Presentation is loading. Please wait.

Presentation is loading. Please wait.

Effect on Model Sensitivities of Combining Transferable Data from Separate Home Interview Surveys Presented to the 11 th Conference on Transportation Planning.

Similar presentations


Presentation on theme: "Effect on Model Sensitivities of Combining Transferable Data from Separate Home Interview Surveys Presented to the 11 th Conference on Transportation Planning."— Presentation transcript:

1 Effect on Model Sensitivities of Combining Transferable Data from Separate Home Interview Surveys Presented to the 11 th Conference on Transportation Planning Applications May 8, 2007 By Jonathan Avner, Wilbur Smith Associates Gregory Giaimo, Ohio Department of Transportation

2 Outline Analysis of Data Transferability Analysis of Data Transferability –Production Rates –Trip Length Analysis –Time of Day Analysis Sensitivity of Models with Transferred Data Sensitivity of Models with Transferred Data

3 Background ODOT undertook a household survey data collection effort in 2000 to support the development of a new generation of travel demand models in the small and medium sized MPOs. ODOT undertook a household survey data collection effort in 2000 to support the development of a new generation of travel demand models in the small and medium sized MPOs. In total, over sixteen thousand households were surveyed (MPO and non MPO areas) that included more than 100,000 trip records. In total, over sixteen thousand households were surveyed (MPO and non MPO areas) that included more than 100,000 trip records.

4

5 Survey Household Locations

6 Data Transferability

7 Previous research has focused on feasibility of avoiding surveying by borrowing other data. Previous research has focused on feasibility of avoiding surveying by borrowing other data. This research focused on combining data to obtain improved parameter estimates. This research focused on combining data to obtain improved parameter estimates. Each area had 1300 to 1900 households surveyed and would be getting the same model design with calibrated parameters. Each area had 1300 to 1900 households surveyed and would be getting the same model design with calibrated parameters. Considered following model components Considered following model components –Trip Production Rates –Trip Distribution (Friction Factor Calibration) –Time of Day

8 Areas Considered for Combination MPO Sm all Larg e Group 1 Group 2 ToledoX LimaXX DaytonX SpringfieldXX AkronXX CantonXX MansfieldXX SteubenvilleX YoungstownxX

9 Trip Production Rate Analysis - Purpose Determine whether datasets could be combined to create larger estimation datasets for better parameter estimation. Determine whether datasets could be combined to create larger estimation datasets for better parameter estimation. Depending on purpose, trip rates are stratified by wealth (vehicles / hh), size (hh size, hh workers, etc.) and possibly area type. Depending on purpose, trip rates are stratified by wealth (vehicles / hh), size (hh size, hh workers, etc.) and possibly area type. With combined datasets able to achieve minimum number of observations per cell with area type stratification (not necessarily without). With combined datasets able to achieve minimum number of observations per cell with area type stratification (not necessarily without). Thus if area type dimension needed, combining study area datasets could be necessary. Thus if area type dimension needed, combining study area datasets could be necessary.

10 Trip Production Rate Analysis – Statistical Analysis The mean trip production rate was compared on a cellular basis for each combination (small, large, group 1, group 2). The mean trip production rate was compared on a cellular basis for each combination (small, large, group 1, group 2). ANOVA (analysis of variance) was used since greater than two samples were being considered. ANOVA (analysis of variance) was used since greater than two samples were being considered. Results were based on looking at F statistic: Results were based on looking at F statistic: –Ratio between the group variability and within group variability –Value close to 1 → accept H o (means are equal) –Value much >1 → reject H o (means are not equal)

11 Trip Production Rate Analysis – Area Type For the larger MPOs, the trip rates were compared between area types to determine need for this dimension. For the larger MPOs, the trip rates were compared between area types to determine need for this dimension. Four area types are used in generation: CBD, Urban, Suburban, Rural Four area types are used in generation: CBD, Urban, Suburban, Rural Average trip rates between area types in the large MPOs were tested using ANOVA. Average trip rates between area types in the large MPOs were tested using ANOVA.

12 Trip Production Rate Analysis – Area Type High F statistics indicated difference between average trips between different area types. High F statistics indicated difference between average trips between different area types. Unique production rates were calibrated for area types or combinations of area types when: Unique production rates were calibrated for area types or combinations of area types when: –F statistic was large between area types; and –Sample size large enough in each cell Households per cell>30 Households per cell>30

13 Trip Production Rate Analysis - Results Chosen combination of study area data would be applied to all trip purposes in the trip generation model Chosen combination of study area data would be applied to all trip purposes in the trip generation model Necessary to develop overall “score” for each combination, since actual ANOVA at a cellular level Necessary to develop overall “score” for each combination, since actual ANOVA at a cellular level –Households in each cell of combination were added together if cell had significant F statistic (accept H o ) –Results below indicate percentage of households that are in cells with similar trip rates. HBW HBS H HBONHWNHOAvgSmall61%29%50%26%61%45% Large72%68%70%69%70%70% Grp 1 68%58%57%89%59%66% Grp 2 51%67%60%66%71%63%

14 Trip Production Rate Analysis - Recommendations AreaCombine ToledoLarge Lima Group 1 Dayton Springfield AkronLarge Canton Mansfield Steubenville Youngstown Group 2 – removed because of overlap with Large Group 2 – removed because of overlap with Large Dayton removed because of independent model development Dayton removed because of independent model development

15 Trip Length Analysis - Purpose Intent of analysis was to find areas where a friction factor curve could be shared between areas. Intent of analysis was to find areas where a friction factor curve could be shared between areas. Same combination datasets were considered: small, large, group 1 and group 2 Same combination datasets were considered: small, large, group 1 and group 2

16 Trip Length Analysis – Statistical Analysis Trips used in analysis were restricted to those with both trip ends within an MPO area and with known locations of trip ends. Trips used in analysis were restricted to those with both trip ends within an MPO area and with known locations of trip ends. Rather than using reported trip length, the skimmed trip length was used in the analysis. Rather than using reported trip length, the skimmed trip length was used in the analysis. ANOVA was used to compare average trip length. ANOVA was used to compare average trip length. Trips were compared two ways Trips were compared two ways –Same trip purpose across areas –Purposes within an area to see if differences existed

17 Trip Length Analysis by Purpose - Results Results indicate that there is significant difference between average trip length between areas in combination datasets. Results indicate that there is significant difference between average trip length between areas in combination datasets. Logical findings given different network characteristics, geographic size of area and other travel related factors. Logical findings given different network characteristics, geographic size of area and other travel related factors. LargeSmall Group 1 Group 2 F-StatSigF-StatSigF-StatSigF-StatSig HBW 18.470.00124.810.0053.090.0036.760.00 HBSH 24.180.0044.920.0010.620.003.780.02 HBO 60.350.0063.520.008.210.0073.870.00 NHBW 12.550.0038.420.0010.540.007.870.00 NHBO 12.770.0034.880.0018.370.0011.520.00

18 Average Trip Length by MPO Area

19 Trip Length Frequency Distribution - HBW

20 Trip Length Analysis by Area - Results Results for comparison of purposes within an MPO area showed little potential for combination. Results for comparison of purposes within an MPO area showed little potential for combination. Consistent with traditional approaches to have unique gravity model for each trip purpose. Consistent with traditional approaches to have unique gravity model for each trip purpose. Toledo HBWHBSHHBONHBWNHBO FSigFSigFSigFSigFSig HBW537.3900.000741.2740.000118.0860.000646.8690.000 HBSH30.4890.00087.8830.0004.7430.029 HBO34.6330.00016.6570.000 NHBW65.6350.000 NHBO

21 Average Trip Length by Purpose by Area

22 Time of Day - Purpose Determine whether datasets could be combined for estimation of time of day factors and directional factors for Time of Day model. Determine whether datasets could be combined for estimation of time of day factors and directional factors for Time of Day model. Coincidence Ratio was used to determine if all areas shared similar daily distribution of trips. Coincidence Ratio was used to determine if all areas shared similar daily distribution of trips. Four time periods were defined: Four time periods were defined: –Over Night (6pm to 6am) –AM Peak Period (6am to 9am) –Midday (9am to 2pm) –PM Peak Period (2pm to 6pm)

23 Time of Day – Statistical Analysis Difference of proportions test was used to compare the proportion of trips made between each area being compared: Difference of proportions test was used to compare the proportion of trips made between each area being compared: –Small with Large –Small only –Large only –Group 1 and Group 2

24 Time of Day - Results From a cursory inspection, it seems all areas could share the same dataset. From a cursory inspection, it seems all areas could share the same dataset. Further review of the results indicates that for HBSH (Period 1), HBW (Period 2), HBO (Period 3) and HBSH and HBO (Period 4) there are significant differences between the small and large datasets. Further review of the results indicates that for HBSH (Period 1), HBW (Period 2), HBO (Period 3) and HBSH and HBO (Period 4) there are significant differences between the small and large datasets. Since all MPOs are included as either small or large, this was the recommended dataset for TOD calibration. Since all MPOs are included as either small or large, this was the recommended dataset for TOD calibration.

25 Time of Day - Results HBWHBSHHBONHBWNHBOSmallLargeSmallLargeSmallLargeSmallLargeSmallLarge Per 1 18.719.932.027.324.123.97.07.617.618.6 Per 2 35.932.83.74.223.223.317.917.49.68.6 Per 3 12.913.033.733.119.418.337.638.837.537.8 Per 4 32.534.330.635.433.334.537.536.135.335.1 HBW – Percent of Trips by PeriodHBSH – Percent of Trips by Period Percent Departure by Period (Shaded = Statistically Different)

26 Additional Analysis Reviewed cell compression scheme suggested by ODOT. Reviewed cell compression scheme suggested by ODOT. –Cells compressed based on rarity of households in survey –Cells with more vehicles than persons were compressed (based on analysis of OKI, MORPC and NOACA survey) Evaluation of compression based on: Evaluation of compression based on: –Number of households in each cell from survey dataset –Difference in trip rate between independent cells and compressed cells Analysis supported ODOT compression techniques. Analysis supported ODOT compression techniques. 0 Wrk 1 Wrk 2 Wrk 3 Wrk 0 Veh 1 Veh 2 Veh 3 Veh 1 HH 2 HH 3 HH 4 HH 0 Veh 1 Veh 2 Veh 3 Veh

27 Additional Analysis Evaluated the potential of a HB School trip purpose. Evaluated the potential of a HB School trip purpose. –Compared Average trip length for school and non school HB activities –Evaluated frequency of trips for sufficient numbers for calibration. –Evaluated distribution of households in cross classification matrix (vehicle ownership x students in household) Determined that a HB School purpose was warranted Determined that a HB School purpose was warrantedMPO Survey HH HH w/Sch Trip TOL2176597 LIM1328302 DAY1950521 SPR1349394 AKR1936559 CAN1319351 MAN1304332 MAN1276249 YOU1251324

28 Trip Rate Sensitivity Analysis Further pursue the impact that various trip generation rates would have on model results Further pursue the impact that various trip generation rates would have on model results Calculate various “feasible” sets of trip rates based on the combined and Toledo stand-alone survey data sets Calculate various “feasible” sets of trip rates based on the combined and Toledo stand-alone survey data sets Smaller sample size in the stand alone data implies a broader range of “feasible” trip rate sets Smaller sample size in the stand alone data implies a broader range of “feasible” trip rate sets

29 Total Households Comparison

30 Trip Rates

31

32

33 Construction of Alternate Trip Rates Calculate Percent Errors for a given confidence interval (rather arbitrarily selected 90%) Calculate Percent Errors for a given confidence interval (rather arbitrarily selected 90%) E = Z*CV/SQRT(N) E = Z*CV/SQRT(N) Develop other feasible sets of trip rates within plus / minus this error percentage of the calculated mean Develop other feasible sets of trip rates within plus / minus this error percentage of the calculated mean

34

35

36 Construction of Alternate Trip Rates Trip rates varied by cross-class cell, however, the overall resultant trip rates were also held within the 90% confidence interval Trip rates varied by cross-class cell, however, the overall resultant trip rates were also held within the 90% confidence interval Various perturbations of the trip rates were created within this range, the two shown are: Various perturbations of the trip rates were created within this range, the two shown are: –Systematic perturbation involving increasing zero Vehicle HH trip rates by exactly the calculated percent error while reducing all other trip rates by 10% of this value –Random perturbation of each trip rate within its percent error range

37 HBW Proportion of the Percent Error Applied to Create Alternate Trip Rate

38 Alternate Reality Socio-Economic Data Given the concentration of variance in certain rare cells of the cross classificatin matrix… Given the concentration of variance in certain rare cells of the cross classificatin matrix… An alternative set of zonal SE data was constructed that placed more HH’s in these cells by: An alternative set of zonal SE data was constructed that placed more HH’s in these cells by: –Reducing Vehicles by 50% in CBD / Urban Area –Increase Workers 16% in all zones –No change in # of HH’s or attraction variables

39

40

41 Test Impact on Measures of Effectiveness (MOE’s) 12 Test Cases Based Upon: 12 Test Cases Based Upon: 6 Sets of Trip Rates 6 Sets of Trip Rates 1.Combined Data, Base 2.Combined Data, Systematic Perturbation 3.Combined Data, Random Perturbation 4.Toledo Data, Base 5.Toledo Data, Systematic Perturbation 6.Toledo Data, Random Perturbation 2 Sets of SE Data 2 Sets of SE Data 1.Base 2.Modified

42 Test Impact on Measures of Effectiveness (MOE’s) Evaluate Various MOE’s: Evaluate Various MOE’s: 1.Link Volume 2.VMT 3.VHT 4.%RMSE or %RMSD 5.Tons of Pollutants 6.Trips 7.Transit Riders

43 Base Model & SE Data Toledo Data Only, Systematic Perturbation, Modified SE Data Volume on New River Crossing

44 VMT

45 VHT

46 %RMS Error and Difference

47 Ozone Precursors

48 Trips

49 Conclusions Randomly perturbed trip rates, even when applied to purposefully skewed SE data showed almost no impact on typical MOE’s Randomly perturbed trip rates, even when applied to purposefully skewed SE data showed almost no impact on typical MOE’s Systematically perturbed trip rates produced slightly lower %RMSD between the SE data scenarios Systematically perturbed trip rates produced slightly lower %RMSD between the SE data scenarios –Base %RMSD: 10.43 –Combined:8.36 –Stand Alone:7.09 These slight differences are minor compared to the models %RMSE values These slight differences are minor compared to the models %RMSE values

50 Conclusions The Toledo stand alone sample was sufficient for the given model (not surprising since it was designed as such) The Toledo stand alone sample was sufficient for the given model (not surprising since it was designed as such) Increasing sample size much beyond the computed minimums wouldn’t have added much Increasing sample size much beyond the computed minimums wouldn’t have added much It was still useful to combine the data sets where practical to give more faith in the low incidence cells It was still useful to combine the data sets where practical to give more faith in the low incidence cells This also allowed the addition of the area type dimension to the smaller areas whose smaller survey sample was not originally designed for this This also allowed the addition of the area type dimension to the smaller areas whose smaller survey sample was not originally designed for this

51 Questions? Please use the Microphone.


Download ppt "Effect on Model Sensitivities of Combining Transferable Data from Separate Home Interview Surveys Presented to the 11 th Conference on Transportation Planning."

Similar presentations


Ads by Google