The purpose of the present analysis was to use longitudinal data collected over 7 years (from 4 surveys) in the Residential Environments (RESIDE) Study (Perth, Australia, 2003–2012) to more carefully examine the relationship of neighborhood walkability and destination accessibility with walking for transportation that has been seen in many cross-sectional studies. We compared effect estimates from 3 types of logistic regression models: 2 that utilize all available data (a population marginal model and a subject-level mixed model) and a third subject-level conditional model that exclusively uses within-person longitudinal evidence. The results support the evidence that neighborhood walkability (especially land-use mix and street connectivity), local access to public transit stops, and variety in the types of local destinations are important determinants of walking for transportation. The similarity of subject-level effect estimates from logistic mixed models and those from conditional logistic models indicates that there is little or no bias from uncontrolled time-constant residential preference (self-selection) factors; however, confounding by uncontrolled time-varying factors, such as health status, remains a possibility. These findings provide policy makers and urban planners with further evidence that certain features of the built environment may be important in the design of neighborhoods to increase walking for transportation and meet the health needs of residents.

Editor's note:An invited commentary on this article appears on page 462,and the authors' response appears on page 467.

Over the past decade, there has been more research examining features of the built environment that enable, facilitate, or are barriers to improved health (15). A strong focus of this research has been an attempt to understand the relationship between the neighborhood built environment and physical activity levels (2, 57), and an effective strategy has been to estimate the association of neighborhood built environment with neighborhood walking (8). In particular, measures of neighborhood walkability and the number and diversity of destinations (land-use mix) that are accessible by walking from home have received considerable attention (913). In a recent review (14), 80% of the studies reported a positive association between walking and the presence and proximity of retail and service destinations. Overall, built environment studies have reported more associations with transport-related walking than with other types of physical activity (11). Since the 1996 US Surgeon General's report (15) was published, promoting walking has been a global priority, with public health agencies actively promoting walking for transportation (16).

However, most studies to date have relied upon cross-sectional analyses; few studies had a longitudinal design (1720). Cross-sectional studies are subject to confounding by unmeasured and thus uncontrolled factors, including factors related to why people choose to live in certain neighborhoods (often called self-selection factors) (21). Longitudinal or repeated-measures data from a cohort of individuals allow consideration of temporal order and the examination of within-person changes in both built environment factors and physical activity behaviors. These within-person relationships, if analyzed appropriately, are not subject to confounding by time-constant (selection and other) factors, as each individual serves as his or her own control (22, 23). Thus, longitudinal data provide better evidence for likely causal relationships between the built environment and physical activity behavior if they are analyzed with appropriate regression models and if there is sufficient variation over time within as well as between individuals.

The Residential Environments (RESIDE) Study provides a unique opportunity to examine longitudinal data from a cohort of individuals and provide stronger evidence of a causal relationship between the built environment and transport-related walking. Importantly, the study has been designed to answer the question of self-selection (21, 24, 25). The primary aim of the present analysis was to use longitudinal data over 7 years (4 surveys) from the RESIDE Study to examine neighborhood walkability and destination accessibility in relation to walking for transportation within a neighborhood (hereafter referred to as transport walking). A secondary aim was to examine these relationships using 3 types of logistic regression models, 2 of which utilize all available data (a population marginal model and a subject-level mixed model) and a third subject-level conditional model that exclusively uses within-person longitudinal evidence, and to compare effect estimates from these different models. The comparison of effect estimates from the subject-level mixed and conditional models enables assessment of bias arising from unmeasured time-constant confounders, such as self-selection factors.

## METHODS

### Sample and data collection

The RESIDE Study commenced in 2003 and is a longitudinal natural experiment including 1,813 people who were building homes in 73 new housing developments across metropolitan Perth, Australia. Details of participant recruitment procedures have been reported elsewhere (26). Briefly, participants who were moving to each development were invited to participate by the state water authority after the land transfer transaction. The following eligibility criteria were applied: English language proficiency, age of 18 years or older, intention to relocate by December 2005, and willingness to complete surveys 4 times over 7 years. Participants were recruited by telephone, and 1 person from each household was randomly selected. Participants were surveyed 4 times: at baseline (n = 1,813) and then approximately 1 (n = 1,467), 3 (n = 1,230), and 7 (n = 565) years later. The numbers included in analyses are slightly lower than these figures because we were missing data for some participants. Almost all participants (99%) moved to their new home between baseline and the 1-year follow-up, and 10% moved after the 1-year follow-up. The University of Western Australia's Human Research Ethics Committee (#RA/4/1/479) provided ethics approval.

### Measures

#### Outcome variable

Participants reported the frequency with which they engaged in transport-related walking within their neighborhoods (defined as a 15-minute walk from their home) over a usual week using the Neighborhood Physical Activity Questionnaire (27). This questionnaire has been shown to have acceptable reliability (27). The frequency of neighborhood transport walking was reduced to a binary yes/no variable for the main analyses, as the majority of participants did no neighborhood transport walking and the distribution was heavily skewed (68% did none, the mean frequency was 1.2 times/week, and the median was 0) and dominated by the none/some dichotomy. Thus, the focus was on factors that influenced whether or not participants did any neighborhood transport walking rather than the amount of walking (for those that did some), and the results therefore directly facilitated the study of built environment interventions aimed at increasing neighborhood transport walking specifically through encouraging the uptake or maintenance of the behavior.

All models were adjusted for age, sex, marital status, educational level, occupation (including whether or not the participant was in the workforce), hours of work per week, annual household income, the number of adults in the household, whether there were children who lived in the home, and whether the participant had access to a motor vehicle (see Table 1). These were included as time-varying factors in all models.

Table 1.

Baseline Sociodemographic Characteristics of the Cohort (n = 1,703), Residential Environments Study, Perth, Australia, 2003–2005

Characteristic
Female sex 59.8
Age, yearsa 39.9 (11.8)
Marital status
Married/de facto married 81.6
Separated/divorced/widowed 7.6
Single 10.8
Educational level
Secondary school or less 39.4
Bachelor's degree or higher 23.2
Occupational level
Professional 28.0
Blue collar 16.5
Clerical/sales/service/other 23.0
Not in workforce 17.4
Work per week, hours
≤19 10.8
20–38 24.4
39–59 42.6
≥60 4.7
Not in workforce 17.4
Annual household income
≤$50,000 25.7$50,000–$69,000 25.0$70,000–$89,000 23.2 ≥$90,000 26.0
1 16.4
2 59.4
3 12.7
≥4 11.5
Children at home 49.0
Characteristic
Female sex 59.8
Age, yearsa 39.9 (11.8)
Marital status
Married/de facto married 81.6
Separated/divorced/widowed 7.6
Single 10.8
Educational level
Secondary school or less 39.4
Bachelor's degree or higher 23.2
Occupational level
Professional 28.0
Blue collar 16.5
Clerical/sales/service/other 23.0
Not in workforce 17.4
Work per week, hours
≤19 10.8
20–38 24.4
39–59 42.6
≥60 4.7
Not in workforce 17.4
Annual household income
≤$50,000 25.7$50,000–$69,000 25.0$70,000–$89,000 23.2 ≥$90,000 26.0
1 16.4
2 59.4
3 12.7
≥4 11.5
Children at home 49.0

a Value is expressed as mean (standard deviation).

#### Objective measures of the neighborhood environment

At each time point, objective built environment measures were generated using geographic information systems. These measures included (standardized) neighborhood walkability measures, such as street connectivity, residential density, and land-use mix, that were calculated for the areas accessible along the street network within 1,600 m from the participants' homes (28, 29). The measure of land-use mix used was “model 2” in the study by Christian et al. (30). The numbers of different types of services (dry cleaners, post offices, pharmacies, and video stores; range, 0–4), convenience stores (delis, general stores, supermarkets, green grocers, seafood shops, gas stations, other food shops, and shopping centers; range, 0–8), public open space destinations (parks, sports fields, and beaches; range, 0–3), and bus stops, as well as the presence of railway stations (yes vs. no), within a 1,600-m service area around participants' homes were also calculated. A commercial electronic database of services and stores (Sensis Pty Ltd., Sydney, Australia) was used to generate these counts based on the geocoded locations of destinations and participants' homes. There is moderate to good agreement between these measures and what is actually on the ground (31).

#### Perceptions of the neighborhood environment

The number of different types of services (dry cleaners, post offices, pharmacies, and video stores; range, 0–4), convenience stores (local shops, supermarkets, green grocers, and gas stations; range, 0–4), and public open space destinations (parks, sports fields, and beaches; range, 0–3) to which the participants perceived they had access within the neighborhood (≤15-minute walk from home), as well as the presence of a bus stop (yes vs. no) or railway station (yes vs. no), were calculated from responses to the Neighborhood Environment and Walking Scale questionnaire. This questionnaire has been demonstrated to be reliable (32).

### Statistical analysis

We used logistic regression models because they are the most commonly used regression method for binary (dichotomous) outcome data and because there are well-established and reliable approaches for fitting the 3 different types of logistic regression models. All models were fitted using SAS software, version 9.4 (SAS Institute, Inc., Cary, North Carolina). The first type of logistic model was a marginal model that was fitted using all available data from the 4 time points for each person. It provided population-average estimates of the associations of factors with neighborhood transport walking (yes vs. no) using generalized estimating equations with robust standard errors. This allowed for repeated measures on the same person over time to be correlated. SAS PROC GENMOD with a repeated statement was used to fit these marginal models. The second model that also used all available data was a conditional model that included a random effect for each subject, which allowed a correlation between repeated measures on the same subject over time. This type of model provided subject-level estimates of the associations of the factors. SAS PROC GLIMMIX with a random statement was used to fit these conditional models. This second type of model is often termed a logistic mixed model because it measures both fixed and random effects. If the factor is time-varying, the estimated effect of the factor from these first 2 types of model is a combined cross-sectional and within-person change-effect estimate (22, 33). The third model was a conditional model that included a fixed effect for each subject (which absorbed the effects of all measured and unmeasured time-constant factors) and provided subject-level estimates of the effects of time-varying factors. SAS PROC LOGISTIC with a strata statement was used to fit this model (commonly termed conditional logistic regression). The effects of time-constant factors cannot be estimated in this third type of model, and because the estimated effect of a time-varying factor is based entirely on within-person changes, there is no potential for bias from measured or unmeasured time-constant confounders (22, 33). Because models with separate counts of destination types for services, convenience stores, and public open spaces did not provide a better fit than did the models with the total number of these types of destinations, only fitted models with the total counts of types of destinations are shown (for both objective and perceived destinations).

## RESULTS

### Baseline sociodemographic characteristics of study cohort

At baseline, the mean age of participants was 40 years; 60% were female, 82% were either married or in a de facto relationship (i.e., living with a partner but not married), 23% had a bachelor's degree or higher, 43% had professional or managerial-administrator occupations, and 25% lived in households that earned a total of \$90,000 per year or more. Many participants worked 39–59 hours per week (43%), had access to a motor vehicle (98%), and lived in a household that had 2 adults (59%) and children (49%) living at home (Table 1).

### Transport walking and characteristics of the neighborhood environment over time

At baseline, 37% of participants did some neighborhood transport walking, with the rates changing to 28% after 1 year, 29% after 3 years, and 36% after 7 years. The mean frequency of transport walking in the neighborhood showed a similar pattern of change over time; the mean of 1.4 trips at baseline decreased to 1.1 trips at 1 and 3 years of follow-up but then returned to baseline levels at 7 years of follow-up.

For many participants, the number of features in the neighborhood environment to which they had access decreased substantially upon relocation to a new housing development (between baseline and the 1-year follow-up) and then gradually increased, but even by the 7-year follow-up, many had not yet returned to their baseline levels (e.g., number of bus stops and railway stations and types of services and convenience stores) (Table 2). Access to types of public open spaces was similar across all time points. Street connectivity increased slightly over time, land-use mix decreased between baseline and the 1-year follow-up and then remained stable, and residential density was relatively stable over all time points.

Table 2.

Objective and Perceived Neighborhood Environment of Participants at Baseline and at 1, 3, and 7 Years of Follow-up in Their New Neighborhoods, Residential Environments Study, Perth, Australia, 2003–2012

Neighborhood Environment Measure Baseline (n = 1,703)

Year 1 (n = 1,273)

Year 3 (n = 1,150)

Year 7 (n = 504)

Mean (SD) Mean (SD) Mean (SD) Mean (SD)
Connectivity z score −0.0 (1.0)  0.7 (1.5)  1.0 (1.4)  1.2 (1.5)
Residential density z score −0.0 (1.0)  −0.3 (0.7)  −0.1 (0.5)  −0.1 (0.5)
Land-use mix z score −0.0 (1.0)  −0.5 (0.9)  −0.7 (0.7)  −0.5 (1.0)
Number of bus stops within 1,600 m 37 (21)  21 (14)  23 (15)  24 (13)
0–14  10  37  35  21
15–29  32  43  43  53
≥30  58  20  22  27
Railway station present within 1,600 m  10
No. of types of servicesa 1.3 (1.2)  0.7 (1.1)  0.8 (1.2)  0.9 (1.1)
No. of types of convenience storesb 2.2 (1.9)  0.7 (1.3)  0.8 (1.4)  1.1 (1.5)
No. of types of public open spacesc 1.2 (0.5)  1.1 (0.4)  1.1 (0.4)  1.1 (0.5)
Total no.d 4.7 (2.9)  2.4 (2.4)  2.6 (2.5)  3.1 (2.6)
0–3  44  79  79  68
4–7  35  14  13  23
8–15  21
No. of types of services to which subjects perceived they had accessg 1.6 (1.6)  0.7 (1.2)  0.9 (1.4)  1.1 (1.3)
No. of types of convenience stores to which subjects perceived they had accessh 2.0 (1.5)  0.9 (1.3)  1.2 (1.5)  1.7 (1.5)
No. of types of public open spaces to which subjects perceived they had accessi 1.6 (0.7)  1.5 (0.7)  1.5 (0.7)  1.6 (0.7)
Total no.j 5.2 (3.1)  3.1 (2.7)  3.7 (3.0)  4.4 (3.0)
0–2  28  62  53  39
3–6  34  24  25  34
7–11  38  15  22  27
Neighborhood Environment Measure Baseline (n = 1,703)

Year 1 (n = 1,273)

Year 3 (n = 1,150)

Year 7 (n = 504)

Mean (SD) Mean (SD) Mean (SD) Mean (SD)
Connectivity z score −0.0 (1.0)  0.7 (1.5)  1.0 (1.4)  1.2 (1.5)
Residential density z score −0.0 (1.0)  −0.3 (0.7)  −0.1 (0.5)  −0.1 (0.5)
Land-use mix z score −0.0 (1.0)  −0.5 (0.9)  −0.7 (0.7)  −0.5 (1.0)
Number of bus stops within 1,600 m 37 (21)  21 (14)  23 (15)  24 (13)
0–14  10  37  35  21
15–29  32  43  43  53
≥30  58  20  22  27
Railway station present within 1,600 m  10
No. of types of servicesa 1.3 (1.2)  0.7 (1.1)  0.8 (1.2)  0.9 (1.1)
No. of types of convenience storesb 2.2 (1.9)  0.7 (1.3)  0.8 (1.4)  1.1 (1.5)
No. of types of public open spacesc 1.2 (0.5)  1.1 (0.4)  1.1 (0.4)  1.1 (0.5)
Total no.d 4.7 (2.9)  2.4 (2.4)  2.6 (2.5)  3.1 (2.6)
0–3  44  79  79  68
4–7  35  14  13  23
8–15  21
No. of types of services to which subjects perceived they had accessg 1.6 (1.6)  0.7 (1.2)  0.9 (1.4)  1.1 (1.3)
No. of types of convenience stores to which subjects perceived they had accessh 2.0 (1.5)  0.9 (1.3)  1.2 (1.5)  1.7 (1.5)
No. of types of public open spaces to which subjects perceived they had accessi 1.6 (0.7)  1.5 (0.7)  1.5 (0.7)  1.6 (0.7)
Total no.j 5.2 (3.1)  3.1 (2.7)  3.7 (3.0)  4.4 (3.0)
0–2  28  62  53  39
3–6  34  24  25  34
7–11  38  15  22  27

Abbreviation: SD, standard deviation.

a Services included dry cleaners, post offices, pharmacies, and video stores within 1,600 m. The range was 0–4.

b Convenience stores included delis, general stores, supermarkets, green grocers, seafood shops, gas stations, other food shops, and shopping centers within 1,600 m. The range was 0–8.

c Public open spaces included parks measuring 2 hectares or more, sports fields, and beaches within 1,600 m. The range was 0–3.

d Total number of services, convenience stores, and public open spaces within 1,600 m. The range was 0–15.

e Perceived access to bus stop within a 15-minute walk from home.

f Perceived access to railway station within a 15-minute walk from home.

g Number of services (dry cleaners, post offices, pharmacies, and video stores) to which the participant perceived he or she had access within a 15-minute walk from home. The range was 0–4.

h Number of convenience stores (local shops, green grocers, gas stations, and supermarkets) to which the participant perceived he or she had access within a 15-minute walk from home. The range was 0–4.

i Number of public open spaces (parks, sports fields, and beaches) to which the participant perceived he or she had access within a 15-minute walk from home. The range was 0–3.

j Total number of types of services, convenience stores, and public open spaces to which the participant perceived he or she had access within a 15-minute walk from home. The range was 0–11.

### Effect of the built environment on neighborhood transport walking over time

Table 3 shows the multivariate model estimates obtained from the 3 different modeling approaches that used objective measures of neighborhood walkability and objective measures of neighborhood public transit and destinations. Table 4 shows estimates from the equivalent model with objective measures of neighborhood walkability and perceived measures of neighborhood public transit and destinations.

Table 3.

Multivariate Models of the Associations of Neighborhood Walkability and of Objective Public Transit and Destination Measures of the Built Environment With Neighborhood Transport Walking Over Time, Residential Environments Study, Perth, Australia, 2003–2012

Neighborhood Environment Measure Analysis Methoda

1b

2c

3d

OR 95% CI OR 95% CI OR 95% CI
Connectivity z score 1.09 1.03, 1.15 1.12 1.04, 1.21 1.13 1.01, 1.26
Residential density z score 1.02 0.92, 1.14 1.03 0.90, 1.18 0.96 0.80, 1.15
Land-use mix z score 1.21 1.12, 1.30 1.29 1.17, 1.43 1.33 1.16, 1.52
No. of bus stops within 1,600 m
0–14 1.00 Referent 1.00 Referent 1.00 Referent
15–29 1.63 1.34, 1.98 1.88 1.49, 2.39 1.99 1.46, 2.71
≥30 1.75 1.39, 2.19 2.07 1.56, 2.74 2.33 1.57, 3.45
Railway station present within 1,600 m 1.34 1.00, 1.81 1.52 1.03, 2.26 1.79 1.02, 3.16
Total no. of types of destinationse
0–3 1.00 Referent 1.00 Referent 1.00 Referent
4–7 1.03 0.87, 1.22 1.03 0.82, 1.29 1.08 0.80, 1.45
8–15 1.29 1.02, 1.64 1.38 1.02, 1.87 1.40 0.93, 2.10
Neighborhood Environment Measure Analysis Methoda

1b

2c

3d

OR 95% CI OR 95% CI OR 95% CI
Connectivity z score 1.09 1.03, 1.15 1.12 1.04, 1.21 1.13 1.01, 1.26
Residential density z score 1.02 0.92, 1.14 1.03 0.90, 1.18 0.96 0.80, 1.15
Land-use mix z score 1.21 1.12, 1.30 1.29 1.17, 1.43 1.33 1.16, 1.52
No. of bus stops within 1,600 m
0–14 1.00 Referent 1.00 Referent 1.00 Referent
15–29 1.63 1.34, 1.98 1.88 1.49, 2.39 1.99 1.46, 2.71
≥30 1.75 1.39, 2.19 2.07 1.56, 2.74 2.33 1.57, 3.45
Railway station present within 1,600 m 1.34 1.00, 1.81 1.52 1.03, 2.26 1.79 1.02, 3.16
Total no. of types of destinationse
0–3 1.00 Referent 1.00 Referent 1.00 Referent
4–7 1.03 0.87, 1.22 1.03 0.82, 1.29 1.08 0.80, 1.45
8–15 1.29 1.02, 1.64 1.38 1.02, 1.87 1.40 0.93, 2.10

Abbreviations: CI, confidence interval; OR, odds ratio.

a All models were adjusted for age, sex, marital status, educational level, occupation (including whether or not in the workforce), hours of work per week, household income, number of adults in the household, children at home, motor vehicle access, and time.

b Marginal population-average model (n = 1,703 participants with 4,630 observations).

c Conditional subject-level mixed model (n = 1,703 participants with 4,630 observations).

d Conditional subject-level fixed-effect model (n = 1,703 participants with 2,858 within-person comparisons).

e Total number of services, convenience stores, and public open spaces within 1,600 m. The range was 0–15.

Table 4.

Multivariate Models of the Associations of Neighborhood Walkability and of Perceived Public Transit and Destinations Measures of the Built Environment on Neighborhood Transport Walking Over Time, Residential Environments Study, Perth, Australia, 2003–2012

Neighborhood Environment Measure Analysis Methoda

1b

2c

3d

OR 95% CI OR 95% CI OR 95% CI
Connectivity z score 1.05 0.99, 1.11 1.06 0.99, 1.15 1.07 0.95, 1.19
Residential density z score 1.04 0.94, 1.15 1.05 0.92, 1.19 0.97 0.81, 1.15
Land-use mix z score 1.16 1.08, 1.25 1.22 1.10, 1.34 1.27 1.11, 1.45
Perceived access to bus stopse 1.35 1.10, 1.66 1.49 1.14, 1.96 1.31 0.92, 1.87
Perceived access to railway stationsf 1.44 1.13, 1.85 1.64 1.17, 2.28 1.80 1.13, 2.85
Total no. of types of destinations to which subjects perceived they had accessg
0–2 1.00 Referent 1.00 Referent 1.00 Referent
3–6 2.07 1.76, 2.43 2.58 2.09, 3.20 2.35 1.81, 3.05
7–11 2.32 1.95, 2.77 3.01 2.38, 3.80 3.11 2.28, 4.25
Neighborhood Environment Measure Analysis Methoda

1b

2c

3d

OR 95% CI OR 95% CI OR 95% CI
Connectivity z score 1.05 0.99, 1.11 1.06 0.99, 1.15 1.07 0.95, 1.19
Residential density z score 1.04 0.94, 1.15 1.05 0.92, 1.19 0.97 0.81, 1.15
Land-use mix z score 1.16 1.08, 1.25 1.22 1.10, 1.34 1.27 1.11, 1.45
Perceived access to bus stopse 1.35 1.10, 1.66 1.49 1.14, 1.96 1.31 0.92, 1.87
Perceived access to railway stationsf 1.44 1.13, 1.85 1.64 1.17, 2.28 1.80 1.13, 2.85
Total no. of types of destinations to which subjects perceived they had accessg
0–2 1.00 Referent 1.00 Referent 1.00 Referent
3–6 2.07 1.76, 2.43 2.58 2.09, 3.20 2.35 1.81, 3.05
7–11 2.32 1.95, 2.77 3.01 2.38, 3.80 3.11 2.28, 4.25

Abbreviations: CI, confidence interval; OR, odds ratio.

a All models were adjusted for age, sex, marital status, educational level, occupation (including whether or not in the workforce), hours of work per week, household income, number of adults in the household, children at home, motor vehicle access, and time.

b Marginal population-average model (n = 1,703 participants with 4,630 observations).

c Conditional subject-level mixed model (n = 1,703 participants with 4,630 observations).

d Conditional subject-level fixed-effect model (n = 1,703 participants with 2,858 within-person comparisons).

e Perceived access to a bus stop within a 15-minute walk from home.

f Perceived access to a railway station within a 15-minute walk from home.

g Total number of types of services, convenience stores, and public open spaces to which the participant perceived he or she had access within a 15-minute walk from home. The range was 0–11.

Both the connectivity and land-use mix walkability components (but not residential density), as well as neighborhood access to public transit, were significantly related to transport walking in the neighborhood, and this was consistent across all 3 analysis methods. Connectivity had an estimated population-average odds ratio of 1.09 per unit change (method 1) and estimated subject-level odds ratios of 1.12 and 1.13 for methods 2 and 3, respectively. Land-use mix had an estimated population-average odds ratio per unit change of 1.21 (method 1) and estimated subject-level odds ratios of 1.29 and 1.33 for methods 2 and 3, respectively. Participants who had 30 or more bus stops within 1,600 m of their homes had odds of walking for transportation that were approximately double those of participants who had 0–14 bus stops, and the presence of a train station within 1,600 m increased the odds of walking for transportation by approximately 50%. The objectively measured number of types of destinations was significantly related to walking for transportation in a dose-response manner (P for trend < 0.05); however, when categorized into levels, the odds ratio comparing neighborhoods with 8–15 destinations types within 1,600 m to those with 0–3 was only approximately 1.3 (P = 0.04).

When the objective public transit and destinations measures (Table 3) were replaced with the measures of perceived access to public transit and the variety of destinations (Table 4), the estimated odds ratios for the walkability measures were slightly attenuated, the odds ratio for railway station access was essentially the same, and although the perceived bus stop access measure was binary (yes vs. no), it remained significant in methods 1 and 2. There was, however, a substantial increase in the strength and significance of the relationship with the total count of types of destinations when the objective measure was replaced with the perceived measure. The odds of walking for transportation for participants who perceived access to 7–11 destinations types within a 15-minute walk from home was more than double that of participants who perceived access to 0–2 destination types.

## DISCUSSION

In our fitted models of population-average and subject-level (within and between persons) associations on transport walking, we found that neighborhood connectivity and land-use mix, the number of public transit stops, and the diversity of destination types that were accessible from home by walking were significantly related to neighborhood transport walking. These longitudinal results were evident in 3 types of regression models and provide stronger evidence than was available from previous findings from cross-sectional studies (11, 14, 34).

Our longitudinal results on measures of walkability support those from several cross-sectional studies (3537), as well as studies of self-selection, neighborhood design, and walking (11, 3840). However, whereas we found that both connectivity and land-use mix were positively related to transport walking, a small longitudinal study of neighborhood design and women's walking measured by pedometers found that a greater land-use mix was associated with less walking (i.e., a negative relationship) but that having fewer cul de sacs (higher connectivity) was positively associated with more walking (41). Our results suggest that the 3 traditional components of walkability (street connectivity, residential density, and land-use mix) are not always equally important determinants of transport walking and perhaps should not always be weighted equally when constructing overall walkability indices (42). In the present study, land-use mix had a greater and more significant relationship than did either street connectivity or residential density. In the few longitudinal studies conducted to date, there has been some evidence to suggest that a greater mix of businesses (25, 43) and more public and pay recreational facilities are determinants of walking and physical activity level (23). The relative importance of the 3 components of walkability, however, may differ across studies and geographic location, as the magnitude of effects per each standard-deviation change (i.e., per unit of z score) is influenced by the degree of heterogeneity in these component measures within the population under study. Perth is a relatively low-density city with little variation in these measures; in the multivariate models, the estimated odds ratio for land-use mix was modest, at approximately 1.25 for a 1-standard-deviation change. Interactions between these variables could not be adequately investigated in our study because of the relatively small variation in these measures and the moderate sample size, but they remain a possibility. For example, as noted in a review report (6), high residential density is likely to be a prerequisite for the presence of mixed land uses because without increased density, there are insufficient people to support local shops, services, and public transit. Newman and Kenworthy (44) observed that both housing and employment densities of at least 35 per hectare are required to support public transit. Thus, future research should examine the relative importance of and the interactions between the 3 components of neighborhood walkability in cities with greater variation in these measures.

As expected, our longitudinal results confirm that access to public transit (buses or railways) is a key determinant of walking for transportation. As the RESIDE Study involves a cohort of people who relocated (between baseline and the first follow-up) to new greenfield land developments, mostly on the outer regions of the Perth metropolitan areas, there was a decrease in the number of neighborhood public transit stops from baseline to the 1-year follow-up. This has previously been shown to be related to a reduction in local transport-related walking (18). The current results, which use data from baseline through the 7-year follow-up and a variety of regression modeling approaches, re-confirm this relationship. Results from 2 transit rail quasi-experiments (4547) showed that the implementation of rail infrastructure may increase walking. Our results support this because although only a small percentage of participants had local access to a train station, access was (independently) related to walking for transportation, with an odds ratio of approximately 1.5.

Our longitudinal results also confirm that local access to a variety of destinations is an important determinant of walking for transportation. However, the relationship was much stronger for the measure of perceived access to types of destinations than for the objective geographic information systems–derived measure. Further, when both the perceived and objective measures were included together in the same multivariate models (results not shown), the perceived measure remained highly significant and the objective measure was no longer significant. This may be because perceptions rather than objective measures are more proximal to the individual's intention, self-efficacy, and decision-making in relation to walking for transportation. Alternatively, the weaker relationship with objective measures could reflect measurement error and/or undercounting of the number of types of destinations using the destination databases available for this analysis (31). Nevertheless, without actual destinations being present, it would not seem possible for participants to hold positive perceptions about their presence, and so the presence of accessible destinations remains important.

Longitudinal studies, which include repeated measures of a cohort of individuals, allow both cross-sectional (i.e., between people) and within-person comparisons. Residential preferences (i.e., self-selection factors) are important in ascertaining the potential causal effect of the neighborhood environment on transport walking. Thus, direct comparisons of subject-level effect estimates from logistic mixed models (method 2) and conditional logistic models (method 3) enable an assessment of the influence of uncontrolled (time-constant) self-selection and other confounders. If the effect estimates are similar, then there can be no bias from uncontrolled confounding by time-constant factors in the logistic mixed model (method 2); therefore, its effect estimate is preferred, as it uses all available cross-sectional and longitudinal evidence. Furthermore, the effect estimate from the corresponding logistic generalized estimating equations model (method 1) is thus also preferred, although of course it provides population-average estimates. If the logistic mixed model (method 2) and conditional logistic model (method 3) produce different effect estimates, then the presumption is that there is confounding by time-constant factors and therefore estimates from the conditional logistic model (method 3), which are not affected by this confounding, are preferred. Note, however, that because the conditional logistic model only uses within-person (i.e., longitudinal) comparisons, its estimate will be less precise (i.e., have a greater standard error).

In our multivariate models, we found similar effect sizes from methods 2 and 3 for all neighborhood environment factors, which suggests that there is little confounding by uncontrolled time-constant factors, including time-constant selection factors. As expected, we also observed more precise effect estimates (i.e., the confidence intervals were not as wide) from method 2 than from method 3. Further, as is generally the case, the effect sizes were slightly larger in the subject-level logistic mixed model (method 2) than in the equivalent marginal model (method 1). That is, the effect of these environmental factors is greater when considered at the individual level than when considered at a population-average level. The estimated subject-level odds ratio represents the estimated increase in the odds of an individual doing transport walking if his or her environmental factor level changes. It is important to note that logistic models applied to cross-sectional data produce population-average effect estimates and hence can only be validly compared with effect estimates from method 1 for longitudinal data.

Note that although comparison of the estimates from the models in methods 2 and 3 does allow assessment of potential confounding by unmeasured time-constant factors, it does not address potential confounding by unmeasured time-varying factors. Uncontrolled time-varying factors that are associated with changes in walking behavior and in the built environment may lead to bias in the estimated association with the built environment. Changing health status and demographic transitions such as marriage, divorce, having children, and leaving the workforce may be associated with changes in walking behavior and are also likely to be associated with changes in built environment if the participant relocates because of the change in health or demographic status. In our models, we adjusted for demographic transitions but not for changing health status, so there is potential for bias from changing health status.

A limitation and further potential source of bias in effect estimates from the present study is the drop-out of participants (especially between the 3-year and 7-year follow-ups). An analysis of the factors related to participant drop-out revealed that drop-out was associated with some demographic variables (age, sex, having children at home) but was not related to prior values of walking behavior (the outcome variable). When drop-out is related to covariates only and not to prior or missing values of the outcome variable, the drop-out pattern is called (conditionally on the covariates) missing completely at random. Estimates from method 1 are unbiased under this pattern of drop-out provided the covariates related to drop-out are included in the models and that there are no further unmeasured covariates related to drop-out, and estimates from the likelihood-based methods 2 and 3 are unbiased under this drop-out pattern and also under the less restrictive pattern of missing at random where drop-out can also be related to the prior value of the outcome variable (22). If drop-out is related to the missing values of the outcome variable, which cannot be checked from the data, the drop-out pattern is called non-ignorable and all methods may produce biases estimates.

In a recent Australian study, physical inactivity ranked fourth in terms of measured population health loss (48), and because this is largely amenable to change, there is a need to identify effective built environment features that encourage increased physical activity level. The present longitudinal study provides stronger evidence than cross-sectional studies of a possible supportive effect of the neighborhood built environment on local transport-related walking. The results suggest that people living in neighborhoods with mixed land uses, connected streets, access to public transit, and a diverse number of destinations (services, convenience stores, and public open spaces) benefit from higher levels of local transport walking. These findings provide policy makers and urban planners with further evidence of features of the built environment that may be important in the design of neighborhoods to increase transport walking and meet the health needs of residents. Future research should examine the relative importance of the components of neighborhood walkability in cities with greater variation in these measures. They should also consider using the analytic methods described here when examining the relationships between the built environment and walking.

## ACKNOWLEDGMENTS

Author affiliations: Centre for the Built Environment and Health, School of Population Health, The University of Western Australia, Perth, Australia (Matthew W. Knuiman, Hayley E. Christian, Mark L. Divitini, Sarah A. Foster, Fiona C. Bull); and McCaughey VicHealth Centre for Community Wellbeing, Melbourne School of Population and Global Health, University of Melbourne, Melbourne, Australia (Hannah M. Badland, Billie Giles-Corti).

The Residential Environments Study was funded by grants from the Western Australian Health Promotion Foundation (grant 11828), the Australian Research Council (grant LP0455453), and an Australian National Health and Medical Research Council Capacity Building Grant (458688). B.G.-C. was supported by a National Health and Medical Research Council Principal Research Fellow Award (grant 1004900). H.E.C. is supported by a National Health and Medical Research Council/National Heart Foundation Early Career Fellowship (grant 1036350), and S.A.F. is supported by a Healthway Health Promotion Research Fellowship (grant 21363).

We thank the geographic information systems team (Nick Middleton, Sharyn Hickey, Bridget Beasley, and Dr. Bryan Boruff) for their assistance in the development of the geographic information systems measures used in this study.

Conflict of interest: none declared.

## REFERENCES

1
Corburn
J
Special section on urban planning and public health
J Am Plann Assoc

2006
72
1
119
121
2
Humphreys
DK
Ogilvie
D
Social inequalities in physical activity: do environmental and policy interventions help to reduce the gap? A pilot systematic review
Lancet

2012
380
S50
3
Renalds
A
Smith
TH
Hale
PJ
A systematic review of built environment and health
Fam Community Health

2010
33
1
68
78
4
Barton
H
Land use planning and health and well-being
Land Use Policy

2009
26
suppl 1
S115
S123
5
Kopelman
P
Jebb
SA
Butland
B
Executive summary: foresight ‘tackling obesities: future choices’ project
Obes Rev

2007
8
suppl 1
vi
ix
6
National Research Council, Committee on Physical Activity Land Use Institute of Medicine
Does the Built Environment Influence Physical Activity? Examining the Evidence

2005
Washington, DC
Transportation Research Board National Research
7
H
Schofield
G
Transport, urban design, and physical activity: an evidence-based update
Transport Res D-Tr E

2005
10
3
177
196
8
Giles-Corti
B
Timperio
A
Bull
F
et al.
Understanding physical activity environmental correlates: increased specificity for ecological models
Exerc Sport Sci Rev

2005
33
4
175
181
9
Durand
CP
Andalib
M
Dunton
GF
et al.
A systematic review of built environment factors related to physical activity and obesity risk: implications for smart growth urban planning
Obes Rev

2011
12
5
e173
e182
10
Lee
C
Moudon
AV
Physical activity and environment research in the health field: implications for urban and transportation planning practice and research
J Plan Lit

2004
19
2
147
181
11
McCormack
GR
Shiell
A
In search of causality: a systematic review of the relationship between the built environment and physical activity among adults
Int J Behav Nutr Phys Act

2011
8
125
136
12
Saelens
B
Sallis
JF
Frank
LD
Environmental correlates of walking and cycling: findings from the transportation, urban design, and planning literatures
Ann Behav Med

2003
25
2
80
91
13
Ferdinand
AO
Sen
B
Rahurkar
S
et al.
The relationship between built environments and physical activity: a systematic review
Am J Public Health

2012
102
10
e7
e13
14
Sugiyama
T
Neuhaus
M
Cole
R
et al.
Destination and route attributes associated with adults’ walking: a review
Med Sci Sports Exerc

2012
44
7
1275
1286
15
US Department of Health and Human Services
Physical Activity and Health: A Report of the Surgeon General

1996
Atlanta, GA
US Department of Health and Human Services, Centers for Disease Control and Prevention, National Center for Chronic Disease Prevention and Health Promotion
16
World Health Organization
Interventions on Diet and Physical Activity: What Works: Summary Report

2009
Geneva, Switzerland
World Health Organization
17
Giles-Corti
B
Bull
F
Knuiman
M
et al.
The influence of urban design on neighbourhood walking following residential relocation: longitudinal results from the RESIDE Study
Soc Sci Med

2013
77
20
30
18
Christian
H
Knuiman
M
Bull
F
et al.
A new urban planning code's impact on walking: the residential environments project
Am J Public Health

2013
103
7
1219
1228
19
Ding
D
Gebel
K
Built environment, physical activity, and obesity: what have we learned from reviewing the literature?
Health Place

2012
18
1
100
105
20
Larouche
R
Trudeau
F
Étude des impacts du transport actif sur la pratique d'activités physiques et la santé et de ses principaux determinants
Sci Sport

2010
25
5
227
237
21
Mokhtarian
PL
Cao
X
Examining the impacts of residential self-selection on travel behavior: a focus on methodologies
Transport Res B-Meth

2008
42
3
204
228
22
Fitzmaurice
G
Laird
N
Ware
J
Applied Longitudinal Analysis

2011
2nd ed
Hoboken, NJ
John Wiley & Sons
23
Boone-Heinonen
J
Guilkey
DK
Evenson
KR
et al.
Residential self-selection bias in the estimation of built environment effects on physical activity between adolescence and young adulthood
Int J Behav Nutr Phys Act

2010
7
70
24
Cao
X
Mokhtarian
PL
Handy
SL
Examining the impacts of residential self-selection on travel behaviour: a focus on empirical findings
Transport Rev

2009
29
3
359
395
25
Handy
S
Cao
X
Mokhtarian
PL
Self-selection in the relationship between the built environment and walking: empirical evidence from Northern California
J Am Plann Assoc

2006
72
1
55
74
26
Giles-Corti
B
Knuiman
M
Timperio
A
et al.
Evaluation of the implementation of a state government community design policy aimed at increasing local walking: design issues and baseline results from RESIDE, Perth Western Australia
Prev Med

2008
46
1
46
54
27
Giles-Corti
B
Timperio
A
Cutt
H
et al.
Development of a reliable measure of walking within and outside the local neighborhood: RESIDE‘s Neighborhood Physical Activity Questionnaire
Prev Med

2006
42
6
455
459
28
Frank
LD
Schmid
TL
Sallis
JF
et al.
Linking objectively measured physical activity with objectively measured urban form: findings from SMARTRAQ
Am J Prev Med

2005
28
2 Suppl 2
117
125
29
Villanueva
K
Pereira
G
Knuiman
M
et al.
The impact of the built environment on health across the life course: design of a cross-sectional data linkage study
BMJ Open

2013
3
1
e002482
30
Christian
HE
Bull
FC
Middleton
NJ
et al.
How important is the land use mix measure in understanding walking behaviour? Results from the RESIDE study
Int J Behav Nutr Phys Act

2011
8
55
31
Hooper
PL
Middleton
N
Knuiman
M
et al.
Measurement error in studies of the built environment: validating commercial data as objective measures of neighborhood destinations
J Phys Act Health

2013
10
6
792
804
32
Cerin
E
Saelens
BE
Sallis
JF
et al.
Neighborhood Environment Walkability Scale: validity and development of a short form
Med Sci Sports Exerc

2006
38
9
1682
1691
33
Allison
PD
Fixed Effects Regression Models for Longitudinal Data Using SAS

2005
Cary, NC
SAS Institute Inc
34
Saelens
BE
Handy
SL
Built environment correlates of walking: a review
Med Sci Sports Exerc

2008
40
7 suppl
S550
S566
35
Frank
LD
Saelens
BE
Powell
KE
et al.
Stepping towards causation: do built environments or neighborhood and travel preferences explain physical activity, driving, and obesity?
Soc Sci Med

2007
65
9
1898
1914
36
Owen
N
Cerin
E
Leslie
E
et al.
Neighborhood walkability and the walking behavior of Australian adults
Am J Prev Med

2007
33
5
387
395
37
Sallis
JF
Saelens
BE
Frank
LD
et al.
Neighborhood built environment and income: examining multiple health outcomes
Soc Sci Med

2009
68
7
1285
1293
38
Khattak
AJ
Rodriguez
D
Travel behavior in neo-traditional neighborhood developments: a case study in USA
Transport Res A-Pol

2005
39
6
481
500
39
Cao
XJ
Exploring causal effects of neighborhood type on walking behavior using stratification on the propensity score
Environ Plann A

2010
42
2
487
504
40
Cao
X
Mokhtarian
PL
Handy
SL
The relationship between the built environment and nonwork travel: a case study of Northern California
Transport Res A-Pol

2009
43
5
548
559
41
Wells
NM
Yang
Y
Neighborhood design and walking. A quasi-experimental longitudinal study
Am J Prev Med

2008
34
4
313
319
42
Witten
K
Blakely
T
Bagheri
N
et al.
Neighborhood built environment and transport and leisure physical activity: findings using objective exposure and outcome measures in New Zealand
Environ Health Perspect

2012
120
7
971
977
43
Cao
X
Mokhtarian
PL
Handy
SL
Do changes in neighborhood characteristics lead to changes in travel behavior? A structural equations modeling approach
Transportation

2007
34
5
535
556
44
Newman
P
Kenworthy
J
Urban design to reduce automobile dependence
Opolis

2006
2
1
35
52
45
Brown
BB
Werner
CM
A new rail stop: tracking moderate physical activity bouts and ridership
Am J Prev Med

2007
33
4
306
309
46
Brown
BB
Werner
CM
Before and after a new light rail stop: resident attitudes, travel behavior, and obesity
J Am Plann Assoc

2008
75
1
5
12
47
MacDonald
JM
Stokes
RJ
Cohen
DA
et al.
The effect of light rail transit on body mass index and physical activity
Am J Prev Med

2010
39
2
105
112
48
Begg
SJ
Vos
T
Barker
B
et al.
Burden of disease and injury in Australia in the new millennium: measuring health loss from diseases, injuries and risk factors
Med J Aust

2008
188
1
36
40

## Author notes

Abbreviation: RESIDE, Residential Environments.