Raimund K. Herz

The Use of the Bicycle

Homepage Manual Contents Navigation

MULTINOMNAL LOGIT-MODELLING MODAL CHOICE

Let us consider now the travellers as random utility maximizers of the conventional type. Suppose a market segmentation had been performed, which had led to the result that housewives from motorized households are a segment of the population which should be treated separately for the purpose of studying the modal choice for shopping trips. From Table II it is known that they use bicycles for 11.9 percent of their shopping trips.

With the same data base as before, two logit-analyses were performed for a random sample of shopping trips of motorized housewives. The well-known logit formula

p_jk = exp U_jk / SUM(i) (exp U_jk)

with utility function

U_jk = a_oj + SUM(m) (a_mjX_mjk + epsilon_jk)

was applied for a binary and a quintuple split.

In the first case, no information which is specific for each of the alternatives, bicycle or other modes of transport, was incorporated into the utility function. The contribution of each attribute X_m to this binary choice was left to the parameters a_o and a_m as they came out of the calibration procedure: attributes with large positive parameter values have a strong positive effect, while negative parameters reveal detrimental factors of influence on the use of the particular mode of transport.

The reason for not using alternative-specific information was that no data of this sort had been collected in the continuous survey on travel behaviour (KONTIV). Another peculiarity of this binary modal choice model is that all attributes are measured on or transformed into a nominal scale, that is, they are dummies.

Table VI shows the dummies and parameters b of the above logit-model transformed into

p_bike = 1 / (1 - exp (U_other - U_bike))

with

U_other - U_bike = delta U_k = b_o + SUM(m) (b_mX_mk + epsilon_k)

Apparently car ownership is most detrimental to cycling, and trip distance is also a very important factor of influence, particularily the range of one to three kilometers, although it is somewhat surprising that shorter and longer distances up to six kilometers are also favourable to cycling. While household size and educational status do not reveal a significant contribution to explaining this modal choice, a large town and adverse topography exert strong negative influences on cycling. The rather big constant term suggests that the specification of the alternative modes and/or utility functions should be improved.¹²

TABLE VI Parameters of the relative utility function delta U_k (sample size: 1000 shopping trips)

* From the list of socioeconomic and situational variables which possibly influence the modal choice of the bicycle, only those were included into this model which at the 0.05 level showed significant differences between housewives cycling and not cycling for shopping.

The log-likelihood ratio as an indicator of the variability explained by the model amounts to 0.57. However, even in the situation most favourable to cycling, the deterministic part of the relative utility delta U remains positive (+0.68) and exp delta U is larger than one. Thus, the deterministic part of the utility function will never lead to the decision to choose the bicycle. If nevertheless 11.9 percent of all shopping trips were made by bicycle, this logit model can cope with this fact only due to its stochastic element epsilon, which is assumed to be GUMBEL-type-l distributed.

These phenomena will often occur when logit models are to explain the choice of an alternative which has a relatively small probability to be chosen, as in this case on the average 0.119 for cycling compared with 0.881 for not cycling, which actually consists of a bundle of alternatives.

The second logit analysis was therefore performed on the modal choice of five alternatives: walking, cycling, driving or riding a car and using public transport. The observed percentages are given in Table II again. This time, mode-specific travel time and cost were included into the utility functions together with six dummies for socioeconomic and situational attributes, which were not mode-specific: educational status and car ownership as before, town size now under 20,000 inhabitants and topography as a one dummy variable (mountainous), age (younger than 40 years) and driving-licence as an additional indicator for car availability. Household size, the weakest factor of influence in the binary model, was removed from the utility functions.

Travel times were taken from the survey for the mode used and supplemented by travel times for the modes not used, calculated on the basis of average overall speeds. Travel costs were calculated on the basis of trip distance and average mileage and fares. For walking and cycling they were assumed to be negligible. Sample size is now 500 shopping trips of housewives from motorized households. Parameters are given in Table VII.

Most results from ;he binary logit-model are confirmed by this quintuple split model, e.g. the bicycle's advantage in small towns, its disadvantage in an adverse topography and the detrimental effect of car ownership on the use of the bicycle for the housewives' shopping trips. It is interesting to note that car ownership discourages even more people from using public transport. Having a driving-licence apparently does not encourage walking: housewives seem to prefer motorized vehicles for their shopping trips. Also, the influence of educational status is relatively weak: housewives with low educational status tend to use public transport or get a ride for their shopping trips. Most interesting perhaps is the fact that instead of cycling, younger housewives prefer other modes of transport for shopping.

The utility of motorized modes of transport is negatively affected by their costs. Apparently the price-elasticity for public transport is smaller than for the private car (for which in the case of getting a ride, the housewife's travel cost was assumed to be half of the cost of driving herself).

The parameters of travel times reveal speed advantages for motorized vehicles. Walking and cycling is time consuming, while the time consumed in a motorized vehicle surprisingly appears to be a benefit (cf. positive signs). Thus, increasing their travel time by some measure of traffic restriction would lead to an increase in utility for motorized modes of transport and increase in their probability of being chosen. But since the expenditure of time is generally not perceived as a benefit from the consumer's viewpoint, this model should certainly not be used for predictive purposes, in spite of the good fit perceived.

TABLE VII Parameters of the utility functions of five modes for shopping trips of housewives from motorized households (sample size: 500 shopping trips)

For this mode. the log-likelihood ratio is 0.52. Its performance with respect to replicating the observed modal split can be seen in more detail from Table VIII. The following percentages of choices correctly predicted by the model (according to maximum utility) are calculated from the figures in the diagonal of Table VIII: cycling: 18 percent correct, walking: 94 percent, public transport: 48 percent, car driving: 65 percent, and car riding: 45 percent correct.

So, unfortunately, it is just the housewife's choice to cycle for shopping which is poorly identified by this model. Apparently. the model needs further specification by including other factors influencing the choice of the bicycle. e.g. weather condition.

TABLE VIII Observed versus calculated modal choices for 500 shopping trips of housewives from motorized households