American Time Use Survey (ATUS)

• ATUS is a continuous, nationally representative survey that collects 24-hour activity diaries to model individual and group behaviors.
• It employs a detailed six-digit activity coding system and a well-being module to support granular spatiotemporal and affective analyses.
• ATUS data drives applications in exposure assessment, automation research, and sociobehavioral studies, enabling robust agent-based modeling.

The American Time Use Survey (ATUS) is a continuous, nationally representative instrument for quantifying daily activity patterns in the U.S. population. Conducted by the U.S. Bureau of Labor Statistics (BLS) since 2003, the ATUS provides granular, 24-hour diaries for modeling individual and group behaviors, supporting research in human exposure assessment, well-being analysis, automation preferences, and agent-based modeling of human activity. Its design, sampling methodology, coding schema, and integration with complementary modules (e.g., the ATUS Well-Being Module) enable advanced spatiotemporal modeling and sociobehavioral analysis (Lund et al., 2019, Ray et al., 10 Jan 2025).

1. Survey Design and Sampling Methodology

The ATUS employs a continuous, stratified random sampling procedure, targeting U.S. households that have recently participated in the Current Population Survey. For each survey wave, approximately 22,500+ respondents submit a 24-hour diary, with sampling weights provided by the BLS to permit construction of nationally representative estimates. Each diary covers one survey day, with activities recorded in minute-by-minute chronology, capturing:

• Six-digit ATUS activity code
• Start and end time (to the minute)
• Co-presence (who else was present)
• Demographics: sex, age, household composition, household income, and related variables

The design ensures coverage of wide demographic variability, with 2021 data reporting 45.5% male and 54.5% female, 68.2% White (with other racial and ethnic groups detailed), 83.9% categorized as middle income, and balanced representation of employment, caregiving, and disability statuses (Ray et al., 10 Jan 2025).

2. Activity Coding Schema and Well-Being Module

ATUS utilizes a hierarchical six-digit activity coding system. The first three digits represent a high-level group (e.g., 200 = "Household Activities"), while the full code specifies the activity in detail (e.g., 200101 = "Interior Cleaning", 200301 = "Food Preparation"). Diaries typically enumerate ~400 distinct activity codes per dataset; durations are recorded in minutes.

The ATUS Well-Being Module (ATUS-WB) supplements the standard diary with affective ratings for three randomly sampled non-personal activities per respondent (excluding sleep and self-care), assessing six affective states—Happiness (H), Meaningfulness (M), Painfulness (P), Sadness (B), Stressfulness (S), and Tiredness (Z)—each rated on a 0–6 ordinal scale (Ray et al., 10 Jan 2025). This dual data structure supports temporal, categorical, and affective analyses.

3. Computational Modeling and Feature Engineering

To enable agent-based and sequence modeling, ATUS data undergoes an extensive preprocessing and feature engineering pipeline (Lund et al., 2019):

• Selection of 12 demographic variables (e.g., TEAGE, TESEX, TELFS, TEHRUSL1, household children, elderly care, disability)
• Construction of:
  • Activity frequency vector $f_i \in \mathbb{N}^{400}$ per respondent, counting code occurrences
  • Time-slice vector $t_i \in \{1\dots A\}^{288}$, encoding the main activity in each 5-minute interval
  • Full feature vector $x_i = [f_i \| t_i \| d_i]$, with $d_i \in \mathbb{R}^{12}$
• Exclusion of diaries with missing key variables or rare codes, and standardization of all features to zero mean and unit variance

This structure enables high-fidelity cohort discovery, clustering, and synthetic diary generation using advanced machine learning techniques.

4. Unsupervised Stratification and Activity Sequence Generation

Lund et al. (Lund et al., 2019) detail an algorithmic pipeline for stratifying respondent diaries and synthesizing coherent, stochastic activity sequences suitable for agent-based modeling:

• Unsupervised Embedding:
  • Build an ensemble of $M = 2000$ extremely randomized trees (via scikit-learn's RandomTreesEmbedding, max depth 5) on $\{x_i\}$.
  • Compute Breiman’s proximity matrix $P \in \mathbb{R}^{N \times N}$; apply t-SNE yielding 2D embeddings; rescale to $(-1,1)$.
• Clustering:
  • Cluster embeddings using DBSCAN, with manual tuning of $\epsilon$ and minPts. About 30% of points may be labeled as noise.
  • Label extension via Extra-Trees, iteratively assigning unlabeled diaries and aggregating small clusters ($K \approx 76$ for activity-stratified classes).
• Synthetic Sequence Generation:
  • For each cluster/class, extract empirically derived distributions for activity start times, durations, and locations using Bayesian Gaussian Mixture Models.
  • Estimate transition probabilities, duration distributions, and location assignments.
  • Monte Carlo sampling generates full 24-hour synthetic sequences, with post-hoc insertion of travel events and normalization to fill exactly 24 hours.
• Validation:
  • Mode similarity (proportion of minutes where modal real and simulated activities agree) and Gini-index correlation (activity diversity over time) are used for assessment. 61% of classes achieve mode similarity and correlation $>0.8$; 95% exceed $0.6$.

The framework supports convolution of synthetic activity sequences with exogenous temporal-spatial exposure fields for environmental or behavioral simulations.

5. Integration with Well-Being, Automation Preferences, and Societal Analysis

Integration of ATUS and ATUS-WB with external datasets, notably BEHAVIOR-1K, enables multifactorial analysis of activity, well-being, and automation preferences (Ray et al., 10 Jan 2025):

• Alignment:
  • Manual mapping of BEHAVIOR-1K tasks (e.g., "mop the floor") to the ATUS six-digit activity schema; exclusion of wage-labor codes.
• Quantification:
  • Aggregation of "Desire for Automation" (DA) Likert ratings from BEHAVIOR-1K, matched to ATUS activity codes.
  • Computation of mean daily time, mean well-being scores (H, M, P, B, S, Z) per activity and subgroup.
• Statistical Analysis:
  • Nonparametric rank correlations (Spearman’s $\rho$, Kendall’s $\tau$) are used to compare the activity rankings across DA, time use (T), and well-being metrics.
  • Relative rank-difference $\Delta_i^{X - GP}$ quantifies subgroup divergences from the general population.

Key findings indicate that, contrary to common assumptions, time spent (T) on an activity does not strongly correlate with automation preference (DA). Instead, lower happiness and higher pain during activities are the strongest predictors of automation desire. Gender and income subgroups reveal specific patterns—women show stronger preference to automate stressful activities, while men prioritize activities associated with low happiness. Middle-income respondents emphasize automation for less meaningful tasks; no clear preferences emerge for low and high-income groups.

6. Applications and Limitations

ATUS data and derived models serve multiple domains:

• Exposure Assessment: Synthetic agent trajectories generated from ATUS enable high-resolution modeling of personal and population-level exposures to spatiotemporal environmental fields, using formulations such as

$$E_s = \sum_{i=1}^N \int_{t=\tau_i}^{\tau_i+\delta_i} E_\text{field}(t, x_i) \, dt$$

discretized in 5-minute slices.

• Social Robotics and Automation: ATUS-derived activity and affect statistics guide the prioritization and design of domestic robots by quantifying user preferences for task automation.
• Behavioral and Sociological Analysis: Standardized time-use and well-being measures enable robust subgroup comparisons and longitudinal trend analysis.

Limitations include a restriction to 24-hour diaries (excluding multi-day rhythms), coarse spatial granularity (only broad location types), manual hyperparameter tuning for clustering algorithms, and current exclusion of deep-learning-based sequence models due to data scarcity for longer time windows.

7. Open-Source Data, Reproducibility, and Future Directions

All primary datasets and processing scripts are open-access. ATUS public-use microdata and the Well-Being Module are hosted by the BLS, and BEHAVIOR-1K results are available on GitHub and an interactive web frontend [https://hri1260.github.io/why-automate-this/]. The generative modeling pipeline for ATUS activity synthesis is described in detail by Lund et al. (Lund et al., 2019), facilitating adaptation to other national time use surveys and spatiotemporal human modeling tasks.

Future research aims to extend temporal modeling beyond one day, refine geolocation accuracy, automate model-selection procedures, and integrate deep generative models if multi-day or longitudinal diaries become available. The ATUS remains central for empirically grounded, high-resolution population activity modeling and interdisciplinary research in exposure science, labor studies, and human-centered automation (Lund et al., 2019, Ray et al., 10 Jan 2025).