Forecasting Demand For a New Mode of Transportation



Citation:

J. D. Davidson (1972) ,"Forecasting Demand For a New Mode of Transportation", in SV - Proceedings of the Third Annual Conference of the Association for Consumer Research, eds. M. Venkatesan, Chicago, IL : Association for Consumer Research, Pages: 294-303.

Proceedings of the Third Annual Conference of the Association for Consumer Research, 1972      Pages 294-303

FORECASTING DEMAND FOR A NEW MODE OF TRANSPORTATION

J. D. Davidson, Senior Operational Research Designer, Air Canada

INTRODUCTION

Starting in mid-73, a subsidiary of Air Canada will be operating a STOL (short take-off and landing) service between downtown Montreal and downtown Ottawa. The service is an experiment sponsored by the Canadian Government using a Canadian built aircraft--the Twin Otter. The O.R. group was asked to help in the marketing of STOL by developing a model for forecasting traffic and predicting the effect of changing the levels of service.

This experiment will offer a fifth mode of travel between the two cities: the existing modes are car, bus, train and CTOL (Conventional take-off and landing). It is difficult to use historical data from the four existing modes to predict traffic on STOL even if it existed because of the fact that STOL will offer a service outside existing experience. In particular, it will offer a shorter door-to-door time than any existing mode and will probably cost more.

Data was gathered by a consumer research company, Market Facts of Canada Ltd, which was commissioned to make a random telephone survey of 21,000 households followed up by 1,055 selected personal interviews. After we had started work on model design we found that Market Facts' parent company in Chicago had already developed two models which would meet our requirements very nicely. We therefore made use of these two models. The method uses attitudinal or "soft" data on such things as meal service and comfort as well as "hard" data on such things as time and price.

One of the major constraints was that the model had to be portable i.e. STOL could be sold as a package (aircraft, avionics, operating and marketing know-how) anywhere in the world. In particular, all the demographic and travel statistics which are available in Canada could not be used as they might not be available for another city pair elsewhere. They were, of course, used as a cross-check, sometimes to our chagrin. The acceptability of the sample is discussed later.

The project was carried out in five stages:

1. Qualitative survey to find out why travellers choose or reject each mode of travel and to judge their reaction to the proposed STOL service.

2. Model design-to build a model which will incorporate attitudinal data and predict the response to the new STOL service.

3. Data gathering by means of a large number of random telephone calls to households in Montreal and Ottawa followed up by personal interviews of a selected sample.

4. Model testing and calibration--tuning the model so that it can reproduce the existing modal split.

5. Forecasting and evaluation--adding STOL to the existing four modes and evaluating the effect of different marketing strategies on the STOL market share.

RESEARCH SURVEYS

We started with qualitative research to determine consumer attitudes to the proposed STOL service. This was done by group depth interviews. We then took the groups for a ride in the Twin Otter and followed up with further group interviews. The change in perceptions was remarkable. As a result of all the interviews we had a lot of qualitative information on how people perceived the advantages and disadvantages of each of five modes of travel. We also gained some knowledge of which attributes are important to travellers and which are regarded as trivial.

The qualitative survey was then followed by a quantitative survey-a random telephone survey of 15,000 households in Montreal and 6,000 in Ottawa. The number of phone calls was chosen to give a statistically acceptable sample of travellers in each city. That gave us demographic and travel information: number of trips made in the last 12 months, mode used on each trip and reason for the trip (business or non-business). The only other data we needed was the population over 18 of the two cities: Ottawa 370,000; Montreal 1.8 million. One interesting fact which emerged: half the population of Ottawa has visited Montreal in the last 12 months. We cannot think of another city pair where half the population of one has visited the other in a 12 month period.

THE INTERVIEW STAGE

As a result of the qualitative research, it was found that all five modes could be described in terms of 13 independent attributes. Each attribute was defined in terms of levels e.g. the attribute "time" had three levels-1 hour, 2 hours and 3 hours. Some attributes had two levels and some had four but the average was three. Hence we had 39 points: 13 attributes X 3 levels = 39.

TABLE

ATTRIBUTES USED TO DEFINE MODES OF TRAVEL

The purpose of the interview was to obtain relative utilities for the 39 points. Used in this sense, a utility is simply a number between 0 and 1. It is further assumed that the pairwise products of utilities reflect a person's trade-off preferences. The interview, which took over an hour, required each respondent to fill in 21 matrices, ranking his preference for pairs of attributes. Suppose we knew what a person's utilities for time and cost were. We could then obtain utilities for pair combinations by multiplication. For example: U(1 hour) x U($8) - .7 x .25 - .175. This assumes that the attributes are independent, something we tried to ensure when deriving them.

MATRIX FOR PAIRWISE COMPARISON OF ATTRIBUTES

Having calculated utilities for each of the twelve combinations in the above matrix we could then rank them from 1 to 12 according to the rank order of the utilities.

In practice, of course, we can't expect the respondent to tell us what his utilities are but it is a simple task for him to rank order the cells in the above matrix according to his preferences. He was asked to do so without thinking of any particular mode. Once we have this rank ordered preferences we can infer his utilities using a technique known as pairwise non-metric multidimensional scaling. The particular computational technique which was used is similar to techniques for non-metric multidimensional scaling which have been developed by Kruskal (1964a, 1964b, 1968)and Johnson (to be published). A recent application may be found in Gibson et al. (1972). Mathematically, the problem is to find utilities for 39 attribute levels such that the pairwise products of utilities have rank orders similar to those provided by the respondent. In the iterative procedure, an initial solution is tested and the utilities modified so that the pairwise products are more like the respondent's rank ordered preferences. The programme seeks to minimize a value called "badness of fit." Iterations will continue in an attempt to reduce badness of fit until:

1 the badness is reduced to an acceptable level

or 2 the badness fails to improve significantly in 5 iterations

or 3 a specified number of iterations is reached

The programme which converts pairwise rank ordered preferences to utilities was developed by Johnson from the procedure described by Kruskal (1968) and is the property of Market Facts Inc. This is the first of the two models. This is an example of conjoint measurement--a technique which allows us to estimate a consumer's value system by observing his behaviour. The way he trades off one attribute against another tells us about his values for both attributes jointly; it might not be possible to measure his values by taking the attributes one at a time. Twenty one attribute pairs were ranked--each attribute was compared with a least two others and important attributes like time and cost were compared with about five others.

THE FORECASTING MODEL

This is the second model. The 1,055 respondents were divided into classes: 4 geographical areas; business/non-business; light/heavy travellers; most often by air/most often by car/most often by other surface mode. This was done so that we could weigh each respondent correctly in order to relate respondent behaviour to the results of the telephone survey. Although the telephone survey was a random sample, the interviewees were not. The sample was structured to obtain equal representation in each class.

To calibrate the model, we had to input the attributes of the four existing modes and ensure that the model produced the correct modal split.

DEFINITION OF EXISTING MODES

Level 1.5 for the trip time by car implies that the time taken lies half way between level 1 and level 2 i.e. 2 1/2 hours. Level 4 for the cost of a car trip implies that the door-to-door cost of a one way trip is $4 per head. Level of 1.9 for the trip time by air represents 2 hours 6 minutes.

This process was repeated for all 13 attributes. Although we had calculated respondents' utilities for levels 1, 2, 3, etc. for each attribute, it was possible to interpolate between these levels and even extrapolate beyond them. When designing the questionnaire, we had tried to choose a range for each attribute which would encompass all possibilities to avoid having to extrapolate. The model assumed straight line interpolation between points. The diagram shows how extrapolation was done.

FIGURE

A respondent's relative likelihood of taking each mode was calculated as follows:

UCar = KC(Uc1Uc2 ... UC13)a

UBus = KB(UB1UB2 ... UB13)a

UTrain = KT(UT1UT2 ... UT13)a

UAir = K"(UA1UA2 ... UA13)a

Here we have adopted the convention that U represents the respondent's utility for the level of attribute 1 for car i.e. his utility for level 1.5 for time. The respondent's "utility" for each mode was calculated by multiplying together the 13 utilities for levels of each attribute. The constant K. is known as the mode effect and is used to make a correction for any unexplained variable. We found that KC=KB=KT=1 but that the mode effect for air was considerably less than 1. The value of KA was determined by trial and error and was required because otherwise the model consistently over-predicted the number of air travellers. Only a fraction of those who would find air attractive actually took the plane. Perhaps the 13 attributes do not completely define each model we think there may be a 14th attribute. Ki was set at zero if a respondent answered no to the question "Would you ever travel to Ottawa by air? (bus? train? car?)."

The value of the exponent a was also determined empirically and came out at .5. There are two major reasons why we expected the exponent to be less than 1. Although we tried to devise 13 independent attributes there is inevitably some redundancy and the exponent makes a correction for things being counted more than once. The second reason is that the algorithm used to convert rank orders to utilities has a tendency to maximize the difference between the lowest level of each attribute and the highest. The exponent therefore corrects for the fact that a respondent's behaviour may operate within a narrower range than that represented by the utilities. Interestingly, in a previous study done by Market Facts Inc., there were 26 attributes and an exponent of .25--twice the number of our attributes and half the exponent. It should be noted that use of the exponent does not change the rank order of anything. A person's response to changes in levels is "damped" (in engineering terms) but rank orders are preserved.

Having calculated a person's "utilities" for each mode, the model then assumes that his trips are made on each mode in proportion to his utilities for each mode. A weighting is then applied to the trips to give the number of trips by persons in that class in the telephone survey. The total number of trips made is then calculated by multiplying by the ratio of population to telephone survey size.

FORECAST

Having defined the four existing modes and calibrated the model, the predictions of modal split compared closely with the actual modal split. The actual modal split was obtained from the telephone survey by taking the most recent trip of each traveller. The comparison between actual modal split and the projection produced by the model is given below. These are not the real number (they have been changed to protect the innocent) but the discrepancies shown are of the same order.

We then inserted the fifth mode--STOL--and defined it in terms of the 13 attributes. The projected market share is shown in the table below. STOL was given the same exponent and mode effect as Air.

Several runs were made, changing the levels of the STOL attributes to determine the effect on traffic. We tried different fares, different trip times, beverage service etc. so that the STOL marketing manager could put together an optimum package and have a couple of tricks up his sleeve if things did not work out as expected.

Curves were drawn showing the change in traffic with change in levels of an attribute. The curve below shows the response of market share for STOL to changes in fare.

An additional benefit was the ability to identify who the most likely travellers on STOL would be. The relative utilities (UC, UB, UT, UA, USTOL) were Printed out for each respondent and the top 10 percent who had the highest utility for STOL could be examined individually. For promotional purposes, the Marketing Manager needs to know where they live, which language they speak, how much they earn, their purpose of travel, etc.

PROTECTION OF MARKET SHARES

PRICE/DEMAND RELATIONSHIP FOR STOL

PROBLEM--OR WHAT WE WOULD DO DIFFERENTLY NEXT TIME

Let there be no doubt that the model has exceeded our expectations and has performed very well in a way that inspires confidence in the user. As in all projects, however, with the benefit of 20-20 hindsight we can see things that might be done a little differently. Progress comes with learning from our experiences.

One of our problems was in forecasting total travel between Montreal and Ottawa. There are statistics for the common carriers (which have to be taken with a large pinch of salt) but none for car which is twice as large as all the other modes put together. The requirement for portability meant that we should have been able to predict total traffic from the results of the telephone survey. We did not have too much faith in those but in the absence of any other data, they would serve the purpose. A cross check with data from other sources led us to believe that the total number of trips estimated from the telephone survey was too high and an examination of the data showed the reason.

FREQUENCY OF TRAVEL DATA FROM TELEPHONE SURVEY

It seems that the appeal of nice, round numbers like 10, 12, 15 and 20 was too much for many people to resist. This is a market research problem. Is there some way we can improve the accuracy of people's recall? Is it realistic to expect people to be able to recall accurately when the number of trips made was more than about 4?

Perhaps a better way would be to ask a respondent about the trips he made in the last month and then forecast annual volume making allowance for seasonality. This implies that external data on seasonality is available and that might affect the portability of the model.

Another problem is that the interaction between interview task, interviewer and interviewee may have resulted in some misunderstanding. Consider the following utilities for trip cost:

TABLE

This respondent was a French-speaking female aged between 18 and 24 with an income between $5,000 and $7,000 who went to Ottawa three times last year by car for non-business reasons. Her preferences seem to be inconsistent with her behaviour.

We would like to develop criteria to accept or reject a respondent based on the credibility of his responses. If a respondent tells us his preferences, what justification have we for rejecting him because it does not make sense to us? Consider this same respondent's utilities for time.

TABLE

It may not look terribly rational but is not nonsense. Should it be rejected? Another approach might be to throw away the top 5 percent of resPondents whose theta--the measure of badness of fit--is highest.

A considerable amount of work and soul-search went into determining the attributes and their levels and into the design of the questionnaire. The questionnaire was pretested, modified, pretested again and modified again before the field work of 1,055 interviews was conducted. Despite that, I believe that even more pretesting over a longer time would significantly improve the quality of the final results. The results of the field work indicate that some of the questions were not properly understood, some were not well defined (how do you define "ambiance"?) and others which were not meaningful could have been dropped. The choice of attributes and design of the questionnaire is a major task which should not be underrated.

Montreal and Ottawa were split into two zones each on the assumption that access and egress times to the airport and train and bus stations would vary by where the respondent lived. It turned out that a two-zone breakdown by city was not useful and that the breakdown should have been much finer--say 10--zones per-city.

I suspect that the model may overestimate the number of travellers that will switch from car to STOL. The reasons is that we have made no allowance for groups. Car is the only mode where four people can travel as cheaply as one. An individual travelling on his own, for example, may choose between the car at $4 and an air fare of $14. If a party of four is travelling together then the choice is between a dollar a head or a total air fares bill of $56. The model thus understates the attractiveness of car for people travelling in groups.

The Marketing Manager of STOL is using this model continuously and finding it invaluable in formulating his marketing plans for STOL. The acid test of how good the model actually is will only come in mid-73 when STOL goes into operation.

Perhaps it should be emphasized once again that this operation is an experiment, not a profit-making venture. With only 11 seats in it, (normally it has 20) the Twin Otter is not economic in this type of service. However, given the operating economics of a viable aeroplane such as the DHC-7, this model would prove to be very valuable in defining a profitable and attractive service. We anticipate that the approach taken in this study will have many more applications within Air Canada, not just for STOL.

REFERENCES

Kruskal, J. B. Multidimensional scaling by optimizing goodness of fit to a nonmetric hypothesis. Psychometrika, 1964a, 29, pp. 1-27.

Kruskal, J. B. Nonmetric multidimensional scaling: a numerical method. Psychometrika, 1964b, 29, pp. 115-29.

Kruskal, J. B. How to use M-D-SCAL, a program to do multidimensional scaling and multidimensional unfolding. Bell Telephone Laboratories, Murray Hill, N.J., March 1968 (mimeographed).

Johnson, R. M. Pairwise nonmetric multidimensional scaling. Psychometrika, to be published.

Gibson. R. E., Neidell, L. A., & Teach, R. D. Performance space analysis for an industrial product. Operational Research Quarterly, 1972, 23, pp. 125-3?.

----------------------------------------

Authors

J. D. Davidson, Senior Operational Research Designer, Air Canada



Volume

SV - Proceedings of the Third Annual Conference of the Association for Consumer Research | 1972



Share Proceeding

Featured papers

See More

Featured

Burnishing Prosocial Image to Self vs. Others

Minah Jung, New York University, USA
Silvia Saccardo, Carnegie Mellon University, USA
Ayelet Gneezy, University of California San Diego, USA
Leif D. Nelson, University of California Berkeley, USA

Read More

Featured

R8. Brand Perceptions and Consumer Support in the Face of a Transgression: Warmth Over Competence

Summer Hyoyeon Kim, University of Kansas, USA
Jessica Li, University of Kansas, USA
Jenny Olson, Indiana University, USA
SHAILENDRA PRATAP JAIN, University of Washington, USA

Read More

Featured

Beyond Needs and Wants: How Networked Hyper-rational Economic Actors “Win” the Deal but “Lose” the Shopping Trip

Colin Campbell, University of San Diego, USA
Hope Schau, University of Arizona, USA

Read More

Engage with Us

Becoming an Association for Consumer Research member is simple. Membership in ACR is relatively inexpensive, but brings significant benefits to its members.