The RECORD Cohort Study 35 36 includes 7290 participants recruited in 2007–2008 during 2-hour-long preventive medical checkups conducted by the Centre d’Investigations Préventives et Cliniques in 4 of its health centers in the Paris metropolitan area 37 38 39 40 41 42 . These checkups are offered for free to all working and retired employees and their families. We recruited without a priori sampling people who were attending the healthcare centers without invitation from our part (convenience sample). The eligible population for these preventive health checkups includes all the currently working, unemployed, and retired salaried workers and their families. In our study counties, this group represents 95% of the overall population 36 . However, the recruitment channels of these healthcare centers are very diverse (people’s own initiative or appointments through the employers, work physicians, social workers, various associations, etc.). The absence of randomization in the recruitment of the participants led to a sample that was not representative of the background population. A previous work showed that a high individual education, a high neighborhood socioeconomic status, and a low building density were associated with higher odds of participation in the RECORD Study 38 . All these factors were included in the models or considered for adjustment to minimize bias.

Eligibility criteria were as follows: age 30 to 79 years, ability to fill out questionnaires, and residence in one of the 10 (out of 20) administrative divisions of Paris or 111 other municipalities of the Paris Ile-de-France region (among a large number of municipalities in the region) that were selected a priori. The districts and municipalities were selected among those that provided a large number of consultants to the medical center in the years prior to the recruitment, and in an attempt to maximize municipality-level socioeconomic disparities and to cover both urban and periurban territories. Of the eligible participants, 83.6% accepted to participate and completed the data collection protocol.

Participants were geocoded based on their residential address in 2007–2008, using the geocoding tool of the French National Institute of Statistics and Economic Studies that ensured an exact correspondence between the spatial coordinates and census tract neighborhoods. Research assistants corrected all incorrect or incomplete addresses with the participants by telephone, and extensive investigations with local departments of urban planning were conducted to complete the geocoding when needed. The study protocol was approved by the French Data Protection Authority. After excluding individuals with missing values for walking (n = 185, see Additional file 1A), 7105 participants from 661 census tracts (TRIRIS areas) were included in the analyses.

The present article reports cross-sectional analyses of baseline data of the RECORD Study. Reporting recreational walking or not was modeled with a logistic regression. The recreational walking time in the neighborhood was analyzed with a logit ordinal model. Coding the neighborhood recreational walking time as a continuous variable led to non-normal residuals in a linear model even after adjustment, while coding it as a binary variable would imply a substantial loss of information.

Multilevel-spatial regression models were estimated 32 48 49 , with two distinct random effects at the census tract neighborhood level incorporated in the same model 32 : both a classical random effect to model unstructured between-neighborhood variations, and a conditional autoregressive Normal random effect to account for spatially structured patterns of variability, i.e., correlation in walking between adjacent neighborhoods. Unstructured and spatially structured neighborhood variations (i.e., two components of the between-neighborhood variance that reflect patterns of variations that respectively show and do not show spatial autocorrelation) were expressed with interquartile odds ratios 32 50 51 . Such ratios quantify on the odds ratio scale the difference in walking between the 25% of individuals in neighborhoods with the lowest prevalence of recreational walking and the 25% of individuals in neighborhoods with the highest prevalence of recreational walking. Assessing spatially structured patterns of variability in walking is useful to generate hypotheses on the underlying social and environmental processes at play 32 33 52 . We therefore derived maps of the spatially structured pattern of variations in walking at the different steps of the modeling, and calculated the proportion of the total between-neighborhood variance that was spatially structured 53 as an indication of whether such spatially autocorrelated pattern in walking accounted for a small part or substantial part of residual between-neighborhood variations.

First, models only adjusted for age and sex were estimated, to assess between-neighborhood variability in walking. Second, models retaining the individual and weather variables independently associated with each outcome were estimated. Third, on the basis of the previous models, all environmental variables were pretested one by one 39 . Fourth, the environmental variables independently associated with each outcome were progressively combined into one model for each outcome. Fifth, interactions were tested between the meteorological variables and the neighborhood variables retained in the models, in order to identify environmental conditions that are supportive of walking even under poor weather (which is relevant given the frequent poor weather conditions in our region). Models were estimated using Winbugs 1.4 54 .

To provide a sense of the overall disparities between individuals that were predicted by the environmental factors retained in the models, using the SAS software, we calculated a risk score of recreational walking as a function of the environmental predictors from the final models (sum of products of coefficients by variable values for each individual); we divided this risk score in four categories based on the quartiles; and we re-estimated the final models after replacing all the environmental variables by the risk score categories. For these categories, we report the median of the odds ratio and the 2.5^th and 97.5^th percentiles as confidence intervals over 2000 bootstrapped samples (analysis performed with SAS 9.3).

The sociodemographic characteristics of the sample are reported in Table 3. Overall, 69% of the participants reported recreational walking. The mean recreational walking time among recreational walkers was 3h31mn (interquartile range: 1h00mn, 4h00mn). Seventy-four percent of the recreational walkers accomplished at least part of this walking activity in their residential neighborhood.

After adjustment for age and gender, between-neighborhood variations were observed for the two walking outcomes (see Table 4). Between-neighborhood variations showed a relatively strong spatial structure for reporting or not recreational walking and for the recreational walking time in the neighborhood (i.e., over 60% of between-neighborhood variations were spatially structured, i.e., were attributable to a spatially autocorrelated rather than unstructured pattern of variability). The spatially structured component of between-neighborhood variations in the odds of reporting recreational walking is represented in Figure 2.

Legend: CrI, credible interval; IqOR, interquartile odds ratio.

^aThe first model only included age and sex. The second model further introduced individual sociodemographic variables and weather variables. The third model further included the environmental variables associated with each outcome (see Tables 4 and 5).

^bRegarding recreational walking or not, the IqOR of 2.07 for the total between-neighborhood variance in the age and sex model indicates that the 25% of participants living in neighborhoods with the highest odds of recreational walking had 2.07 times larger odds of recreational walking than the 25% of participants residing in neighborhoods with the lowest odds of recreational walking. In the same model, 62% of the total between neighborhood variance was attributable to the spatially structured component of neighborhood variations. When progressively adding covariates to the models, spatially structured variations decreased to a large extent but spatially unstructured variations did not, resulting in a decreasing share of the total between-neighborhood variability that was spatially structured.

Individual sociodemographic and weather variables were then introduced into the models (Tables 5 and 6). Low white collar workers and blue collar workers had lower odds to report recreational walking over 7 days. Higher average temperature and lower rainfalls were associated with a higher prevalence of recreational walking and with a higher neighborhood recreational walking time.

Legend: CrI, credible interval; OR, odds ratio.

^aAt each step of the modeling, only the additional variables tested that were independently associated with the outcome were retained in the model.

Legend: CrI, credible interval; OR, odds ratio.

^aAt each step of the modeling, only the additional variables tested that were independently associated with the outcome were retained in the model.

Regarding the first outcome (Table 5), the likelihood of reporting recreational walking was higher in neighborhoods with a medium-high or high level of education compared to low education neighborhoods. Moreover, residing in an air traffic exposure area was related to a decreased probability of recreational walking, after adjustment for all the factors listed in Table 5.

Regarding the second outcome (Table 6), a higher neighborhood education level, the presence of green/open spaces of quality, and a higher density of destinations nearby were independently associated with a higher neighborhood recreational walking time. In the opposite direction, exposure to air traffic was related to a reduced walking time in one’s residential neighborhood.

No interaction was documented between the environmental variables and the weather variables retained in the final models.

As shown in Table 4, accounting for the environmental factors led to a notable reduction in the spatially structured between-neighborhood variations in walking, especially for the neighborhood recreational walking time but also for reporting recreational walking or not (i.e., the interquartile odds ratios quantifying the spatially autocorrelated pattern of variability decreased when environmental factors were added to the models). On the opposite, the environmental factors did not explain spatially unstructured between-neighborhood variations in walking (the corresponding interquartile odds ratios did not decrease when including environmental factors into the models). Before adjustment for environmental factors (top part of Figure 2), higher odds of recreational walking were documented in Paris and in the first crown of counties around Paris, especially in the westernmost county of the first crown of counties which is the most urbanized after Paris. As shown in the bottom part of Figure 2, the residual spatially structured patterns of variations in recreational walking were substantially altered after accounting for the higher density of destinations in Paris and in the first crown of counties, for the more frequent exposure to air traffic in the western and northern counties of the second crown of counties, and for the more frequent presence of green/open spaces of quality in the western counties of the first and second crowns of counties.

When all the environmental variables in each model were replaced by a risk score, it was found that, compared to the environments with the lowest environmental supportiveness of walking, the odds of reporting recreational walking were 1.17 (95% CI: 1.13, 1.20) times, 1.55 (95% CI: 1.44, 1.62) times, and 1.59 (95% CI: 1.56, 1.62) times higher in the three categories of increasing environmental supportiveness. The odds of reporting a higher recreational walking time in one’s neighborhood were 1.20 (95% CI: 1.13, 1.27) times, 1.52 (95% CI: 1.44, 1.60) times, and 1.81 (95% CI: 1.73, 1.87) times higher in these three categories of increasing environmental supportiveness.

Strengths of the study include the multilevel-spatial modeling investigation performed, the numerous individual, environmental, and meteorological determinants considered, the determination of environmental variables in street network buffers using Python scripts that ensure replicability, and the two complementary variables analyzed to capture overall walking and walking in one’s residential neighborhood. Study limitations include its cross-sectional design and the declarative nature of walking data (a shortcoming that we address with GPS and accelerometers in the RECORD GPS and MultiSensor Studies 55 56 ). Concerns related to declarative walking data include the fact that the different population groups may have a distinct definition of recreational walking, a different recall rate of recreational walking episodes, a different perception of the social desirability of regular walking, and a different accuracy in walking times assessment. Overall, measurement error may either dilute the associations reported or systematically bias some of them, e.g., those between neighborhood education and walking. Another limitation is the lack of adjustment for neighborhood selection factors that likely confound the reported associations 10 57 58 59 . For example, a high motivation for recreational walking leading certain participants to choose to reside, when moving, in a neighborhood with a lot of green spaces would confound the relationship between the spatial accessibility to green spaces and recreational walking.

Moreover, participants had to rely on a subjective definition of their neighborhood (i.e., no instructions were provided to the participants) to report (i) the perceptions of their neighborhood (that were used to determine the neighborhood ecometric variables) and (i) the location of walking. Regarding the first aspect, it seems relevant to adhere to the participants’ definition of neighborhood to assess their psychological relationship with it. Regarding the second aspect, however, using a subjective definition of neighborhood for the assessment of walking may have biased the analyses on the recreational walking time made in one’s residential neighborhood. The identified environmental determinants of a higher recreational walking time performed in one’s neighborhood may reflect perceived neighborhoods of larger sizes for the participants living in these environments (which would lead to a larger share of their walking activity in their perceived neighborhood). While we admit a possible bias, we do not believe these relatively small variations in the average size of perceived neighborhoods to be sufficient to drastically influence the reported time of recreational walking made in one’s residential neighborhood. What matters instead is whether there is or not a park around the residence that substantially increases recreational walking nearby the residence.

Finally, in addition to selection processes in the recruitment of the RECORD Cohort (convenience sample) 38 , low educated and low income participants had higher odds to be excluded due to missing data in walking times (Additional file 1A). Adjustment of regression models on walking for education and income likely reduced selection biases. However, we could not examine whether participants and non-participants in RECORD, and whether participants excluded or not due to missing values in walking times differed in their true walking behavior.

For the two outcomes investigated, two independent associations were reported with neighborhood socioeconomic factors (with neighborhood education), three with the physical environment (one of them with green/open spaces, two with air traffic), and one with the service environment (density of destinations) (see the classification of environmental factors reported in Table 1).

Previous literature did not systematically observe associations between the socioeconomic environment and recreational walking. For example, one study documented an association between residing in a high socioeconomic status neighborhood and the practice of recreational walking even after adjustment for perceived environmental factors 60 . However, another study found that area socioeconomic status did not predict walking for recreation after adjustment for other environmental variables 61 . In our study, neighborhood socioeconomic effects were captured by education rather than income, in line with most previous research in RECORD on clinical risk factors of cardiovascular diseases 37 39 40 62 63 . A potential explanation is that a high average education of local residents may foster a general climate of values favorable to healthy lifestyles.

In relation to the physical environment, the presence and high quality of green/open spaces were associated with a higher recreational walking time in one’s residential neighborhood. Additional file 1B demonstrates that the relationship between the presence/quality of green/open spaces and recreational walking in one’s residential neighborhood was not attributable to same-source bias (i.e., to the fact that the ecometric variable was based on the aggregation of information provided by the participants who themselves reported their walking time). This finding is coherent with a recent review 11 that indicated that a positive relationship between parks/open spaces and recreational walking was reported in 44% of the cases 61 64 65 . The fact that an association was documented with the presence/quality of green/open spaces (shared perception, as aggregated through the ecometric approach) rather than with the objective surface of green spaces suggests that the quality of green/open spaces is important in addition to their availability, a conclusion on the importance of the quality of recreational facilities that also emerged from previous literature 11 .

As also related to the physical environment, living in the air traffic area around the two international Paris Region airports (as objectively assessed) was inversely associated with recreational walking in two models. No previous study investigated the association between air traffic exposure and walking. A plausible hypothesis is that noise from air traffic around international airports discourages recreational walking, a situation of environmental injustice. While exposure to air traffic was associated with lower odds of recreational walking, other environmental nuisances such as the presence of a highway, road traffic pollution, or waste treatment facilities nearby the residence were not.

A recent review of previous literature identified recreational walking to be less dependent on destinations than utilitarian walking 11 . Apart from the density of destinations, previous studies have reported exercise/recreational walking to increase with housing 66 , residential 8 , and population 20 density and with composite walkability indices 67 . In our French context, in relation to the service environment, a high density of destinations and services nearby the residence was associated with a higher recreational walking time in one’s neighborhood (effects of density on recreational walking were better captured by service density than by densities of population, buildings, intersections, and public transportation stations). A potential interpretation is that there is no perfect separation between recreational and utilitarian walking and that hybrid walking episodes exist.

Although weather condition was associated with recreational walking, none of the weather variables interacted with the identified environmental effects. It suggests that these characteristics of the socioeconomic, physical, and service environment are supportive of recreational walking regardless of the weather.

It is, however, important to keep in mind that a large number of environmental factors that were tested were not associated with recreational walking in the final models. Overall, of the 50 associations with environmental factors that were tested (25 factors × 2 outcomes), 33 associations were documented when each environmental factor was introduced separately in the model without adjustment for other environmental factors (but with adjustment for individual and weather variables); but only 6 associations were documented when environmental predictors were mutually adjusted for. In addition to the factors not associated with walking discussed above, it should be noted that no association was documented with the social-interactional environment or with the symbolic environment. All the environmental factors that were associated were either already identified in previous literature (neighborhood socioeconomic status, density of destinations, green/open spaces); or identified in the different models, thus relatively consistent (neighborhood socioeconomic status, air traffic exposure); or related to plausible hypotheses (all factors, including air traffic exposure).

The associations between environmental factors and walking were of modest magnitude and had relatively wide confidence intervals. In the final models, none of the associations identified showed an odds ratio above 1.5 (or below 0.67). It should be noted, however, that the overall influence of the environments identified in the initial models was shared between the different environmental factors when they were considered simultaneously (i.e., the strength of the relationship of each environmental factor with walking was reduced when the different factors were adjusted for each other). An indicator of the overall disparities that were predicted by the different environmental factors suggests that the odds of reporting recreational walking and the odds of a higher neighborhood recreational walking time were, respectively, 1.59 times and 1.81 times higher in the most vs. the least supportive environments (lowest and highest quartiles of cumulated effects of the different environmental factors). Such an overall quantification of environmental influences on walking may be useful to convince policymakers that, in addition to information and health promotion campaigns, interventions that address the environmental determinants of walking are relevant.

More environmental predictors and stronger associations were documented for the recreational walking time in the neighborhood than for overall recreational walking, which was expected because in the first case both the exposure and the outcome refer to the residential neighborhood. Finally, it was found that environmental factors explained to a greater extent spatially structured than spatially unstructured variations between neighborhoods. This finding empirically confirms the usefulness of spatial-multilevel regression models that are able to isolate, in the unexplained between-neighborhood variation, a part of it that is spatially structured, which is useful to generate hypotheses. Such spatially structured neighborhood variations in walking were not completely eliminated by the environmental factors retained in the models. Our future work will have to build on the map of residual spatial variations in walking reported in the bottom part of Figure 2 to generate additional explanatory environmental hypotheses.