highway trip matrix development: current best practice · highway trip matrix development: current...

Highway Trip Matrix Development: Current Best Practice

Denvil Coombe SATURN USER GROUP

19 October 2011

Introduction

Usual assignment user classes:Cars on employers’ business trips (in work)Cars on other trips (non-work)TaxisLGVsHGVs

The focus in this presentation is on car trip matricesThis presentation is not a comprehensive specification: the focus is on the important or novel aspects and matters of conventional or well-established practice are omittedThe ideas and advice contained in this presentation have been developed over the last few years during work for Transport for London and other clients

19 October 2011 The DENVIL COOMBE Practice 2

Car Trip Matrices: Steps 1 to 5

1 Collection, editing and expansion of intercept survey data2 Collection, editing and reconciliation of count data3 Synthesis of matrix cell values in the non-interviewed

directions4 Creation of partial (‘observed’) trip matrices5 Analysis of the accuracy of the partial trip matrices at sector

level


Car Trip Matrices: Step 6

6a Assembly of synthesized trip ends6b Assembly of generalized cost matrices6c Assembly of trip cost distributions6d Trip matrix synthesis6e Assembly of external-to-external trip matrices


Car Trip Matrices: Steps 7 to 11

7 Factoring of daily or period trip matrices to assignment hours8 Adjustments to the prior trip matrices in the light of the prior

trip matrix tests (see next slide)9 Matrix estimation to ensure greater consistency of the trip

matrices with the count data10 Adjustments to the prior trip matrices if the changes brought

about by matrix estimation are regarded as significant11 Adjustments to the prior trip matrices in the light of the

journey time validations


Prior Matrix TestsStage Test Comparison Measure Criterion Acceptability

guideline Final model

Total assigned flows and counts across RSI, calibration and validation screenlines, by time period.

Flow differences

< 5% All or nearly all

Partial trip matrices

A Flows and counts of trips across RSI enclosures, for peak/inter-peak periods separately or 12 hours or 24 hours, depending on periods used for gravity model calibration and trip synthesis.

Flow differences


Synthetic trip matrices

B Flows and counts of trips across RSI enclosures, for peak/inter-peak periods separately or 12, 16 or 24 hours, depending on periods used for gravity model calibration and trip synthesis.

Flow differences

< 7.5% All or nearly all

Prior trip matrices

C Total assigned flows and total counts in both directions across RSI, calibration and validation screenlines, for each modelled hour.

Flow differences



Step 1: collection, editing and expansion of intercept survey data (1)

Enclosures (cordons) and screenlines must be ‘watertight’For screenlines, movements which could partially route around the ends of the screenlines should be omittedGaps on major roads because surveys were not feasible or permitted or abandoned should be treated as follows

Either use old RSI data re-expanded to new counts, providing that land-uses have not changed materially Or synthesize RSI trip records by means of Select Link Analysis using an existing modelFor both sources, low weights should be applied in the averagingprocess in the creation of the partial trip matrices (Step 4)



Flows on minor residential streets might form a material proportion of the total cordon or screenline flows but flows on individual roads might be too low to justify RSIsThe trip end estimates will be estimates of total trips and so the partial matrices should be comparable Therefore gaps on minor roads should be treated as follows

Group gap sites serving the same area and obtain the total flowExtract appropriate trip records from RSI sites on nearby roads and expand to the total gap flow – exclude trips with Os or Ds which are unlikely to be served by the gap roadsMerely expanding the trip records at the surveyed sites to the total flow at the surveyed sites and associated gaps would yield distorted trip patterns



Actual and demand flowsCounts are ‘actual’ flows (in SATURN terminology) and are less than demand flows at peak timesIf expansion factors are based on counts (actual flows) at timeswhen significant queues form upstream, the demand in the partialmatrices will be under-estimatedAs the partial matrices are used to control the synthetic trip matrices, the resulting ‘prior’ trip matrices may under-estimate demandThe solution is to build the partial trip matrices for periods which are sufficiently long for ‘actual’ and ‘demand’ flows to be equal


Step 2: collection, editing and reconciliation of count data (1)

Counts are required for:expanding samples of new roadside interviewsre-expanding samples of old roadside interviewscalibrating matrices by means of matrix estimationvalidating the model



General pointsThe traffic counts to be used for calibration and validation need to be specified at an early stage and should be of the same generalqualityCalibration screenlines need to focus on synthesized or unobserved movementsValidation screenlines need to be independent of RSIs and calibration screenlinesA set of rules should be established against which the suitability of any count may be judgedSingle-day MCCs are required for vehicle proportions but these should be indexed to ATCs conducted for two weeks to give more accurate totals – but note that vehicle proportions are still based on a single day’s count



Treatment of gapsFlows on all major roads must be countedFlows on all minor roads may be estimated based on some sample counts and a road typologyEither counts or estimates are required for ALL roads crossing cordons and screenlines



Accuracy of counts (95% confidence intervals)Automatic Traffic Counts (ATCs)

5%Manual Classified Counts (MCCs)

Cars: 10%LGVs: 24%HGVs: 28%

DMRB 12.2.1, paragraph 3.2.19, advises that counts with 95% confidence intervals wider than 15% should not normally be used in calibration or validation: grouping counts reduces the confidence intervals



Counts for expansion of RSI trip recordsInterviews should be expanded by vehicle type: this requires Manual Classified Counts (MCCs)Interviews should be expanded to the population from which they were sampled: this requires MCCs on the day that the interviews were conducted (DMRB 12.2.1, paragraph 3.2.21)A number of further adjustments are required:

to convert from the day of survey to an average weekdayto convert to a specific neutral monthto account for day to day variabilityto account for diversions on the day that the interviews were conducted

These adjustments require Automatic Traffic Counts (ATCs) for two-weeks, generally including the day that the interviews were conducted (DMRB 12.2.1, paragraph 3.2.20)



Counts for matrix estimationCounts at the RSI sites should be used as constraints to preventother constraints altering the assigned flows at these points to an unacceptable degreeBut the main purpose of matrix estimation is to refine the estimates made of movements which were not intercepted in the surveys: this means that counts are required on screenlines covering intra-sector movementsMatrix estimation should be applied to individual vehicle type matrices rather than to the total vehicle matrix (which would be too much of an approximation): this means that MCCs are required –but note that vehicle type counts are usually one-day countsTo enable matrix estimation to produce reasonably correct total numbers of trips, ATCs are also required



Counts for matrix estimation (continued)Counts of cars may sometimes need to be used at the individual link level, but they should be used mainly at the mini-screenlinelevelCounts of LGVs and HGVs must be used at the mini-screenline or full screenline levelCounts should, in principle, be weighted to reflect their accuracy

counts to which factors have been applied to adjust to the neutral month or to the model base year will be less accurate than counts taken in the neutral month in the model base year



Counts for re-expanding old RSI trip recordsDMRB advice on use of old data is not entirely clear

paragraph 2.10.2 in DMRB 12.2.1 says that where the original traffic data are more than 6 years old and comprehensive new data cannot be collected, a ‘present year validation’ is requirednothing appears to be said about re-use of old data by re-expansion

One approach would beto carry out two-week ATCs at the sites from which old data are to be used and re-expand to the new countsnot to conduct single day MCCs on grounds that the added value would be low



Counts for model validationDMRB 12.2.1 (paragraph 4.4.35) requires validation against counts which are independent of those used in matrix development and calibration and advises that counts with 95% confidence intervals greater than 15% should not normally be usedCounts for model validation need to be independent of

the counts used to expand the RSIsthe counts used as constraints in matrix estimation



Counts for model validation (continued)They need to be of the same general quality as the other counts

ATCs will enable validation of total flows on individual linksMCCs on their own (without ATCs) are not useful for individual link flow validation: MCCs need to be indexed (controlled) to two-week ATCsIndexed MCCs may enable validation of car flows on individual links but will not provide sufficiently accurate data for validation of LGV and HGV flows on individual links: validation of these vehicle types will have to be done at a mini-screenline level


Step 3: synthesis of matrix cell values in the non-interviewed directions

Common practiceMorning peak interview records

reverse the Os and Ds in the interview trip records for the morning peakadjust the purpose split to match the evening peak purpose splitcontrol to the evening peak count

Evening peak interview recordssimilar to the morning peak process

Inter-peak interview recordsreverse the Os and Ds and control to the count in the non-interviewed direction

For home-based trips, a better approach is to make use of return time probabilities obtained from a household survey


Step 4: creation of partial trip matrices (1)

The format of the partial trip matrices will need to accord withthe format employed in the trip synthesis (in Step 6)Current best practice is specified in the ERICA5 Manual although it may be preferable to develop and use other softwareThis method derives weighted averages where there is more than one estimate available for particular cell values, with theweights being based on indices of dispersion (normalised variances) Key section is 3 b viii in the ERICA5 Manual on combining different sources of error – however, this specification, as written, is incomplete and further information is required in order to implement the method


Step 4: creation of partial trip matrices (2)

Undertake prior trip matrix Test A

Investigate the causes of all exceedances before proceeding

Stage Test Comparison Measure Criterion Acceptability guideline

Partial trip matrices

A Flows and counts of trips across RSI enclosures, for peak/inter-peak periods separately or 12 hours or 24 hours, depending on periods used for gravity model calibration and trip synthesis.

Flow differences



Step 5: analysis of the accuracy of the partial trip matrices at sector level (1)

TAG Unit 3.10.2 advises“1.6.11 The process of “introducing observed data” must then make allowance for the statistical accuracy of that data, based essentially on sampling theory (see guidance in DMRB 12.1). Thiscould be done along the following lines. For each observed cell of the matrix…, the “prior” value would be tested as to whether it lay outside of the confidence region of the observed data: if so, the prior data will need to be modified.”At zonal level, the prior (synthesized) value will almost always lie within the 95% confidence intervals and so little or no use would be of the ‘observed’ data



TAG Unit 3.10.3 advises“1.5.21 However, by taking a weighted average of the observed andsynthetic matrices empty cells are eliminated and greater weight is given to cells where there are more observed trips than expected from the locally-calibrated synthetic model. Relative weights should reflect the relative accuracy of the two forms of estimates. If these are not known then a rough guide would be to use 90% of the observed estimate and 10% of the synthesised estimate.”Again, no mention of the level of spatial detail at which these processes should be carried outNo advice on how to determine weights which reflect the relative accuracy of the ‘observed’ and synthesised values (although this can be done)The 90%/10% advice means that the lumpiness of the ‘observed’ matrices (due to sampling variability) would be retained (undesirably)



Common practiceIt is not uncommon for analysts to use ‘observed’ trip matrices at zonal level in one of the three approaches advised in WebTAG

In the Unit 3.10.2 approach, this would lead to little or no use of the ‘observed’ trip estimatesIn the first method in Unit 3.10.3 which weights the estimates by reliability, greater weight would be attached to the ‘observed’ data than the synthesisedIn the second method in Unit 3.10.3, the 90%/10% method, little use would be made of the ‘synthesized’ data



Common practiceIt also not uncommon for analysts to compress the zonal level partial matrices to sector level before using them to ‘control’ the synthesized matricesUsually, the sector systems used are those defined by the RSI cordons and screenlinesIn this approach, typically

the majority of the sector level cell values will have 95% confidence intervals which are too wide to be useful because the numbers of trip records are too fewthe minority of sector level cell values will have satisfactory 95% confidence intervals, often based on more trip records than necessary



A better approachDefine a 95% confidence interval which would be regarded as acceptable – say, 20% to 30% of the cell valueRe-design the sectors with the aim that each sector level cell has at least sufficient trip records to exceed the 30% threshold and nomore than is necessary to achieve the 20% thresholdIf necessary, define movements (rather than sector level cell values) in terms of one sector to many others, many sectors to one sector, or many sector to many othersVary the sector and movement definitions by purpose, as necessaryUse the resulting estimates of trips making the defined sector level movements as constraints in the matrix synthesis (Step 6)


Step 6: synthesis of complete trip matrices (1)

Complete trip matrices need to be synthesized becausethe trip matrices derived from the RSI survey data are partial and estimates of the movements not intercepted in the surveys are requiredat the zonal level, the sampling variability of the ‘observed’ (partial) trip matrices will be very large and some means of smoothing outthat variability is required



Complete trip matrices may be synthesized usingeither a gravity modelor a destination choice model

The aims are the same, namely to calibrate model parameters so that

the partial matrix trip cost distributions are replicatedthe statistically reliable sector level movements in the partialmatrices are reproduced

The assumption is made that a model which meets these aims will provide satisfactory estimates of non-surveyed movementsThe calibrated distribution model is to be used only to synthesise base year matrices and is not intended to be used forforecasting future year demands



General principlesOnly inputs to the distribution model which are acceptably accurate should be used

The corollary is that use of inputs that have been developed by applying inaccurate factors should be avoided - this consideration applies particularly to the level of segmentation of the trip ends

The segmentation of the trip distribution model should be commensurate with the level of the segmentation of the inputs that can be derived with adequate accuracy

This means that a high level of segmentation should not be used in the distribution model if adequately accurate trip ends cannot be produced at that high level of segmentationThe distinctions in model parameters for the various segments in such a model may be spurious and at least partly due to errors in the trip ends



General principles (continued)The RSI data should be used in a statistically sound way

This means that the data should not be disaggregated by sector, time period and trip purpose to the point where the matrix cell values are not statistically reliable. Because spatial detail is important for the trip matrices in an assignment model, priority should be given to the detail of the sector system over the other two dimensions.

Use of inaccurate factors should be avoided in the derivation of the matrices for assignment from those created by the distribution model

Working in units of person trips in PA format at the 24-hour level will require factors to convert to vehicle trips in OD format for theassignment hours



Options for trip synthesisFor periods (morning peak, inter-peak, evening peak), or 12-, or 16- or 24-hour daysBy trip purpose, either the normal home-based employers’business, commute, education, shopping, and other and non-home-based employers’ business and other or some grouping of these segments (such as assignment user classes)In either person trips or vehicle tripsIn either OD or PA format

The choice depends only on the availability of reliable inputs



6a Assembly of synthesized trip ends6b Assembly of generalized cost matrices6c Assembly of trip cost distributions6d Trip matrix synthesis6e Assembly of external-to-external trip matrices


Step 6a: assembly of synthesized trip ends (1)

Options for car vehicle (driver) trip endsAn existing demand model

Input trip ends are not useful (usually total mode and singly constrained trip attractions will be modified by the demand model)Output trip ends have errors due to planning data, trip rates, trip end model, and demand modelUsually units are person trips in PA format at 24-hour level but these have to be converted to vehicle trips in OD format at period or hourly level for assignment

An existing highway assignment modelWill usually differ markedly from comparable trip ends output by a demand modelArguably more accurate than the demand model trip ends assuming that the highway assignment matrices have been adequately calibrated



Options for car vehicle (driver) trip ends (continued)NTEM

NTEM produces estimates of total trips across all modes and all times of dayNTEM splits trip ends between modes based on mode share data from the NTS in the 1990s for trip generations and from LATS 1991 for the trip attractions – the mode split parameters have limited spatial variation and there is a fixed mode split for each purpose for each person type (not cost dependent)



Options for car vehicle (driver) trip ends (continued)Least bad trip ends might be those from an existing calibrated highway assignment model, usually available in OD format for either peak periods or peak hours plus an average inter-peak hourIf NTEM trip ends have to be used, consider use of other data, such as Census JtW data, to adjust for variations in accessibility by mode



Options for car vehicle (driver) trip ends (continued)Purposes

An existing demand model and NTEM are likely to have a normal (detailed) trip purpose segmentationHighway assignment trip ends may be available for only a more limited set of user classes (often in-work and non-work)Using NTEM or demand model purpose information to disaggregate highway assignment trip ends is unlikely to yield accurate trip ends by purpose – better to stick with the user class segmentation



Options for car vehicle (driver) trip ends (continued)Periods

peak period trip ends might be available directly from the existing highway assignment matricesthere is no ambiguity about the costs that should be used for each periodthe use of factors to convert daily matrices to period matrices, which are necessarily approximate, is avoided



Options for car vehicle (driver) trip ends (continued)Purposes and periods

Trip ends by in-work and non-work user classes by period may well be the most reliable availableA gravity model calibrated for two demand segments has been shown to yield synthesized non-surveyed movements which are similar to those produced by a model calibrated for six demand segmentsTwo purposes and three periods requires the same number of gravity models as six purposes at the 24-hour levelUse of two demand segments means that movement constraints derived from the partial trip matrices can be defined at a greater degree of spatial detail



A least bad option for car vehicle (driver) trip endsFrom an existing calibrated highway assignment modelFor morning and evening peak periods and the inter-peak periodFor in-work and non-work purposes (user classes) only In OD formatBut they might need to be converted to the new model’s zoning system

less easy in OD format than in PA formatensure weights are as appropriate as possible



Check against partial matrix trip endsPartial matrix trip ends inside a cordon should be plausibly less than the total trip ends from the existing HAMFactor intra-cordon trip ends if necessary – factors will be a matter of judgment but develop rules for systematic application



In summaryAvoid the temptation to manipulate the available trip ends – the application of every factor adds further error and some factors only lead to results of spurious accuracyTake account of the availability of generalized costs – use of costs which relate clearly to specific hours or periods with period trip ends is more rigorous than use of peak and inter-peak costs (in some way) with trip ends for 12, 16 or 24 hour periods


Step 6b: assembly of generalized cost matrices

Generalized costsRelate to the trip ends as directly as possibleInclude time and all money costs – fuel, non-fuel (for in-work trips), congestion charges, toll charges and parking chargesDerive fuel and non-fuel costs from TAG Unit 3.5.6 for new model base month and year and use network average speeds (from an existing model or survey) for operating cost calculationsConvert money costs to time units using up-to-date values of time from TAG Unit 3.5.6 for new model base month and yearSkim from either an existing model or an initial network for thenew modelIf available, surveyed times, such as those from TrafficMaster, are likely to be better than modelled times


Step 6c: assembly of trip cost distributions

Trip cost distributions (TCDs)Are required for the calibration of the distribution modelsShould be derived by demand segment and time period from the partial trip matrices for the basic calibrationShould also be derived for enclosures (cordons) and screenlines for refining the calibration


Step 6d: trip matrix synthesis (1)

General principlesUse either a gravity model or a destination choice model, whichever is computationally convenientCalibrate the deterrence function to

first, match the partial matrix TCDssecondly, match the TCDs for the individual cordons and screenlines

Use an algorithm to find the parameter values whichminimise the sum of the squares of the differences by cost bin between the ‘observed’ and synthesised TCDsmatch the mean trip cost and the standard deviation of the TCD

Use a 3D Furness to constrain the synthesised matrices toorigin (or production) and destination (or attraction) totalsstatistically reliable sector level movements from Step 5



Gravity modelThe basic formulation is:

where is the number of trips between zone i and zone jis the number of origins in zone iis the number of destinations in zone jis a function of the generalized cost of travel between zone i and zone j, the ‘deterrence function’

The deterrence function usually takes the form:

)(.. ijjiij CfnDOT =

ijTiOjD

)( ijCfn

ijCijij eCCfn αβ *)( =



Gravity model (continued)The fit to TCDs may be improved by modifying the costs using a distance-based cost damping function (TAG Unit 3.10.2)Log normal formulation has also been shown to fit TCDs wellThe fit for long-distance trips needs particular attention

Typically, long-distance trips will be less than 5% of the total trips but could contribute about a third to the total trip-kmsThese trips will display high fuel cost elasticities in the demand model calibration, unless costs are damped, and will disproportionately influence the overall elasticity



Undertake prior trip matrix Test B, in two formsA version constrained to Os and Ds (2D Furness) will show how well the deterrence function is able to reproduce surveyed volumes at cordon and screenline levelA version additionally constrained to the Step 5 movements should show very similar results to Test A

Check all exceedances before proceeding



Synthetic trip matrices

B Flows and counts of trips across RSI enclosures, for peak/inter-peak periods separately or 12, 16 or 24 hours, depending on periods used for gravity model calibration and trip synthesis.

Flow differences


Step 6e: assembly of external-to-external trip matrices

External-to-external tripsWill be difficult to synthesize with any accuracyMay be better omitted from the trip synthesis and sourced from elsewhere


Step 7:factoring of daily or period trip matrices to assignment hours (1)

Peak hour factorsAre best derived from partial matrices constructed for the periods used in the trip synthesis and for the peak hoursRatios of peak to period trips should be derived using the movement definitions derived in Step 5Estimate initial values of peak hour factors for non-surveyed movements based on the factors for the surveyed movements


Step 8: adjustments to the prior trip matrices in the light of the prior trip matrix tests (1)

Apply peak hour factors to car trip matrices, add in matrices oftaxi, LGV and HGV trips, assign, and undertake prior trip matrix Test C

If acceptability guideline is not met, consider amendments totrip endsdistribution model parameterspeak hour factors


Prior trip matrices

C Total assigned flows and total counts in both directions across RSI, calibration and validation screenlines, for each modelled hour.

Flow differences



Step 8: adjustments to the prior trip matrices in the light of the prior trip matrix tests (2)

Test C compares ‘actual’ observed flows (counts) with ‘actual’assigned flows (for the assignment hours)Demand flows (for the assignment hours) should also be tabulatedThe general scale of the difference between the ‘actual’ and demand flows should be plausible, that is, likely to yield realistic levels of queueing in the networkAt this stage, the plausibility of the difference between ‘actual’and demand can only be a matter of judgement – some evidence will emerge from the journey time validations (see Step 11)


Step 9: matrix estimation to ensure greater consistency of the trip matrices with the count data (1)

DMRB 12.2.1 advises that:matrix estimation should not be used if the differences between the count data and the modelled flows (resulting from the assignmentof the prior trip matrix) are within the survey accuracies (paragraph 4.3.34)if matrix estimation causes “significant” changes in the prior matrix, the prior matrix should be reconsidered (paragraph 4.3.41)

The implication is that matrix estimation should not be used unless the confidence intervals of the counts used as constraints are very small



Counts need to be grouped to form mini-screenlines for two reasonsA Applying groups of counts as constraints should avoid matrix

changes being made which merely compensate for network and routeing errors

B Individual counts are unlikely to have sufficiently small 95% confidence intervals and grouping counts will yield smaller confidence intervals



Some tensionsAs the network and matrices are refined, the modelled routes should become more reliable and therefore, in principle, the mini-screenlines could be shortenedHowever, shortening the mini-screenlines means that the confidence intervals will increase, whereas, as the network and prior trip matrices are improved, it is to be expected that the changes required to be brought about by matrix estimation will be smaller, which means that confidence intervals should decrease which means that mini-screenlines should lengthen



My current adviceGroup counts together as much as possible without mixing together roads serving very different areasAvoid as far as possible using individual counts as constraints unless they are unusually accurate or account can be taken of their relative accuracy in the matrix estimation process

The danger is that matrix estimation distorts the prior trip matrix to match a set of counts which vary randomly to a material degree about their true values



In principle, weights could be applied to reflect the accuracy or reliability of the inputs

Prior trip matrices – greater weight could be attached to those movements which have been constrained to the sector level partial matrix movements from Step 5Traffic counts Trip ends, although it seems unlikely that these will be sufficiently accurate to be used as constraints at all



Validation and monitoringAs matrix estimation is designed to refine the trip matrices, of the DMRB validation tests, only the screenline flow comparisons are relevantThe changes brought about by matrix estimation should be monitored in the following ways

scatter plots of matrix zonal cell values, prior to and post matrix estimation, with regression statistics (slopes, intercepts and R squared values)scatter plots of zonal trip ends, prior to and post matrix estimation, with regression statistics (slopes, intercepts and R squared values)trip length distributions, prior to and post matrix estimation, with means and standard deviationssector to sector level matrices, prior to and post matrix estimation, with absolute and percentage changes


Step 10: adjustments to the prior trip matrices if the changes brought about by matrix estimation are regarded as significant

If the changes brought about by matrix estimation are regarded as significant, consider further amendments to the prior trip matrices by amending

trip endsdistribution model parameterspeak hour factors


Step 11: adjustments to the prior trip matrices in the light of the journey time validations

Journey time validations, in the form of time/distance graphs, will show the following

If the slopes of the modelled and observed lines are parallel, this will indicate the cruise speed element of the model is about rightIf the modelled and observed lines show material jumps in time at the same distances, this will indicate that delays are being modelled in the right places to the correct magnitudeIf the above two conditions are met, the modelled and observed lines should start and end at the same places and that the general level of demand in the trip matrices is about rightIf the first condition is met but the second and therefore the third are not, some adjustments to the level of demand in the trip matrices should be considered


Concluding Remarks

The three most important inputs to the development of highway assignment trip matrices are

Traffic counts – high quality counts are required for trip matrix development, calibration and validationIntercept survey data – the synthesis of the complete trip matrices will be based on these data, so high quality surveys and good coverage are very desirableTrip ends – unless a demand model exists, recourse may need to be made to the national data and/or models - avoid the temptation to over-manipulate the available trip ends – the application of every factor adds further error and some factors only lead to results of spurious accuracy


highway trip matrix development: current best practice · highway trip matrix development: current...

Documents