An invitation to respond to the survey was sent by mail (including the paper-based questionnaire and a link to online questionnaire) to 6000 residents aged ≥ 18 years of two suburban postal areas in each of five Finnish cities (Helsinki, Vantaa, Kuopio, Vaasa, and Oulu), two city center postal areas in Kuopio, and six city center postal areas in Helsinki (Fig. 1). Residents were selected by simple random sampling by the Finnish Digital and Population Data Services Agency (official Finnish population registry). Each suburban postal area had a quota of 500 residents (max. one/household), and both city centers had a quota of 500 residents. The included cities are among the fifteen most populated cities in Finland, and their population in 2021 ranged from 68,000 (Vaasa) to 658,000 (Helsinki). All the suburban areas were dominated by apartment buildings and located a few kilometers from the city center. The first invitation was sent in October 2021, and after one reminder, 2072 questionnaires were returned by mid-December 2021. We excluded 15 questionnaires due to (1) duplicate responses (n = 6); (2) respondent having recently moved or living in other than the registered address (n = 6); and (3) the recipient not filling out the questionnaire themselves (n = 3). The final sample size was thus 2057 (response rate 34%).

The locations of the postal areas included in the study and their response rates

Socio-demographic factors included self-reported gender (male or female; “other” excluded due to < 30 responses, causing issues with convergence), age (in years), relationship status, having children in household, education, employment status, and annual household income. Categories for all predictors are shown in Table 1. Dwelling factors were also self-reported and included satisfaction with dwelling (on a scale 1 “very satisfied” to 5 “very dissatisfied”), type of building, ownership, living floor, and having a green view and the frequency of looking at it.

Objective environmental factors included travel-related urban zones, reflecting both population density and transportation facilities [23] (Table 1). To define area-level socioeconomic deprivation, we used a summary score based on unemployment rate, educational attainment (proportion of those aged > 18 with primary or lower secondary education at highest), and household income (mean income proportional to the number of household, treated as additive inverse), calculated in 250 m × 250 m grids [24]. For each indicator, we derived a standardized z-score and used the mean value across all z-scores [25]. Higher values indicate higher socioeconomic deprivation.

Green spaces were based on CORINE land cover 2018 data from which we identified areas with a minimum size of 1.5 ha that are suitable for recreation [26]. Blue spaces comprised of lakes [27], rivers, and seas [28]. We used a 1-km buffer zone around each respondent’s home location which reflects recreational possibilities at the neighborhood-level within walking distance [29]. The share of blue spaces was highly skewed and, thus, collapsed into categories of none (< 0.1%), some (0.1‒10%), and more than 10%.

To assess possibilities to influence neighborhood decisions, the respondents were asked to rate their possibilities to influence the decisions regarding their residential area, on a scale 1 “very good,” 2 “good,” 3 “neither good nor poor,” 4 “poor,” 5 “very poor,” and 6 “not interested.” For the analyses, we combined options 1 and 2 (“good” or “very good”), 4 and 5 (“poor” or “very poor”), while “neither good nor poor” and “not interested” remained as separate categories. The question was designed by the project researchers.

Subjective environmental factors were asked with a question “How satisfied are you with the following aspects of your residential area?,” followed by a list of fourteen attributes (Table 1). The rating options were 1 “very satisfied,” 2 “satisfied,” 3 “neither satisfied nor dissatisfied,” 4 “dissatisfied,” and 5 “very dissatisfied.” Because responses 4 and 5 were rare, and to ease interpretation and model complexity, these were all recoded into binary variables reflecting whether the respondent was satisfied with the attribute (options 1 and 2) or not (options 3–5).

Neighborhood liveability was enquired by asking “In your opinion, how comfortable is your residential area?,” with the options 1 “Very comfortable,” 2 “Comfortable,” 3 “Neither comfortable nor uncomfortable,” 4 “Uncomfortable,” and 5 “Very uncomfortable.” Comfort and liveability have partially the same translation in Finnish (viihtyisyys). Environmental comfort refers to a holistic cognitive evaluation of the effect of surrounding environmental conditions on oneself [30]. Thus, it conceptually largely overlaps with liveability. Liveability has been more typically used in urban planning literature regarding neighborhoods (vs comfort in indoor conditions), and we therefore use the term “liveability” throughout this paper [13]. For the analyses, we categorized the outcome into a binary measure indicating whether the respondent found their residential areas as liveable (options 1 and 2) or not (options 3–5).

We first assessed descriptively the explanatory factors and neighborhood liveability. For the multivariate models, explanatory variables were added in the following steps, reflecting the literature on factors associated with neighborhood evaluations: (1) socio-demographic, (2) dwelling-related, (3) objective environmental factors, and (4) subjective dwelling and environmental factors, including participation in the decision-making process of the neighborhood. The model created in the last step (4) was considered our main model. Prior to the analyses, we screened the explanatory variables for multicollinearity using the scaled generalized variance inflated factor, suitable for both continuous and categorical variables, to ensure all values were below 2.5. We based the inference on effect sizes using odds ratios (OR) and their 95% confidence intervals (CI) and p-values (0.05 as the approximate threshold for “statistical significance”).

The multivariate models were specified with R software [31], using both a fixed effects and mixed models with a random intercept for every postal area. In the last analysis step, the model without random intercept showed a slightly better fit with the data based on Akaike’s and Bayesian information criteria, and hence, this approach was selected for the main models. Moreover, the outcome was initially specified as “ordinal,” but due non-convergence with including age and household income in the models, we specified the outcome as binary and checked that the results were consistent with the ordinal models. In all analyses, the sample was weighed to match the age and gender distributions in the target postal areas.

After building the main model, we assessed moderation by factors potentially reflecting more settled status in the neighborhood by adding one interaction term at a time for the following: having children, participation in decision-making, living with a partner, age (categorized into 18–34, 35–64, and 65 + years), dwelling ownership, and employment status (with categories with few cases merged), and those subjective and objective environmental factors that had the strongest association with neighborhood liveability (OR close to 2.00 or greater/0.50 or smaller).

We checked the robustness of our main results with several alternative model specifications. First, the main model was re-estimated in the following ways: without sample weights, adding a random intercept for the postal areas, specifying the outcome as ordered categorical (with probit link), and changing the threshold in the outcome to the highest category. Second, we re-ran the analyses using only the suburban sample (n = 1392) to see how much of the relationships was affected by urbanicity (the central urban sample was too small to run the final model on its own). Third, we used a 300-m buffer to assess green and blue space, reflecting recreational opportunities at almost immediate vicinity from home location. Fourth, we added only one subjective environmental factor at a time to the main model to get an idea how much their potential overlap might have affected their effects in the main model. Finally, we specified a model with all the perceived environmental factors without the objective environmental factors to investigate whether and how much the associations were affected by them.