2026-06-30 Executive Team

[joecastiglione]

Agenda

• Technical updates
  • Explicit Telecommute Implementation (WSP)
  • EET Investigation of Inconsistencies / Output Analysis (RSG, Outer Loop)
• Administrative updates (partners only)
  • RFQ status
  • Survey results

Notes

Explicit Telecommute Model Update (Sijia)

Overview

Sijia presented on progress on two new model components, now renamed for clarity:

• Telework Arrangement Model — a multinomial logit model applied to all workers, predicting whether a worker has any in-home work activity on the simulation day (yes/no). Runs after CDAP and before tour/trip models.
• Telework Duration Model — applied only to hybrid and teleworking workers, predicting how much time they spend working at home. Supports either an MNL specification or a probabilistic (empirical distribution) draw, configurable via YAML, with support for segmented distributions (e.g., by person type).

Both models are estimated using SANDAG's post-COVID 2022 Household Travel Survey and are designed to avoid significant runtime or memory impact at full model scale.

Model Design Notes

• Worker-day outcomes are classified into four categories: commuting, hybrid, teleworking, or not working.
• The in-home work threshold is set at 3+ hours of telecommute time (consistent with Phase 9 methodology), to exclude short "spillover" work episodes (e.g., answering emails after returning home).
• Source code (teleworkarrangement.py) includes unit tests and is currently under internal review (David H).

Preliminary Results

• On average, ~41% of workers have some in-home work activity, led by full-time workers (43%), then part-time (27%) and university students (24%); driving-age students showed 0%.
• Using the 3-hour threshold and CDAP, ~6% of workers are hybrid and ~35% are pure telework — notably high, possibly reflecting the survey's proximity to the COVID period (2022).
• Significant predictors for telework arrangement include person type, presence of preschool children at home, higher income, age (35+ less likely), telecommute frequency, non-mandatory CDAP interactions, and business/services industry. Workplace-mode-choice logsum was directionally correct (negative) but insignificant.
• For telework duration, person type, income, hybrid-vs-telework status, and workplace-mode-choice logsum (positive and significant here) were key predictors.

Data Quality Issues Identified

• ~11–12% of non-home-based workers did not report a workplace location; addressed via a missing-workplace dummy variable.
• ~10% of workers reported "never telecommute" weekly but logged 3+ hours of telecommute on the survey day — likely a question-interpretation issue. The team kept the reported frequency variable unchanged to preserve consistency with other SANDAG-based estimation work.
• ~3% of workers reported both long telecommute hours and long out-of-home work tours the same day (some 7–8 hours each), suggesting some telecommuting may occur at non-home locations (e.g., co-working spaces) mislabeled as telecommute time. Still under investigation; may prompt reclassification.

Discussion Highlights

• Joe asked whether the 3-hour in-home threshold should vary by person type, given low durations for some segments (e.g., 24% for university students)
• Sijia explained how the signs on the accessibility parameters were logical: workplace-mode-choice logsum is negative (easier commutes reduce telework), while non-mandatory accessibility is positive (good local access encourages staying home to run errands).
• Joe questioned why only "business services" emerged as a significant industry; Sijia noted other industries tested were insignificant or counterintuitive, possibly because industry effects are already captured by correlated variables like telecommute frequency.
• Joe asked about validating outputs against observed distributions in a full-scale SANDAG application; Sijia agreed this is feasible though not formally in scope.

Next Steps

• Feedback deadline: 2 weeks from today. Reviewers should add comments directly to the shared Google Drive slide deck; Joe will adjust sharing permissions as needed and add a reminder to the agenda.
• Sijia will continue investigating the long-telecommute/long-work-tour overlap cases and finalize model estimation.
• Documentation of both models is a planned deliverable following estimation finalization.

Other Items

• EET update: deferred to a future meeting (~2 weeks out).
• RFQ/survey results: Joe will follow up asynchronously; a separate discussion can be scheduled if needed.