in the CFA on the validation sample for both urban and rural road segments. Two-week test-retest reliability ranged from moderate to almost perfect for all except the decoration factor in rural areas.
A priori Caughy et al Constructs
The three constructs from Caughy et al  (physical incivilities, territorality, and social spaces) were developed from a neighborhood audit conducted in Baltimore, Maryland. For the physical incivilities construct, we collected similar items to the original construct, with the addition of visible dogs. All of the a priori items loaded on both urban and rural segments, with the exception of the presence of visible dogs which was important in rural but not urban segments. When we moved to the EFA, the physical incivilities construct derived from the data included four of the original eight items, in addition to pedestrian oriented lighting and no trespassing sign. The physical incivilities construct appears empirically represented by these data. In prior research, physical incivilities was associated with increased levels of crime  and pregnancy-related behaviors .
The other two constructs from Caughy et al  did not hold in our data. For the territorality construct, we included similar items to the original index with the exception of two items. We did not collect whether residents reacted to the presence of raters, as rating was performed from a car rather than by walking. We also did not collect presence of security bars, as almost no homes had security bars in our study area. Similar to Laraia et al , we added several other items pertaining to signage when we explored the territorality construct. For the social spaces construct, we collected similar original items and expanded it to include presence of porches and sidewalks, also similar to Laraia et al .
These differences may be why the constructs did not hold in this geographic area. Both of these constructs (e.g., territorality and social spaces) were developed for use in the urban northeast US, where population density, park accessibility, and foot traffic patterns differ from the suburban and rural southeast US. Both constructs rely on specific types of indicators (e.g., short walls) and natural opportunities for social interactions (e.g., playgrounds) that were not as often present in our region. Documenting that these constructs do not function as expected is an important finding of this work, and suggests future development work in these areas. Empirically identifying items more specific to a latent construct shared across different types of neighborhoods for each study area may be an important undertaking not only to understand how the latent construct may manifest itself, but also as a data reduction technique that will minimize error.
A priori Pikora et al Constructs
Considering the walking and bicycling framework by Pikora et al , the original instrument developed from this framework in Australia was titled SPACES  and was adapted to the US in other studies [29–32] to study walking and bicycling. Working from this instrument, we developed and modified items on the audit with applicability to rural and urban areas of central North Carolina. The adaptation of the SPACES instrument, such as revising questions, dropping items with low prevalence, may have been why the a priori constructs did not hold. Several items were dropped rather than modified from the original SPACES instrument due to measurement concerns, including whether the path formed a direct route or continuous route to destinations (functional construct), driveways and permanent obstructions in the path or lane (safety construct), and pollution (aesthetics construct). In addition, the original SPACES audit was conducted on foot, whereas our audit was conducted from a vehicle. Although the original constructs did not hold, many of the items mapped to newly derived data driven constructs. These changes to the instrument were done for use in our geographic area, but in turn may have compromised a direct test of replicability. It is possible that these original constructs may hold in geographic areas more similar to Perth, Australia where the tool was first developed.
Our analysis suggests some items that may be redundant or that did not contribute to any factor, including several secondary GIS measures (short road segment and steep segment) and a number of audit based measures including type of front yard (#6), security warning sign (#10), industrial land (#13), agricultural land (#14), home business (#18), vacant or underdeveloped land (#19), graffiti (#29), footpath (#33), trees (#35), road oriented lighting (#36), neighborhood crime watch (#43_13), beware of dog or invisible fence signs (#43_15), and signs for cars regarding bike/pedestrian (combined index from #43_2, #43_6, and #43_7). Several of these items were newly added, to capture features in more rural environments, but were found to not contribute to any underlying factor. However, they still may be important in other parts of the country and still may be predictive of physical activity as independent items.
Jago et al  also used a modified SPACES instrument to assess neighborhoods in Houston, Texas. Similar to our work, they dropped items with low variability and items that did not load on any factor. Using principal component analysis, they also dropped items that loaded on more than one factor. The remaining four data driven components accounted for 49% of the variance and included walking/cycling ease, tidiness, sidewalk characteristics, and street access/condition. These constructs also differed from the originally envisioned SPACES constructs, and the authors noted that this may be due to differences in measurement (self-reported ideas vs observed environmental data). These findings, together with our own, may indicate that a universal audit instrument may lose local heterogeneity.
Considerations for Data Collection and Analysis
In the process of collecting, processing, and analyzing data using our PIN3 Neighborhood Audit data, several important lessons were learned that should be considered by others using this type of process. First, the factor analysis could not easily accommodate nominal variables with more than two levels. Thus, for questions such as condition of resident units, resident grounds, vacant/underdeveloped land, or public spaces, we added a "mixed condition" choice to describe conditions with extreme differences on the same road segment. While this may have better captured the characteristic of the road segment, it changed the previously ordinal variable to a nominal one. In many cases we had to collapse these variables into fewer categories. Future studies should consider whether nominal response options with more than two levels could be reworded to overcome this limitation of current software.
Second, the issues of missing data must be considered, whether due to incomplete answers or intended skip patterns. In our dataset, we had virtually complete data because we used a handheld device to collect the data. Where data were missing that was not due to a skip pattern, we treated it as missing at random. The PIN3 Neighborhood Audit tool included 7 intentional skips and in our analysis we either coded the missings to "not present" or treated the missings as another response option, depending on the specific variable. As these audit instruments evolve, consideration should be given to whether or not items with forced skip patterns are used. Intended skip patterns are acceptable, but at present the resultant variables will be nominal.
Third, consideration should be given to the type of rotation used in the factor analyses. Using orthogonal rotation forces the constructs to be statistically uncorrelated which is advantageous for subsequent use in statistical modeling as well as its simplicity and conceptual clarity . In contrast, the oblique rotation allows factors to be statistically correlated. While this may better represent reality, since these neighborhood constructs may be intercorrelated, it also adds statistical complexity to future analyses. A priori, we had decided to derive the factors orthogonally because it simplified future analysis and this is what is presented in the results. However, we also derived correlated constructs using the Promax (oblique) rotation. The factor loadings were similar to those from Varimax (orthogonal) rotation, and Pearson correlations between constructs were low (range: -0.21, 0.16). Thus, using either an orthogonal or oblique rotation, our final data driven constructs were consistent.
Limitations and Strengths
This study was conducted in four counties in central North Carolina and the road segments collected were based on where the PIN3 participants lived. This study included a large number of road segments (26% of all street segments in a 4-county area that included 7,150 miles of roads) with enough geographic diversity to explore urbanicity. However, our sample was neither random nor complete for the geographic area. Thus, it is not known how these results might generalize to other areas. While we expanded the use of the audit instrument to more rural areas, these segments were located within proximity to more urban areas; it is not known if results would change if this audit was conducted in more isolated rural geographies. Also, neighborhood observation represents a cross-sectional snapshot of the community and may miss neighborhood dynamics that change over time . A strength of this study is that we collected test-retest reliability to further describe the measurement properties of our derived constructs. This study also collected items based on an a priori framework from Caughy et al  and SPACES .