Combining disparate surveys across time to study satisfaction with life: the effects of study context, sampling method, and transport attributes

While the concept of happiness has long fascinated philosophers, positive subjective well-being as an academic field of study saw formal and widespread development starting in the 1970s (Diener 1984). As an individualized measurement of well-being (Mokhtarian 2019), SWB serves as a reflector of critical societal metrics such as economic development, social progress, and government policy (Diener 2000). However, the definition and measurement of SWB are more complex relative to traditional social indicators like gross domestic product (GDP). Conceptually, SWB has been defined to have two main components: hedonic and eudaimonic well-being. Hedonic well-being (HWB) refers to pleasure attainment and pain avoidance, while eudaimonic well-being (EWB) is based on the idea of self-actualization (Ryan and Deci 2001). In this study, we focus on a component of SWB, life satisfaction, which represents individuals’ conscious evaluation of their lives (Pavot and Diener 1993). Traditionally, life satisfaction has been considered a component of HWB; however, Huta and Ryan (2010) showed that life satisfaction may be related to both hedonic and eudaimonic perspectives of SWB.

Life satisfaction (along with other SWB components) is typically measured using either single-item or multi-item methods, each of which has differing strengths and limitations (Diener 1984). Single-item measures refer to short, clear survey items that can be implemented independently of other items. The simplicity of single-item measures requires less effort from respondents and survey developers, which makes them more suitable for inclusion in non-SWB focused surveys. In contrast, multi-item measures, such as the Satisfaction with Life Scale (SWLS), are composed of a group/set of survey questions, thus allowing researchers to check internal consistency and/or quality of responses, and to obtain a richer, more-nuanced, measure. Recent literature has shown that single-item measures of life satisfaction perform similarly to the multi-item measures in SWLS studies (Atroszko et al. 2017; Cheung and Lucas 2014; Jovanović, 2016); and furthermore, Diener (1984), one of the SWLS developers, recommends the use of single-item scales when a brief measure of global well-being is needed, noting that they provide adequate validity and reliability for such purposes. Accordingly, we use a single item to measure respondents’ satisfaction with life, in a series of travel-behavior oriented surveys that will be analyzed in this study.

Here, we summarize findings from the literature on the effects of (1) demographic characteristics, (2) travel-related attributes, (3) personality and attitudinal variables, and (4) contextual variables (e.g., GDP per capita) on SWB, focusing particularly on these four subgroups of explanatory variables as these were the ones available in the study at hand. Furthermore, many of these variables are commonly studied explanatory factors for life satisfaction in the literature, in addition to domain-specific factors such as health, occupation, and community/friendship. Although the latter have also been found to influence SWB measures, we do not explore them further given that they are not within the scope of this study.

This study utilizes a fused dataset of five California-based, transport-oriented cross-sectional surveys covering a 27-year period from 1992 to 2018. These five surveys were selected for this specific analysis as they all contained the same life satisfaction question, which is the key dependent variable in this study, and were implemented within California, creating a degree of homogeneity of geographic context. One co-author of this paper was responsible for or integrally involved with the survey design, development, and implementation processes across all surveys considered for this data fusion process, while other co-authors were responsible for or integrally involved with one or more of them. Table 1 provides an overview of key characteristics of the surveys; here, we discuss these characteristics, providing additional context regarding the original goals and distinctive features of each survey.

The earliest survey, conducted in 1992, was deployed with the overarching goal of studying telecommuting, then defined as the concept of working from home or a location closer to home, thus eliminating the need to travel to and from work (i.e., commuting). Telecommuting was hailed as an exciting game-changer in the early 1990s, and seen as a strategy for reducing congestion and emissions. The survey was deployed to six of twenty-seven departments in the City of San Diego, California, and all regular employees within those departments were surveyed. This convenience sampling method was not used in any of the remaining five surveys, and resulted in a very high response rate of 44% (Mokhtarian and Salomon 1996).

The next survey in the series was conducted six years later in 1998, and was deployed with the intention of measuring the existence and impact of positive attitudes toward travel itself (particularly local daily travel), in contrast to the conventional view of travel as a disutility, undertaken purely for instrumental reasons of reaching a desired destination. This survey used address-based simple random sampling across three judgmentally-selected San Francisco Bay Area neighborhoods with diverse land-use, travel, and demographic patterns; the overall survey had a response rate of 24% (Curry 2000). Although surveys involving the same investigator that were conducted in 2003, 2006, and 2009 were also considered, the next one with enough commonalities to be included in the fused dataset of this current study occurred in 2011.

The 2011 survey focused on multitasking during commute travel (i.e., attitudes and behaviors related to travel-based multitasking). This survey used several sampling methods (see Table 1) with the goal of obtaining sizable (rather than representative) shares of all pertinent means of travel (modes). Survey distribution channels included physical distribution on public transit, employee/student email lists from organizations including a large university, a large email list of Sacramento-area commuters interested in alternatives to solo driving, email and mailing addresses purchased from commercial marketing companies, and survey links posted on transportation agency and corporation websites. Where measurable, response rates across the various sampling channels varied from 0.23% (for the email addresses purchased from a marketing corporation) to 18.2% (for the physical distribution on public transit; Neufeld and Mokhtarian 2012).

The next survey included in the dataset was conducted in 2015, and focused on the mobility choices of Generation Y (Millennials, born in 1981–1997) and Generation X (born in 1965–1980). This study used an online opinion panel, and applied a quota sampling approach using targets for gender, age, race, ethnicity, household income, and presence of children in an effort to ensure diverse representation of the population in California. This sampling method resulted in a high response rate of 46.3%, which is not surprising as the invitations for this survey were sent to members of an online opinion panel. These individuals had previously self-selected into the panel, and were available to complete surveys in exchange for incentives in the form of reward points, gift cards or airline miles (Circella et al. 2016).

The final survey included in the fused dataset was fielded in 2018, which is the second wave of the 2015 survey. The 2018 survey aimed to study the impacts of emerging technologies (e.g., ride-hailing services, autonomous vehicles) and transportation trends through a unique longitudinal approach. Specifically, part of the sampling frame of the 2018 survey came from the 2015 survey respondents. Among the re-contacted respondents, 246 people completed the 2018 survey. The rest of the 2018 survey respondents were recruited from two sources: a stratified, address-based random sample, and a quota sample from an online opinion panel (Circella et al. 2019, see Table 1 for the response rate).

The basic process we undertook to fuse the surveys discussed in the preceding section can perhaps serve as a framework for others interested in such an approach. We first developed a question “inventory” spreadsheet, which included all survey questions and possible response choices across the surveys. Next, we categorized the survey questions using a four-level hierarchical classification system, the first level of which specified a general, broad category for each survey question. General categories included attitudes, socio-demographic characteristics, travel attributes, land use characteristics, and specific survey focused questions. Next, we categorized each survey question according to its specific topic; for example, general values, environment, lifestyle, travel, time use, work, etc. The remaining categories in the hierarchy simply specified in greater detail what aspect of the topics are covered in each question. This hierarchical labeling system then allowed us to group questions that conveyed essentially identical meanings based on content and possible answer choices, while allowing for minor wording and formatting differences. In the next section, we provide descriptive statistics for the dependent variable in this study (i.e., satisfaction with life), as well as the common explanatory variables used across the surveys, and detail some of the adjustments that were made to facilitate the data fusion process.

For the purposes of this analysis, we restrict our sample to commuters only (i.e., workers who travel to and from work), to increase the homogeneity of context across all survey datasets. Furthermore, by filtering out the non-workers, we are able to retain more travel-related (specifically, commute-related) variables in the model. After removing inattentive and incomplete cases, the final fused dataset comprises 7,514 valid cases for use in this study (see Table 1 for sample size by survey). Each of the five surveys obtained life satisfaction ratings by asking respondents to rate the statement “I am generally satisfied with my life” using a five-point Likert-type response scale ranging from strongly disagree (1) to strongly agree (5).

Figure 1a presents the life satisfaction rating distribution for each survey, and illustrates that there is an increasing trend over the years with the exception of the survey conducted in 2015—which only used an online opinion panel to sample respondents, and also had the narrowest age range. This observation prompted us to compare the group of all opinion panel respondents (part of the 2011 and 2018 sample [Fig. 1b], and all of the 2015 sample) to everyone else, and we found that there was a significantly lower mean life satisfaction (t = 9.326, d.f. = 5450.8, p < 0.001) for the opinion panel group. However, because recruitment via an opinion panel (particularly in 2015; see the “Cross-sectional surveys” section and Table 1) is somewhat confounded with belonging to age groups whose SWB might be expected to be lower than average (Millennials having entered the workforce during a major recession, and Gen Xers being near the bottom of the U-shaped relationship of age to SWB as described in the “Demographic characteristics” section), we revisit this effect in the “Model estimation and analysis” section, using generalized ordered probit models that control for age and other variables. The overall mean for the fused dataset was 3.92 (out of 5).

Figure 2 presents the life satisfaction rating distribution by region. We see that among the three California regions, Northern Californians have the highest average life satisfaction, followed by respondents from “other” California regions and Southern Californians.

Table 2 summarizes descriptive statistics for the explanatory variables used in the models, which are presented in the “Model estimation and analysis”. Prior to executing the models, it was necessary to transform several variables to obtain a consistent scale across surveys. For example, the original five surveys have different household income categories to reflect income distributions at the time of survey implementation. For this modeling effort, we used the mid-point of each income category from the original surveys and converted this to the equivalent purchasing power in June 2018 (the implementation date for the last survey in the fused dataset). To do this, we used the consumer price index (CPI) from the Bureau of Labor Statistics to convert all category midpoints to “2018 dollars”, after which we classified the converted household incomes into the six income categories used in the 2018 survey, in keeping with the need for consistency across surveys. The income distribution after conversion is shown in Table 2. This is detailed here as an example of the types of consistency conversions necessary when fusing cross-sectional surveys across time.

To further simplify model development, we treat income and education as continuous variables in the models. The original age statistics for each survey are listed in Table 2. In the modeling portion of this paper, we use mean-centered age to provide more natural interpretations of the impact of age (by considering changes from the mean age rather than changes from 0 years old) and to introduce a quadratic term into the models (Enders 2016). Table 3 shows the slightly differing statements representing the five attitudinal variables retained in the fused dataset; respondents were asked to rate the statements using a five-point Likert-type response scale. We also note that similarly modest differences in wording exist not only in the attitudinal statements, but also in survey questions that obtain demographic and travel-related characteristics. We acknowledge that the subtle differences in wording between the surveys may influence respondents’ final responses, but note that such differences will be common consequences of fusing disparate datasets. We also exercised what we believe to be conservative judgments about the extent of differences that we considered acceptable; i.e., we eliminated questions/statements that we considered likely to result in rating differences across surveys. As a result, all retained variables are believed to convey the same meaning to respondents across surveys, although slight variations in wording are still present.

Overall, the full sample consists of slightly more females than males (54% vs. 46%). More than 60% of respondents hold a bachelor’s degree or higher. Fewer than 20% of respondents live alone, and most of them (97%) hold a driver’s license. About two-thirds of the survey respondents come from Northern California (66%), compared to 25% Southern Californians and 9% Californians from other regions.

In this study, satisfaction with life is measured using a five-point Likert-type response scale from “strongly disagree” to “strongly agree”. As such, ordinal logit (OL; ordered logit) models would serve this analysis well. However, the dataset used in this analysis violates the parallel lines assumption of OL models, a violation that frequently occurs in practice (Williams 2006). The parallel lines assumption requires corresponding coefficients (with the exception of the intercept) to be identical across different levels of the dependent variable.

Therefore, we use a less restrictive form of OL, the generalized ordered logit (GOL) model. GOL models relax the parallel lines restriction on the explanatory variables that violate this assumption, while keeping coefficients for the remaining explanatory variables in a parsimonious form (Williams 2016). We also considered using the multinomial logit (MNL) model, since MNL does not impose the parallel lines assumption across explanatory variables; however, MNL does not take into account the ordinal nature of the dependent variable. Furthermore, in empirical practice, the MNL specification would result in substantially more model coefficients than GOL. As a result, to balance parsimony, conceptual fidelity, and interpretability, we selected GOL models for use in the analysis presented here. The gologit2 specification in Stata/IC15.1 was used for model development (Williams 2006).

The GOL model specification can be written as follows: where \({Y}_{i}\) represents the life satisfaction of respondent \(i\); \({\alpha }_{j}\) is the constant term associated with response j; \({X}_{i}\) is a vector of explanatory variables; \({\beta }_{j}\) is the corresponding vector of coefficients associated with response j (some, but not necessarily all, elements of which may be equal across some values of j); and \(M\) is the number of life satisfaction categories, which is five in this study; \({\alpha }_{j}+{X}_{i}{\beta }_{j}\) can be interpreted as the observed propensity for life satisfaction to be greater than response j. GOL models are different from OL models in that they allow \({\beta }_{j}\) to vary with \(j\); otherwise, the specifications for GOL and OL models are identical. The probability of each rating category is:

From this it can be seen that an increase in \({\alpha }_{1}+{{\varvec{X}}}_{i}{{\varvec{\beta}}}_{1}\) will unequivocally decrease the probability of the lowest life satisfaction response, and an increase in \({\alpha }_{M-1}+{{\varvec{X}}}_{i}{{\varvec{\beta}}}_{M-1}\) will unequivocally increase the probability of the highest response, but the effect of increases in \({\alpha }_{j}+{{\varvec{X}}}_{i}{{\varvec{\beta}}}_{j}\) on the middle three responses is ambiguous (Greene 2018). For this reason, we will interpret variables that have consistent effects across life satisfaction levels (i.e. for which the \({\beta }_{j}\)’s are equal across j) in terms of increases or decreases in the propensity for life satisfaction, which is unambiguous. For variables that have differing effects across life satisfaction levels (i.e. for which the \({\beta }_{j}\)’s vary across j), we will calculate the average marginal effects to illustrate their impact on each life satisfaction response level.

The final GOL model developed for this analysis relaxes the parallel lines assumption for household income and the three context-control variables (i.e., GDP per capita, unemployment rate, and the opinion panel indicator), as we found that there are different effects of household income and context-control variables across life satisfaction levels. Table 4 shows the final model results, and indicates that overall, the full-sample model has an acceptable model fit with a \({\rho }^{2}\) (equally-likely base) of 0.30. We will interpret the explanatory variables found to have significant predictive power for life satisfaction ratings. In addition, Table 5 presents the marginal effects of the explanatory variables whose coefficients have been allowed to relax the parallel line assumption, namely, household income, GDP per capita, unemployment rate, and the opinion panel indicator. Specifically, we present two groups of statistics—actual probability changes and percentage changes in the probability—since each offers meaningful but different insights in view of the unbalanced shares of the five responses. For example, a large incremental change in probability could represent a small percentage change if the baseline share is large, while conversely, a small incremental change could be a large percentage of a small share.

We look first at household level demographic characteristics. As expected, the model shows that increased household income levels tend to increase the propensity for satisfaction with life. Interestingly, the magnitudes of the coefficients decline as life satisfaction levels increase. This suggests that those with lower life satisfaction tend to have greater returns on their satisfaction propensity as their income levels increase. The trend is consistent with the marginal effects (percentage changes in probability). Specifically, a one-level increase in household income results in larger percentage changes in probability for lower life satisfaction levels than for higher life satisfaction levels. The second household-level variable is household structure, which shows that households with more members have an increased propensity to experience greater satisfaction with life relative to those who live alone. Consistent with numerous SWB studies showing the importance of social relationships in general, and close family ties in particular, living with other family members on net brings both practical and emotional support for the burdens of daily life.

Turning now to individual level demographic characteristics, we see that age has a U-shaped relationship with life satisfaction propensity. This indicates that individuals’ life satisfaction tends to have a declining trend during their early life stages, with the lowest life satisfaction occurring around the age of 44, on average. After this turning point, people have a greater tendency to be satisfied with their lives. The trend, which is consistent with studies in the literature (Beutel et al. 2010), is conceptually intuitive, as we can conceive that those in the middle-aged portion of life may have a greater number of stressors—career building, marriages, children, ailing parents—that may result in anxiety and decreased life satisfaction (hence the stereotypical mid-life crisis, e.g., Rosenberg et al. (1999). Figure 3 intuitively illustrates the average changes in the probabilities of the five response levels as age varies from 18 to 93. The probability of strongly agreeing with the life satisfaction statement forms a U-shaped curve across adulthood. In contrast, the probabilities of the other responses do not form such curves due to the restriction that the probabilities of the five response levels sum to one. Still, the figure indicates that people in middle age have the lowest propensity to be highly satisfied with their life and the highest propensity to be less satisfied. In terms of gender, the model indicates that, in general, females tend to have a higher propensity for life satisfaction than males (Welsch and Kuehling 2017). We see that individuals with higher levels of education tend toward greater life satisfaction. Regarding occupation, we see that those in managerial positions have an increased propensity to be more satisfied with their life relative to those in other occupations.

Regarding transport attributes, we see that increased commute time tends to have a negative effect on life satisfaction. Commuting is a recurring event, often conducted under time pressure and in less-than-pleasant circumstances, which occupies much of an individual’s “travel budget” and serves as a spatio-temporal anchor for many other activities. In the U.S., the vast majority of commuting is performed by driving an automobile, and especially for drivers but also for many others, commute time cannot be used as productively as may be desired (Shaw et al. 2019). For these reasons, among others, it is not surprising that longer commutes have a deleterious effect on travel well-being (Smith 2017) and overall life satisfaction (Hilbrecht et al. 2014).

On the other hand, we also see that having a driver’s license has a positive effect on life satisfaction propensity. As an instrument of motility (De Vos et al. 2013), license possession can increase one’s mobility by providing travel flexibility and an increased radius of potential travel. As witnessed by the literature on the mobility of the elderly, simply being able to drive can intuitively increase overall life satisfaction (Banister and Bowling 2004). Thus, we see that commuting specifically has an opposing effect to that of travel freedom generally—probably because of the lack of freedom found in much commuting.

The three contextual variables—GDP per capita, unemployment rate, and sampling method—have differing effects across life satisfaction levels. In general, people from regions with a higher GDP per capita are more likely to feel satisfied with life. Also, GDP per capita has the most substantial effects (i.e., the largest positive coefficient) on people who have a high propensity to feel satisfied with life. However, based on the marginal effects shown in Table 5, a one-unit ($1,000) increase in GDP per capita also results in a slight increase in the probability of the lowest life satisfaction level. Interestingly, the unemployment rate positively associates with life satisfaction, i.e., a higher unemployment rate is related to higher life satisfaction. One potential explanation resides in recalling the target group of this study, i.e., employed people. Considering their unemployed peers, people who do not lose their jobs when the unemployment rate is high may be more appreciative of their life than at other times. The marginal effects in Table 5 further show that a one-percentage-point increase in the unemployment rate will result in a probability increase of the medium–high level of life satisfaction (“Agree”) and probability reductions for the other life satisfaction levels.

In addition to the specific context-control variables of GDP per capita and unemployment rate, we considered including context-control indicators to capture the average impacts of certain unobserved contextual factors. First, we tested using region and year indicators as substitutes for the specific contextual variables in the full-sample model. See Table 7 in the Appendix for the model results. Compared to using specific contextual variables, the interpretability of the context-control indicators is relatively weak. For example, in the context indicator model, we see that life satisfaction has an increasing trend over the years,² while people from southern and other California regions have a higher life satisfaction than people from northern California. However, the sources of, or reasons for, these life satisfaction differences are unknown. Further, the fit of the context indicator model is lower than that of the model using specific contextual variables. We also tested including the context-control indicators together with the specific context-control variables; however, the two indicators were insignificant when GDP per capita and the unemployment rate were also in the model.

Regarding the sampling method, online opinion panel respondents tend to have lower life satisfaction propensities relative to respondents recruited using the other sampling methods (mainly address-based sampling). One possible explanation is that the two groups of people have different purposes for survey participation. The online panel respondents are people who previously registered at some websites for survey participation with rewards, while the other sampling methods (mainly address-based sampling) recruit essentially volunteer respondents (even though most of the surveys provided some small incentives). With this in mind, we can see that the online panel respondents may have more financial pressure and possibly more ennui, while the volunteer respondents are more likely both to be in a benevolent mood when they started the survey, and to have their sense of well-being further improved by the knowledge of being helpful by completing the survey.

To further analyze the effects of contextual variables, we executed a restricted version of the model with contextual variables removed (Table 4). As we can see, coefficients for the remaining variables do not change dramatically, which suggests the model’s stability and the independent explanatory power of the two groups of variables.³ The likelihood ratio test for the removal of a block of variables decisively rejects the null hypothesis that all the contextual variables are irrelevant (\({\chi }^{2}\) = 274.88, d.f. = 12, p < 0.001), which indicates the importance of contextual influences in life satisfaction modeling.

As noted, the attitudinal model (Table 6) is developed using a reduced version of the fused dataset, comprising only three surveys (the 1998, 2011, and 2015 surveys), and thus facilitating the inclusion of a larger set of common variables. To be specific, the attitudinal model includes all explanatory variables used in the full-sample model (as shown in Table 4) as well as seven additional variables that encompass travel-related attributes and attitudes. The attitudinal and full-sample models have similar model results (coefficient and significance) for the shared explanatory variables; accordingly, for economy of presentation, we here focus on interpreting the seven new variables.

Two of the seven additional explanatory variables are travel-related attributes that relate to whether respondents have physical conditions or anxieties that limit their use of transit or walking (i.e., Physical limitation—transit and Physical limitation—walk). Results indicate that those with physical limitations have lower life satisfaction propensities, likely due to restricted mobility. Further, the physical limitations may also reflect the presence of health problems that are not directly measured by our surveys, but which conceivably have negative effects on life satisfaction.

The five additional explanatory variables are transport-oriented attitudinal statements, which reflect general values as well. The first statement, “Like large yard,” captures a residential preference for living in locations that allow large yards and lots of space between homes. This statement is usually associated with a pro-suburban attitudinal factor in prior works (e.g., Kim et al., 2019). The positive coefficient indicates that people who prefer living in suburban areas have increased propensities to be satisfied with their lives. Not surprisingly, the attitudinal result is consistent with the behavioral result; for example, Sander (2011) found that those who live in less urban areas have higher levels of happiness. A possible explanation of the positive relationship between pro-suburban attitudes and life satisfaction is that the suburban life is still an aspiration for many Americans, and may therefore be perceived to offer a more satisfactory living environment. The second attitude, “car as a symbol”, reflects the respondent’s vanity with respect to owning a car. The results show that respondents who regard cars as status symbols tend to be less satisfied with their lives, perhaps because there will always be others with more status, to whom it is disappointing to compare oneself. Similarly, Olivos et al. (2020) have shown that upward social comparisons have negative effects on life satisfaction.

The last three attitudinal statements (“Don’t mind being stuck in traffic,” “Travel is wasted time,” and “Commute is stressful”) capture general preferences toward time use, travel liking, and commuting. Those who don’t mind being stuck in traffic have increased propensities for life satisfaction, while—consistently—those who see travel as a waste or commuting as stressful have reduced propensities. These are expected results, especially considering that such attitudes may reflect not only a specific affect toward travel, but also an optimistic/pessimistic outlook on life in general. Note that these attitudes substantially reduce the impact of commute time per se: when attitudes are excluded (as in the second model in Table 6), the coefficient of commute time becomes markedly more negative in compensation, which is not surprising. Nevertheless, even controlling for these attitudes, commute time remains strongly significant.

Thus, we see that the inclusion of these five attitudinal statements allowed for the interpretation of life satisfaction from more dimensions (e.g., personality, time use, travel liking). In Table 6, we also present a model identical to the one discussed here, except with the attitudinal variables removed. Similar to the results for the full-sample model, we note that the remaining explanatory variables have relatively stable coefficients after the attitudes are removed, and also that the likelihood ratio test for the removal of a block of variables decisively rejects the null hypothesis that all the attitudinal variables are irrelevant (\({\chi }^{2}\) = 177.06, d.f. = 5, p < 0.001), which indicates the importance of attitudes in life satisfaction modeling.