Cities and Wellbeing
Last post, I discussed the pace of life in cities and how life feels that it moves faster in larger cities. However, we saw that there is a bit of ambiguity in how this effect is measured, and more importantly, we saw that the literature is all over the place in explaining why this pattern is observed.
Today we are taking a look at the related topic of wellbeing in cities. I’ll acknowledge at the outset that this is a politically charged topic. Many people “know” that large, dense cities are good for wellbeing, in that they foster more social opportunities. Many others “know” that large, dense cities are bad for wellbeing, since the stress of crowding is harmful. Once again, though, we will see that the research is ambiguous. Most likely, both sides of this debate err by overestimating the importance of size and density in human wellbeing.
As always, today’s work is part of a Living Literature Review grant from Giving Coefficient. All conclusions are my own and do not represent Giving Coefficient’s views.
Large Cities and Depression
One of the first challenges we have is that we don’t have a consistent definition of wellbeing, and so the results that we are discussing today are often not directly comparable. I have chosen to take an expansive view of wellbeing, so much of this challenge is the result of my own decision rather than any shortcoming in the literature. Anyway, we’ll start with the question of whether people tend to be more depressed in large cities.
Miles, Coutts, and Mohamadi (2011) tackle the question by examining several neighborhoods in Miami. They find that living in a neighborhood with higher residential density is associated with fewer depressive symptoms. They find that a greater amount of green space is associated with fewer depressive symptoms, but that this effect wasn’t statistically significant in the full model. They also find that a higher density of auto commuters is associated with more depressive symptoms.
Miles, Coutts, and Mohamadi (2011) measure depressive symptoms with the Center for Epidemiological Studies-Depression scale, a self-reported survey that is commonly used to measure depression in the general public, as described by Radloff (1977).
The findings of Miles, Coutts, and Mohamadi (2011), as described above, are specific to a multivariate regression model. That is a linear regression taken on all of the variables simultaneously. They also do ordinary least squares regression on each of the variables individually, and for that, they do not find a significant relationship between depressive symptoms and either housing density or auto commute density.
It is worth pointing out that residential density and auto commute density are themselves highly correlated. The correlation coefficient is 0.84; a value of 1.0 would mean that there is a perfect linear relationship between them, a value of 0.0 would mean no correlation whatsoever, and a value of -1.0 would mean a perfect linear relationship, but in the opposite direction. As explained by Kim (2019), multivariate regression with multicollinearity (that is, independent variables that are highly correlated) leads to unreliable results. Here is a detailed overview by Jim Frost of the multicolinearity problem and ways to address it.
Another weakness of Miles, Coutts, and Mohamadi (2011) is that it does not address selection effects or other causality issues. Selection effects mean that perhaps, for whatever reason, people who are prone to depression tend to gravitate to neighborhoods with lower density or more auto commuters, rather than low density or more auto commuters being the cause of depression. Hoogerbrugge and Burger (2022) is one of the minority of studies that attempts to address the selection bias problem. They find that subjective wellbeing in cities in the United Kingdom is worse than in rural areas, but that this effect can be explained, at least in part, by less satisfied people being more likely to make a rural-to-urban move.
Also on the subject of causality, one especially devilish problem in regression and machine learning is the hidden variable problem. This is the situation in which other variables, not studied in the model, are the true explanation for the dependent variable. Elidan et al. (2000) address this phenomenon in the context of probabilistic models and suggest some complex solutions. In the particular case of Miles, Coutts, and Mohamadi (2011), they find that housing density is weakly correlated with economic deprivation and negatively correlated with housing stability, and so maybe these two variables, rather than density per se, is the true cause of the apparent association between density and depression. The data and analysis simply don’t allow us to tell.
In summary, while Miles, Coutts, and Mohamadi (2011) find that housing density and auto commute density are, respectively, negatively associated and positively associated with depressive symptoms, the paper suffers from so many conceptual problems that the conclusion is not convincing.
Before we get back to city size, Galea et al. (2005) is worth discussing for studying the relationship between housing and neighborhood quality and depression. They find,
In adjusted models, persons living in neighbourhoods characterised by poorer features of the built environment were 29%-58% more likely to report past six month depression and 36%-64% more likely to report lifetime depression than respondents living in neighbourhoods characterised by better features of the built environment.
The study was conducted in New York City, with data collected via telephone surveying in 2002. The original purpose was to assess the impact of the September 11 attacks on mental health, with proximity to the World Trade Center as a key neighborhood variable, but the purpose shifted between the sampling dates and publication.
Evans (2003) also tackles the built environment and mental health. He doesn’t address density directly, but he does address some features of the built environment that are associated with density. For example, he finds that “[h]ighrise housing is inimical to the psychological well-being of women with young children” and “loud exterior noise sources (e.g., airports) elevate psychological distress but do not produce serious mental illness”. Noise pollution may be one of the important hidden variables in these density/mental health studies and may be worth taking a closer look at later.
Park, Koo, and Kang (2026) are interested in the impact of density on mental health, particularly in light of the COVID-19 pandemic. Considering 25 neighborhoods in Seoul, they find,
First, the results revealed that population mobility density, residential density, and public transportation congestion were consistently positively associated with depressive symptoms across both study periods (2020 and 2021). Second, social participation significantly moderated the relationship between urban density characteristics and depressive symptoms, regardless of the pandemic phase.
This study also fails to address selection or hidden variable issues. Additionally, it would have been helpful to have a non-COVID year as a control, so that we could see how many of these results were in fact driven by the pandemic.
City Size and Relationship Quality
Another aspect of wellbeing is social wellbeing, or the quality of one’s interpersonal relationships. Mouratidis (2018) studies several neighborhoods in Oslo, Norway, some of which are suburban and some of which are compact. He finds statistically significant associations between compactness and several metrics related to social wellbeing:
Opportunities to meet new people,
Frequency of meeting with relatives and friends,
Number of close relationships,
Support from close relationships (significant at p=0.05 but not 0.01),
Personal relationships satisfaction.
To address the self-selection problem, Mouratidis (2018) augments the statistics with structured interviews with participants. The interviewees stated that they had more time and opportunities for socializing in compact areas and some stated that they chose to move specifically for that reason. How exactly this mitigates the self-selection problem, I am not sure, but at least he addresses the issue.
Dempsey, Brown, and Bramley (2012) tackle the question using questionnaires and focus groups. They find that higher densities in United Kingdom studies do indeed foster greater access to services and facilities. However, they find that less social interaction tends to occur in denser neighborhoods.
City Size and Wellbeing
Now we move on to the slippery concept of wellbeing, starting with Hebert et al. (2023) in a report for the planning, design, and research firm Happy Cities, which is centered around Charles Montgomery’s Happy City: Transforming Our Lives Through Urban Design (Montgomery 2013). Maybe I’ll discuss that book later.
Hebert et al. (2023) uses surveys to assess seven types of wellbeing, which are then aggregated into three metrics: general wellbeing, social wellbeing, and neighborhood wellbeing (actually, neighbourhood wellbeing; the firm is based in Canada). They surveyed 1886 people in the Vancouver, British Columbia area, for which they were able to determine postal codes for 1565.
A major challenge is that respondents are not equally representative of the population. Some groups, such as women and older people, are overrepresented, meaning that a greater proportion of them answered the survey than live in the area. For public opinion polls, this is a well-known problem. A common—though not always sufficient solution according to Mercer, Lau, and Kennedy (2018) for the Pew Research Center—is weighting. This techniques put a heavier value on respondents for underrepresented demographics so that the representation of each demographic in the survey matches their prevalence in society at large. No weighting was done in Hebert et al. (2023), and it is not clear how much of a difference this would have made.
The first major finding is that the wellbeing metrics are not significantly correlated with density. That was found by dividing the neighborhoods into five density zones. The second major finding is that wellbeing and housing type (i.e. single family or multifamily) mostly don’t correlate when controlling for income. The exception is that units less than 300 square feet and basement units are associated with lower wellbeing, even when controlling for income.
Ala-Mantila (2017) considers the spatial aspects of subjective wellbeing and asks the question, are dense, non-car dependent neighborhoods or sparse, car-dependent neighborhoods more conducive to wellbeing? She considers two wellbeing metrics: happiness and quality of life. Happiness is measured by the question, “Over the past 4 weeks, for how much of the time have you felt happy?” Quality of life is measured by the question, “We ask you to think about your life in the past two weeks, how would you rate your quality of life?”, and the surveyee is asked to choose from five ratings.
These two metrics sounds like they should be very similar, but after controlling for socioeconomics, Ala-Mantila (2017) finds that quality of life is higher in pedestrian-oriented zones, while happiness is higher in car-oriented zones.
We’ll give the last word to Astell-Burt and Feng (2015), who conduct a review of 103 studies on the effect of the built environment on wellbeing and mental health. Of the studies that find a relationship between the built environment and wellbeing, the most important feature identified was access to green space. The studies mostly were cross-sectional (i.e. they measured a single point in time, making it impossible to assess causation) and relied heavily of self-reported data. No studies used randomized control trials.
Conclusion
The reader will, I imagine, be underwhelmed with today’s selection of studies. I did not go looking for weak or inconclusive studies. That’s just what what the lay of the land is. The subject of wellbeing and the urban form simply doesn’t have the data that would be necessary to reach strong conclusions. Some of the common problems are that the that the studies don’t allow us to assess causation and that they have all sorts of risks of hidden variable bias.
There are two other problems I want to highlight. First—and I again admit that I did this myself by structuring the post as I did—terms such as mental health, happiness, quality of life, and (subjective) wellbeing are frequently used interchangeably. However, they all have somewhat different meanings, and even for one specific term, the meaning can be fuzzy. As Ala-Mantila (2017) finds, even the direction of the regression (positive or negative) can depend on seemingly subtle differences in the terms. It is enough to make me wonder whether the results are even meaningful.
A second problem lies in the word “significant”. I’ve used the term statistically significant many times here and in previous posts, but what does that mean exactly? Significant is the likelihood that the observed result would have arisen from random chance if there was no relationship. That likelihood is called the p-value. For example, suppose we want to determine if a variable X has any influence on a variable Y. If we do our test and find a p-value of 0.01, it means that there is a 1% chance that the correlation we see, or something stronger, would have arisen by chance if there was in fact no correlation. A p-value of 0.05 is typically used to denote statistical significance. Wasserstein, Schirm, and Lazar (2019) criticize the practice of using p-values this way, but it’s what we have.
Statistical significance, however, does not necessarily mean significance in a practical or policy sense. With enough data, even the tiniest relationship can rise to the level of statistical significance. However, metrics such as self-reported happiness might be able to tell us that people are happier with X than with Y, but they offer no insight into how much happier, which is one of several shortcomings that make subjective wellbeing unusable for policy.
Perhaps in the future, enough data, methodological advancements, and improved rigor will allow us to unambiguously answer the question of how the urban form governs happiness. But if there is an effect that is of great importance to policy, I strongly suspect that it would be obvious by now.