The Promises and Pitfalls of OD Table Estimation from Ground Counts

Estimating origin-destination (OD) tables from ground counts can be controversial.  In many cases the number of variables of the estimation problem greatly exceeds the number of data points.  This means that additional information must be gotten from somewhere, and the quality of that information is often suspect.

QRS II contains many different methods for performing OD table estimation, all of which are dubbed “refinement” techniques.  QRS II is set up, naturally, to refine a gravity-model (or logit-model) created OD table — that is, one that has some underlying theoretical foundation.  So when doing a conventional refinement with QRS II, the aforementioned additional information comes from theory.  However, it is possible to bypass the gravity model entirely and feed QRS II’s refinement step any seed OD table that the heart desires.

I am quite comfortable with refining OD tables for short- to medium-range travel forecasts.  For example, a work-zone traffic management plan could be tested with a refined OD table, because there is only a short time span between the no-construction and construction cases.  There is little reason to believe that the factors causing errors in the gravity model would change appreciably over just several months.  Refining OD tables that are intended for long-range travel forecasts can be hazardous; the reason for errors today are unlikely to be the same as the reasons for errors twenty to thirty years hence.

A high quality seed OD table is a must, but I am unaware of any standards as to how good a seed OD table should be.  A seed OD table can come from a model, but it can also be gotten from ground data.  Recent innovations in vehicle re-identification technologies have opened up the possibility of inexpensive empirical OD tables for small areas and corridors.  NCHRP Report 765 illustrates how reasonable OD tables can be created for small areas by matching overall turning-movement percentages.  An OD table built from cell phone records is another potential source.

My favorite bad example is a OD table estimation program that allowed, even encouraged, users to input a seed OD table consisting of all 1’s.

The least-squares methods that I most prefer find a compromise between the seed OD table and ground counts.  Any bias in the compromise is controlled by the various weights of the estimation process.  QRS II provides just one weight for the OD table and a separate weight for each ground count station.  A thoughtful setting of the weights is critical to the quality of the estimates.  I almost always set the weights such that the differences between the model and the ground counts are no better than the amount of error in the ground counts themselves.  Oh yes, there can be quite a bit of error in any given ground count.  Thus, the estimation will adjust the seed OD only to the extent that can be justified by the quality (and consistency) of the ground counts.

Personally, I believe transparency is an important feature of any forecasting method I use.  All of the refinement methods in QRS II are explained in the Reference Manual and a bunch of journal articles I have authored.  It should not take long for a motivated individual to figure out what QRS II is doing.  However, my attempts to find clear documentation of methods in other software packages have often been frustrating.  Here is the rub:  different methods produce vastly different OD tables.  We cannot have full confidence in an OD table if we have no idea how it was calculated.  I find it interesting that people who would never take a prescription drug without learning of its side effects are willing to stake their professional reputations on computer software that is essentially undocumented.

Perhaps my biggest disappointment thus far is the slowness of the calculations on large problems.  I have waited days, even weeks, for very large refinements to finish.  I recently put in a lot of effort in Version 9 of QRS II to speed things up considerably.  One particular new option reduces precision slightly, but can speed up calculations by as much as a factor of 50.  However, there are still practical limits as to the size of an OD table that can be estimated, simply because of time constraints.

Considerable advice from me on OD table estimation from ground counts can be found in NCHRP Report 765.

Alan Horowitz, January 7, 2015