Context: On July 15th I defended my master’s thesis named: “Rail Ridership Response to Planned Highway Closures: A Revealed-Preference Analysis of the Dutch Network”. This thesis, worth 30 ECTS, is the last piece of my formal education and rewarded me with a degree in Transportation, Infrastructure and Logistics at Delft University of Technology. During the thesis I worked at NS, the Dutch Railway operator, to obtain the data necessary for this analysis. The thesis was awarded with a 9.5/10 on the Dutch grading scale.

Read the full thesis here: PDF

Read the paper here: PDF

Highways need maintenance and maintenance requires the closure of the road. Although this is often done by closing some but not all of the lanes, Dutch practice often has the national highway authority, Rijkswaterstaat, close off entire highways for multiple days. This method is like ripping off a band-aid: the maintenance is performed quicker but the disruption is larger. This disruptions changes the relative attractiveness of the car towards the train and could therefore spur a modal shift during these disruptions. For NS these disruptions therefore raise the same operational question: how many extra passengers should we expect and where will they show up? Get it wrong and the train is either overcrowded or you pay for unused capacity.

This question was the starting point for my master’s thesis at TU Delft, carried out as an internship with NS. The aim was simple to state but considerably harder to answer: What is the causal effect of planned highway closures on intercity rail ridership in the Netherlands?

Building a closure dataset that did not exist

The first obstacle was that no public record of highway closures exists in the Netherlands. To get around this, I used open traffic intensity data from the loop detector network and inferred closures from sustained drops in vehicle counts on affected road segments. This produced a dataset of 215 full-day closure events, accounting for 882 closure days across 2024 and 2025, cross-validated against known maintenance windows. Roughly three quarters of all origin-destination pairs in the rail network were exposed to at least one closure during the study period.

Isolating a causal effect

Simply comparing ridership on closure days to ridership on ordinary days is not enough. Ridership varies by season, by day of the week, by year, and closures themselves are not randomly distributed across the network. To isolate the causal effect, I used a restricted two-way fixed effects panel model on NS’s automated fare collection data, covering 1,743 origin-destination pairs. The model controls for corridor-specific and time-specific patterns using calendar dummies for day of week, month, holiday periods and year, so that what remains is attributable to the closure itself rather than to the season or the day.

The key explanatory variable is treatment intensity, a measure of how much longer a car journey takes because of the closure relative to normal conditions. This was interacted with the closure’s duration, split into three bins: two to three days, four to seven days, and eight days or more.

Duration is what matters

The headline result is that duration, not just detour severity, governs whether people respond at all. Closures of two to three days, typically weekend maintenance windows, produce no significant change in ridership. The same is true for closures of four to seven days. Only once a closure runs for eight days or longer does a significant ridership increase appear, and from that point the size of the increase scales with how much worse the detour makes driving. At the median treatment intensity for this duration bin, ridership rises by around 3.7 percent, and the effect grows as the detour becomes more severe.

This pattern lines up with literature and basic human behaviour. A short closure can be mitigated by cancelling plans, working from home, or accepting a detour. Longer closures forces you to actually reconsider your routine as working from home for a week might not be possible, or maybe you have family members to tend to.

Not everyone responds the same way

Averages hide a lot which is why I also performed heterogeneity tests aimed to identify how passenger groups respond individually as well as seeing how response varies by time-of-day. The extra ridership during long closures is driven largely by passengers without an established rail habit, people paying full fare for a journey they would not normally make by train. Students and business travellers on employer-paid cards also respond more than the average, which makes sense given that neither group faces a marginal cost at the point of travel that would otherwise discourage the switch. Commuters on personal subscriptions barely move at all, since they have already made their transport choice and a temporary closure does little to change it.

The timing compounds the problem for NS. The extra passengers arrive mostly during peak hours, precisely when the network already has the least spare capacity. A closure effect that looked, on paper, like a moderate 3.7 percent average uplift, turns out in practice to concentrate on the trains that are already busiest.

Real world tests

A model is only useful if it holds up outside the data used to build it. I built a prediction pipeline that turns a planned closure’s characteristics into a forecast of the expected ridership uplift, and tested it against ten closures from 2026 that the model had never seen. For a 44-day closure on the A28, the pipeline predicted an 8.3 percent uplift against an observed, control-adjusted increase of 6.7 percent, comfortably within the confidence interval. Short closures on the A12 correctly showed predicted effects too small to distinguish from ordinary day-to-day noise, confirming that no capacity response is warranted for them. The exception was a closure of the Vlaketunnel on the A58, a single-bottleneck corridor with no comparable alternative route, where the observed response was two to three times larger than predicted. This closure exposed the boundaries of the model. corridors with a single point of failure behave differently from the well-connected highway network that the model was trained on.

For NS, the practical takeaway is a three-tier way of reading any planned closure. Inside the bounds of the estimation sample, the forecasts are accurate enough to inform how many extra trains and staff to plan. For short or low-intensity closures, the correct response is to plan nothing extra, because no measurable ridership change should be expected. For closures in single-access corridors that fall outside the model’s experience, the forecast should be treated as a floor rather than a ceiling, a signal that the corridor deserves closer attention rather than a number to plan against directly.

Planned highway closures should be treated just like events where a different travel pattern can be expected and a tailored response is required. Highway closures are communicated well in advance and a response based on the output of this thesis can be easily formulated, the question is if NS actually has the capacity to respond to the ridership increase, especially during the peak days.