A cluster analysis of how Americans spend their weekends
What are your plans for the weekend? While it is easy to imagine that Americans spend their weekends differently, this is a challenging question to empirically validate.
Summarizing the average American’s weekend plans is an easy exercise in descriptive statistics. On average, the typical American spends 10 hours sleeping, six hours socializing and relaxing, two hours on household activities and one hour eating and drinking. This picture of the average weekend is shown below.
This plot describes the population average, but provides little relevant information about how actual individuals spend their time. The idea that different people spend their weekends in different ways is unsurprising but it is hard to quantify. Unsupervised machine learning, also referred to as cluster analysis or clustering, offers a solution to these types of unstructured problems. The method applies distance metrics and linear algebra to recognize patterns, and can quantify if different groups of Americans spend their weekends in different ways.
Data on time use was obtained from the American Time Use Survey (ATUS), which measures how Americans spend their time doing various activities such as work, household activities, volunteering, leisure, and socializing. The ATUS is the most comprehensive measure of time use and obtains a representative sample of how Americans spend their day.
The data and sampling
The cluster analysis used 16 years of ATUS data totaling to a sample of over 201,000 unique Americans. Filtering out responses to only include time spent on the weekend left a sample of approximately 100,000 unique people. Cluster analysis is a computationally expensive procedure, and including the entire sample in the analysis was not feasible. To maintain a representative sample of the American population, survey weights were used to draw 25,000 individuals to be included in the analysis.
The final cluster solution was able to identify six unique clusters, five of which were characterized by uniquely different activity patterns compared to the general population. The sixth cluster appeared to characterize the “average” American weekend.
Broadly speaking, considerable differences in time spent on work, on social and leisure activity, on household activities (i.e. home maintenance) and on taking care of other members of the household were uncovered. For example, one cluster spent 6.6x time working compared to the population average and another spent 7.7x time caring for other household members. These results suggest that there are distinct groupings of weekend activities and that these groupings can be empirically identified.
Clustering algorithms rely on calculating distance between data points and variables. There are several established distance metrics such as the Euclidean, Manhattan, Mahalanobis and many others. As a starting point, it is useful to examine the distances between variables to understand if an underlying relationship exists.
The distance matrix confirms some simple assumptions. For example, those who work more on the weekend spend less time sleeping. Similarly, those who work more on weekends have less social and leisure time. While these simple comparisons confirm the assumption that patterns exist they do not reveal any underlying complex structure.
Hierarchical clustering extends the use of distance metrics to understand the similarities and differences between data points (i.e. the people), rather than the distance between variables. A powerful way to visualize this distance is by plotting a dendrogram of the clusters. Hierarchical clustering groups observations by iteratively reducing the number of clusters — combining nearby data points with each iteration. Dendrograms are a way of visualizing these relationships and identifying existing groupings. In this context, branches that are closer to one another represent people with more similar weekends.
Frequently, variables are scaled prior to cluster analysis. In this particular context, variables are already recorded on the same scale so no centering or transformations were included and variables were left in their original untransformed format.
Identifying the final number of clusters is a nuanced and subjective exercise. That being said, mathematical and quantitative guidelines exist to aid this decision. To determine the number of groups, the Calinski-Harabasz index, Silhouette width and Gap-statistic were consulted.
These metrics, combined with the visualization provided by the dendrogram suggest that Americans’ weekend activities can be divided into six distinct clusters. Of the six final clusters, five represent distinct groupings while the sixth appears to be more general.
Hierarchical clustering is only one of many clustering methods. We considered the popular K-means algorithm as well as model-based clustering. A hierarchical approach offered the best fit to this particular data set. Within the Time Use Survey, many respondents indicate zero minutes spent on several activities. The K-means algorithm assumes all variables have an equal variance and does not handle zero-inflated variables well. These assumptions do not fit the specifics of time use. Similarly, model based clustering has more rigid assumptions that variables come from a multivariate normal data generating process. Hierarchical clustering avoids these assumptions and is a good fit for this particular clustering problem.
After assigning each person to a cluster, different groupings of weekend activities begin to emerge. Shown below, while all six clusters spent approximately equal time sleeping, there are distinct differences in the amount of time spent on work, on household activities, on socialization, and on caring for other members of ones’ household. In the plot, each row represents a distinct cluster and colors denote different activities.
To get a better understanding of each cluster, the plot below focuses on the most distinct differences in time spent between each cluster. Differences in these four primary activities account for the majority of differences in how Americans spend their weekends. These results show that the algorithm is able to identify those who spend their weekends predominantly working (Cluster 1), socializing (Clusters 2 & 3), performing household activities (Cluster 4) and caring for other members in the household (Cluster 5).
One of the values of cluster analysis is the ability to identify who makes up segments of the broader population. By linking cluster assignment to data from the Current Population Survey (CPS) a better picture of differences between clusters can be drawn.
Why explore demographic data?
Linking demographic data to cluster assignment has many different uses that range from marketing to academic research. For example, a long line of research has shown that women spend more time taking care of household members compared to their male counterparts. Investigating demographic associations with cluster assignment could be used as a tool to validate that women disproportionately spend more hours on unpaid work. In a different context, marketing firms could improve targeted advertisements by identifying segments that spend their weekends on different types of activities.
Comparing age, income, sex, children, marital status, and educational attainment between clusters revealed several patterns. While each cluster is made up of all demographics, demographics within each of the first five clusters differ from population averages. The visual below shows how age and income are not evenly dispersed between clusters. Some clusters tend to be younger, others are largely middle-aged and others skew older. Similar differences exist for different levels of household income.
Similarly, we see uneven dispersion in sex, marital status, children, and education levels. Some clusters skew towards female, others contain more married individuals, and others with more education. Notably, Cluster 5 contains almost exclusively individuals that have children in the household.
Comparing cluster assignment against demographic variables allows us to characterize which Americans make up the majority of each cluster. A brief demographic description of each cluster is provided below:
The cluster analysis showed that different Americans do in fact spend their weekends in different ways. The methodology was able to identify six unique clusters, five of which were characterized by uniquely different activity patterns compared to the general population. The sixth cluster appeared to characterize the “average” American weekend.
These findings could be applied in several ways. From a marketing perspective the ability to identity segments of the population may help improve targeted advertisements. From the perspective of a researcher, these findings align with prior work that shows gender differences in time allocated to unpaid work. This is one example of how clustering may be applied to identify existing social inequities.
Describing weekend time use is a methodological challenge — because the goal of such an analysis is to illustrate the relationship of many variables, there is no single variable to “supervise” the analysis. This work showed how clustering, an “unsupervised” approach, can be used to empirically validate that there are distinct patterns in how Americans are spending their weekends.
2020 May
Find the code here: github.com/joemarlo/ATUS-cluster-analysis