Estimation¶
Common design¶
- Sample. All survey-based models are estimated on the estimation half of the travel survey; the land-use models of residential location and car ownership on ACS PUMS households.
- Weights. The survey is stratified by home location, which every model except residential location takes as given, so these models are estimated without weights. Weights are used only where the sample is selected on the outcome: the PUMA level of residential location (whose sample is selected on origin and destination inside the zone-system area) and the work tour mode choice (whose sample over-represents rail commuters; weighted exogenous sample maximum likelihood).
- Location probabilities. A survey location known by traffic zone or tract may cover several model zones; its probability is the sum of the probabilities of those zones, weighted by housing units, enrolment or jobs.
- Choice sets. School and activity-location models use the full choice set of 1,588 zones; workplace choice the chosen zone and 99 sampled zones; the PUMA level of residential location all 97 PUMAs, and the dwelling level 30 sampled alternatives.
- Specification search. Each model is built in blocks of variables from a basic model. A block is kept when its likelihood-ratio test is significant at 5 %, it lowers the BIC and its signs are as expected; non-nested forms are compared by BIC.
- Standard errors are robust (sandwich).
- Hold-out. The hold-out half was consulted more than once in the specification search of several models, so its fit is not a fully independent test.
- Software: Python Biogeme for every model of the estimation package.
All estimated models¶
The full results of every model, coefficient by coefficient, are in Estimated models.
\(N\): observations; \(K\): parameters; LL: log-likelihood at the estimates; \(\bar\rho^2\): adjusted rho-square. Hold-out: log-likelihood per mover (residential location), \(\rho^2\) (car ownership), log-likelihood per record (workplace, schools, locations) or total log-likelihood (tour formation and mode) on held-out data.
| Model (alternatives) | \(N\) | \(K\) | LL | \(\bar\rho^2\) | Hold-out |
|---|---|---|---|---|---|
| Land use and workplace | |||||
| Residential location: PUMA (97)1 | 15,163 | 14 | −34,670.3 | 0.500 | −2.297 |
| Residential location: dwelling (30 sampled) | 15,163 | 18 | −47,478.6 | 0.079 | −3.115 |
| Car ownership (4) | 91,348 | 56 | −78,613.6 | 0.379 | 0.375 |
| Workplace (100 sampled) | 9,707 | 24 | −21,376.5 | 0.521 | −4.8112 |
| Schools (full choice set) | |||||
| K–8 (1,144 zones) | 1,509 | 13 | −2,885.6 | 0.721 | −1.953 |
| High school (719 zones) | 668 | 6 | −1,217.0 | 0.722 | −1.810 |
| College (141 zones) | 673 | 9 | −1,810.4 | 0.454 | −2.789 |
| MAST: locations (1,588 zones) | |||||
| Main activity, four purposes (pooled) | 14,341 | 62 | −49,427.4 | 0.532 | −3.370 |
| Stop: shopping | 3,979 | 23 | −9,982.6 | 0.659 | −2.527 |
| Stop: recreation | 5,064 | 20 | −15,874.9 | 0.574 | −3.205 |
| Stop: accompanying | 3,515 | 22 | −12,488.1 | 0.517 | −3.396 |
| Stop: other errands | 8,055 | 20 | −24,653.3 | 0.584 | −3.028 |
| Outing from the workplace: first stop | 1,180 | 10 | −3,112.3 | 0.663 | −2.707 |
| Work business: usual-workplace workers | 1,539 | 18 | −8,432.1 | 0.310 | −5.554 |
| Work business: other persons | 1,259 | 16 | −6,819.3 | 0.292 | −5.625 |
| Work trips of home workers | 366 | 5 | −1,640.9 | 0.430 | −4.389 |
| School without record; daycare (pooled) | 846 | 9 | −3,798.5 | 0.440 | −4.396 |
| MAST: tour formation and mode | |||||
| Outing from the workplace (2) | 8,613 | 8 | −3,061.5 | 0.486 | −3,093.2 |
| Mode of the outing (2) | 1,177 | 7 | −360.6 | 0.549 | −299.8 |
| Return home (2) | 28,349 | 33 | −13,697.0 | 0.301 | −13,771.8 |
| Tour mode: work (6)3 | 9,126 | 51 | −4,313.1 | 0.626 | −4,284.0 |
| Tour mode: school (6) | 2,651 | 34 | −972.5 | 0.651 | −1,096.5 |
| Tour mode: shopping (6) | 3,206 | 44 | −1,150.2 | 0.713 | −1,107.0 |
| Tour mode: other (6, nested) | 10,405 | 57 | −4,916.7 | 0.623 | −5,049.8 |
| Joint shopping (5) | 452 | 15 | −112.8 | 0.780 | −94.0 |
| Joint recreation and other (6) | 1,296 | 25 | −450.3 | 0.710 | −459.1 |
| Transit-tour legs (5)4 | 6,773 | 35 | −3,421.5 | 0.621 | −3,291.0 |
1. Weighted by the expected in-area share of origin and destination; inclusive value of the dwelling level with coefficient 1. 2. Full choice set of 1,587 zones with jobs. 3. Choice-based sample weights (weighted pseudo-log-likelihood). 4. Car driver available only with a household car that is free over the tour window.
The synthetic population (CVAE) is not a likelihood model; its fit is reported in validation. The activity distributions of MAST are 3,946 weighted relative frequencies of the estimation half.
Linking logsums¶
Three logsums connect the models:
- Mode-choice logsum of a person between two zones for a purpose, from the tour mode model; it enters the stop models as the logsum detour.
- Non-work accessibility of a zone, the destination-choice logsum over all 1,588 zones for shopping, recreation, accompanying and other trips, weighted by trip shares (0.329, 0.279, 0.148, 0.243). It enters car ownership as the accessibility gained from a car.
- Inclusive value of the dwelling level in the PUMA level of residential location.
Re-estimating the models yourself¶
Every model can be re-estimated from its data with Python Biogeme; see the estimation package.