Resident Destination Choice

Size Term

This is the natural logarithm of the attractions at a destination zone, which itself is a linear combination of employment variables and their coefficients. These coefficients are simultaneously estimated with other parameters. Since it is required that these coefficients be positive, the utility formulation specifies coefficients that are then exponentiated. Successful estimation requires that one of these attraction variable coefficients is fixed. The estimation provides a coefficient for the size term, which theoretically should be between 0 and 1.

Mode Choice Logsums

Using mode choice root logsums typically presents a problem in destination choice. The root logsum value is dominated by the auto nest which washes out effects from non-dominant modes. For example, drastically improving the transit accessibility to a particular zone will not affect the probability of choosing that zone if the transit logsums are dwarfed by the auto logsums. To circumvent this issue, separate auto, transit, and non-household auto logsums from the respective nests are used in the TRM specifications (a non-household auto could be a neighbor’s car for example). The logsums are segmented by vehicle sufficiency market segments for added explanatory power. Work trip logsums are further distinguished by high- and low-income segments. Finally, note that there is no auto logsum component for the zero-vehicle households market segment.

Others

• 4D measures: Several 4D measures such as transit accessibility, hospital accessibility, walk accessibility, employment densities are considered.
• Time Coefficient(s): The auto travel times corresponding to the latest model skims are attached depending upon the time of day the respondent made the trip. A time coefficient is estimated in each specification.
• Intra-Cluster effects: Intra-Cluster coefficients (‘HomeCluster’ in the tables below) are estimated if the home and the chosen zone are within the same cluster (nest).
• Intra-Zonal effect: A coefficient is estimated to boost the utility of choosing the same zone as the origin.
• Cluster Nest coefficients: A nest coefficient for each of the 12 clusters is estimated.
• Cluster ASCs: Likewise, alternative specific constants (ASCs) are estimated for each cluster.

The rich utility specification allows the model to capture the decision-making process of choosing destination zones and capturing cluster-to-cluster flows.

Estimated coefficients and t statistics

ASCs were estimated for all but one of the clusters. Several had values close to zero with poor significance and were dropped. The ASCs are very small in magnitude, confirming that the model’s explanatory power is derived from other variables and is sensitive to changes in model inputs.

Intrazonal and home cluster effects are strong. Residents of Cary, Durham, Chapel Hill, and Garner often live and work in the same cluster. Similar effects are also observed for the peripheral clusters toward the northeast and south.

Root mode choice logsums are significant for workers from zero-auto households, capturing a coupling with access to public transit modes. Using root mode choice logsums for the other auto sufficiency markets caused unrealistic signs as well as much poorer model fit. This can be explained by the fact that these logsums are primarily driven by auto skims, which are already a part of the utility equation (the “Time” term).

Hospitals are major sources of employment in the region, and accessibility to hospital jobs helps drive work destination attractiveness. Transit and walk accessibilities are also drivers of a zone’s attractiveness for work tours.

Most of the cluster nest coefficients are significantly different from 1, which means the nested structure is justified. Research Triangle Park is a special cluster (very low residential land use, very high technology employment, etc.). Its coefficient is effectively 1, which implies an MNL-type sub-nest.

The size variables are stratified by income using the percentage of low and high paying jobs in each zone. This is how the model pairs high-income workers (e.g. in the Regency area of Cary) with high-paying jobs in RTP. This stratification is critical for accurate work flows in the region.

Double constraint and attraction model

All trip types (other than work) are singly-constrained. This means that the row sums of the resulting trip table will match starting productions, but column sums will not necessarily be proportional to the amount of employment in each zone. As an example, for two zones with the same employment, the more accessible zone will attract more trips.

For work trips, a traditional assumption to make is that each job of the same type must attract the same work trips - even if the zone is in a remote location. The justification for this assumption has weakened in recent years with the rise in telecommuting, flex schedules, and other changes, but the TRMG2 model still implements double constraint for work trips. In the context of destination choice models, this is achieved by assigning each zone a “shadow price”. This extra term in the utility equation is adjusted in an iterative fashion to match predicted attractions.

These predicted attractions come from a regression model estimated from the survey data.

During model application, predicted attractions are always scaled to match predicted productions before applying double constraint. As a consequence, the attraction model should predict total attractions that are close to predicted productions. This is referred to as PA balance (“production/attraction balance”). If it does not, adding employment to a zone will have unexpectedly high (or low) impacts on total zonal attractions.

The coefficients in the table above were multiplied by 2.45 during model calibration to achieve the appropriate PA balance of 1.0.

Estimated coefficients and t statistics

ASCs were estimated for all but one of the clusters, but only the significant ones were retained. These ASCs are very small in magnitude, confirming that the model’s explanatory power is derived from other variables and is sensitive to changes in model inputs.

The model has a sophisticated approach to travel time. Travel time is more important up to the first 30 minutes. Excess time beyond this threshold gets a slight discount. This reflects a difference in the behavior for these longer trips, which are dominated more by what is at the destination rather than the time to get there.

Home cluster effects are strong for residents of Cary, Garner, Raleigh, Southern Durham and the southeastern periphery. People in these clusters often stop in the same cluster for their intermediate activities.

The public transit mode choice logsum is significant for certain segments like zero-vehicle households. Non-household auto logsums (primarily taxis and TNCs) are also significant.

The number of households is a major factor in the calculation of the size variable, reflecting the inclusion of personal visits (e.g. to a friend’s house) in this trip purpose. Logically, the retail and service employment also help drive trips to specific destinations. Income-based stratification of employment did not make a difference in model performance.

Most cluster nest coefficients are significantly different from 1, so the nested structure is justified. As with W_HB_W, RTP is an exception to this.

Estimated coefficients and t statistics

School location choice is driven by travel time and school enrollment. Some of the cluster nest coefficients are significantly different from 1, so the nested structure is justified for those clusters. The remaining zones are treated as top-level choices (MNL).

Estimated coefficients and t statistics

As with the previous models, cluster-based ASCs are small and only applied to a few clusters, which implies a model that is appropriately sensitive. Travel time is again more important up to the first 30 minutes. Excess time beyond this threshold is perceived as being less onerous.

Attractions are driven by retail and service employment, which is appropriate for shopping/maintenance trips.

Estimated coefficients and t statistics

A zone’s accessibility to a hospital is a strong predictor in this model. This captures the important clustering of doctor’s offices, clinics, and pharmacies around major hospitals.

Travel time above 30 minutes has almost no impact on destination choice. This is an encouraging result. People traveling long distances to see a specialist are not impacted by travel time in a major way.

Attractions are driven by retail and service employment, with an added boost from service employment related to the presence of hospitals.

Estimated coefficients and t statistics

This model has a strong fit and is primarily based on K12 enrollment and travel time (as expected). Additionally, the negative coefficient on the Durham and Northeast clusters means that (all else equal) zones in that cluster are less likely to be chosen. The positive intrazonal coefficient is expected given the tendency for people to attend schools near their home.

Estimated coefficients and t statistics

The penalty on travel time above 30 minutes is reduced as seen in multiple trip types. The positive coefficient on walk accessibility means that these trips are more attracted to zones in areas that are more dense. Home cluster effects are strong, meaning that most of these trips stay within the home cluster. Employment, school enrollment, and residential population all are significant components of the size term. This reflects the catch-all nature of this trip type.

Estimated coefficients and t statistics

The attenuation on drive time above 20 minutes is not as strong as other trip types. This indicates that these trips are likely just trying to reach the nearest attraction that satisfies the trip purpose. Similarly, home cluster effects and the intrazonal term are strong meaning that people stay nearby. These effects are reasonable, reflecting that people are not inclined to travel longer distances for short activities.

Logsums for household auto, non-household auto, and transit all play an important role in determining where someone travels. Similar to the N_HB_OD_Long trip, the size term is made up of employment, K-12 enrollment, and household population.

Estimated coefficients and t statistics

Drive time has the expected negative sensitivity, with times above 30 minutes penalized to a lesser extent than the first 30 minutes. The generic intra-cluster effect is strong, which is expected given shorter trip lengths of NHB trips. Attractions are driven by a variety of employment as well as school enrollment.

Estimated coefficients and t statistics

The non-work auto trip model looks similar to work, but the inclusion of parking logsums captures a major difference between the two trip types: work trips are not impacted by parking variables while non-work trips are.

Estimated coefficients and t statistics

Both the total time of the trip and the number of transfers are significant predictors in the attractiveness of destinations accessible by public transit modes. Transit accessibility of the destination also positively impacts destination choice, which matches expectations.

Estimated coefficients and t statistics

As expected, walk distance poses a highly significant deterrent to destinations that are farther away from the origin. An extra penalty is imposed on destinations that are more than a mile away. This captures the sharp drop off in trip lengths beyond one mile in the survey.

Intra-zonal and intra-cluster effects are strong, which also reflects the short-range nature of non-motorized trips. Finally, the walk accessibility of the destination is important for zero vehicle households. This captures an important reality: these households cannot use autos to make trips.

W_HB_W

The W_HB_W purpose is unique because of shadow pricing employed in the model application. Regression models were estimated for the trip attractions that are then used to compute the shadow prices. Therefore, it is expected that the percentage of trips attracted to each cluster match between the model and the input attractions as shown below. The small values of the shadow prices however indicate that the original model without these shadow prices was also reasonably accurate.

The following tables show the cluster-to-cluster percentage comparisons between the model and the survey. For estimation, only the home to work leg of the trip was used in the dataset, whereas the model PA totals consist of both the forward and return trips (in PA format). Therefore, it is not possible to compare exact numbers but rather look at the percentages of flows from cluster-to-cluster.

The three tables below show the survey cluster-to-cluster percentages, the model percentages and the differences between the survey and the model respectively in a matrix format. Note that the last row in each table shows the percent of trips attracted to each cluster. There are differences for ‘w_hb_w’ because the shadow pricing matches the attractions derived from the attractions model as opposed to the survey numbers themselves. For a couple of the clusters, the survey attractions are different from the input attractions. For instance, the input attractions indicate that 5.51% of trips are attracted to the RTP cluster whereas the survey indicates that 7.84% of trips are attracted to the RTP cluster.

Despite the shadow pricing showing a deviation from the raw survey numbers, the cluster-to-cluster percentages between the model and survey match very well, and the differences are small. This is further illustrated in the chart, where each point represents a cluster pair. The survey percentages and model percentages from the tables below are used for the x and y axis respectively. Most of the points are close to the 45-degree line. The regression line is also plotted, with a slope of 0.96, an intercept of nearly 0 and a R-squared of 0.94, indicating an excellent match. Note that only points that were observed in the survey are plotted.

W_HB_O

Similar plots are shown for W_HB_O. This purpose has no shadow pricing and the match to the survey cluster-to-cluster flow is excellent.

W_HB_EK12

The model cluster-to-cluster flow patterns match the survey closely for the rest of the purposes and only the charts are shown. In each chart, each point represents one cluster pair.

N_HB_OME

N_HB_OMED

N_HB_K12

N_HB_OD_Short

N_HB_OD_Long

All the above graphs indicate that the cluster-to-cluster patterns revealed by the survey are adhered to in the model (with only a minimal set of adjustments), which in turn makes the model more robust and inspires confidence in the model’s predictive capability.