Why Tokyo commuters may choose crowded trains: the cost of arriving at the wrong time

A crowded train can still be the preferred option

Why board a packed train when a less crowded service is available? A preprint by Shota Fujishima and Ryosuke Matsumoto, posted on November 2, 2026, offers a model-based explanation: avoiding an inconvenient arrival time may matter more than avoiding crowding.

The study examines Tokyo’s Seibu Shinjuku Line using observed travel choices. Its estimates suggest that commuters accept crowded services partly to avoid arriving earlier or later than desired. That is not evidence that crowding is harmless, nor an experimental demonstration of what causes commuters to choose particular trains. It is an interpretation of observed behavior under an economic model.

Connecting survey responses to trains

The researchers use individual records from Japan’s 2015 Public Transport Survey for Metropolitan Areas, matched with contemporary train timetables. Their analysis includes 2,740 commuter-pass users, representing an expanded population of 140,473 commuters across 27 origin stations and 89 trains arriving by 10:30 a.m.

The modeled commuters travel to Takadanobaba or Seibu-Shinjuku, treated as one central destination, including passengers who transfer there. Other passengers contribute to crowding, but their choices are not explained by the model.

Express and local services offer different combinations of journey time, arrival time and crowding. Fares did not vary by time of day during the study period, helping the researchers focus on those nonmonetary trade-offs.

A crucial distinction is between when someone arrives and when they would prefer to arrive. Observed arrival times already reflect compromises. Instead of using them as a proxy for preferences, the authors construct desired arrival times from reported work start times, onward rail travel and travel from the final station to work. They also assume a preferred 15-minute workplace buffer, drawing on an external 1999 survey. The baseline simplifies these differences into a common desired arrival time of 8:10 a.m.

Crowding is an outcome, not just an input

A train becomes crowded because passengers choose it. If people switch to a quieter service, that service becomes less quiet. Treating crowding as fixed would miss this feedback.

The authors therefore model train choice as a population game: each commuter’s cost depends on their chosen train and how many other passengers use it, rather than on those passengers’ identities. Costs combine journey time, early arrival, late arrival and crowding exposure along the route.

The model assumes commuters occasionally reconsider their choices using a probabilistic rule that favors lower-cost options. A mathematical property called a potential game makes the long-run distribution of passenger allocations tractable. The researchers estimate preferences from a large-population approximation to that distribution. This adjustment process is assumed, not directly observed in the one-day survey.

What the estimates mean

In the baseline model, one minute of arrival earlier than desired carries a cost equivalent to 0.39 minutes of ordinary in-vehicle travel. One minute later than desired corresponds to 0.49 minutes. Here, “late” means later than the constructed desired arrival time, not necessarily late for work.

By comparison, one minute aboard a train at 100% rated capacity adds a crowding cost equivalent to 0.04 minutes of ordinary travel. Rated capacity includes both seats and standing room; 100% does not mean merely that every seat is occupied. These figures express estimated discomfort in comparable time units, not actual extra journey minutes or monetary payments.

The broad ranking survives alternative specifications, but the crowding estimate is fragile. When commuters are grouped more finely by work-start-related desired arrival times, estimated late-arrival costs rise substantially. The crowding parameter reaches its constrained lower bound of zero in three of four such specifications. That does not establish that passengers feel no crowding discomfort; it shows how sensitive its inferred value is to the scheduling assumptions.

Important limits on precision and policy

The baseline confidence intervals exclude zero for all four cost components, but they capture uncertainty conditional on the expanded population and model. They omit survey-sampling and expansion-weight uncertainty, potentially understating overall uncertainty.

Train identities are inferred rather than recorded directly. Some constructed crowding rates exceed 300%, which the authors consider likely to reflect matching errors. The model also assumes crowding discomfort rises linearly and does not distinguish sitting from standing. Its findings concern one line and historical travel patterns.

The policy implication is therefore a question to investigate, not a demonstrated remedy. Off-peak rewards may ask commuters to accept inconvenient arrivals; flexible or staggered work hours could change that constraint. This version does not quantify either policy’s effects or their interaction. Its contribution is to show why evaluating crowding policies requires attention to when people need to arrive, as well as how uncomfortable their trains are.