The results obtained when estimating the general spatial model (
1) for the three modes of transport considered (PRIV, PUB, and WALK) show that the spatial error term is never significant, and the spatial lag term is significant only for private and public transport. This is not an unexpected result, since walking or cycling journeys only cover a short distance and are supposed to depend on the characteristics of the locality itself, but not on those of the neighbouring localities. Therefore, we estimate the appropriated reduced models: The spatial lag model for private and public transport and the non-spatial linear regression model for walking/cycling:
where
X is a vector including land use factors (gross population density, commuting self-containment, street intersections density, and street density), characteristics of public transport (distance to transit, destination accessibility, and public transport supply) and high and medium socio-economic profiles via the dummy variables Cluster1 and Cluster2 which take values of 1 if the locality belongs to Cluster1 or Cluster2, respectively, and 0 otherwise.
Given the presence of the spatial lag, models (
4) and (
5) are estimated using spatial two-stage least squares (S2SLS) and model (
6) by OLS, always under the general assumption that the error term
u may present unknown heteroskedasticity and/or spatial autocorrelation. Accordingly,
Table 5 and
Table 6 show the estimations for the coefficients and the levels of significance for three standard errors calculated under different assumptions: Homoscedasticity and no autocorrelation (Classical), heteroskedasticity (White correction), and spatial heteroskedasticity and autocorrelation (SHAC) [
64]. We believe that it is preferable to go for robust results if they lead to different levels of significance. As can be observed, the conclusions concerning the significance of the variables barely change, though the results provided by the SHAC standard errors are slightly more conservative.
4.1. Main Findings and Discussion
Table 5 and
Table 6 show that land use factors influence the patterns of modal split in commuting in the AMBA, although there are important differences depending on the mode of transport. Private transport use depends on all land use variables except commuting self-containment capacity. Thus, the share of private transport is larger in those localities with lower population densities, lower street intersection density, and higher street density. By contrast, public transport use is positively linked to population density and street intersection density and negatively linked to self-containment capacity. Finally, non-motorised trips depend positively on commuting self-containment and street intersection density and is negatively linked to the density of roads.
In general, the results obtained are similar to those of previous studies carried out in the US and Europe in the sense that we also find evidence of the effect of land use factors on modal split [
4,
5,
7]. However, our findings about the relative importance of these factors present some differences. While our results confirm that street design variables are highly relevant in explaining modal choice as happens in the US, on the other hand, they show that density still plays an important role in determining the share of private and public transport in the AMBA, which is more in line with the results obtained by Guerra et al. [
21] for Mexico than with the conclusions of Cervero et al. and Naes [
5,
7].
The spatial lag is a statistically significant variable for private and public transport which means that there are spatial effects that are not captured by factors included in the model. According to Equation (
2), this result implies that the use of both private and public transport depends not only on the characteristics of a given locality but also on those of its neighbours. If the urban characteristics of a given locality (low population density, low intersection density, and high street density) favour private vehicle use, the fact that adjacent localities have these same characteristics will result in an increase in vehicle use throughout the area. Commuters in a given locality may also benefit from the fact that the urban characteristics of adjacent localities (high population density or high street intersection density) meet the minimum requirements for having efficient public transport. Conversely, the presence of public transport in a given locality may benefit the adjacent ones. This result suggests the existence of spillover effects associated with transport network characteristics, which have effects that extend beyond the limits of one locality and encourage use of public/private transport in adjacent areas, although it could also in part depend on the data scale. We should bear in mind that mobility does not recognise borders and goes beyond administrative limits. Commuters usually cross several localities and our result suggests that the mode of transport used depends partly on the characteristics of all of them.
If we look in detail at the total impact of land use factors, transport characteristics, and socio-economic profiles on the modal split in commuting (see
Table 7), we can see that population density has a significant impact only in relation to both private and public motorised transport. The magnitude of the total impact is similar but with opposite signs, i.e., low densities are conducive to private transport use and high densities to public transport use. These results coincide with previous research linking higher density areas to better transit service (increased demand makes it more cost effective) and a reduction in private transport due to worse driving conditions (reduction in traffic speed and parking supply, increase in traffic congestion) [
4,
6,
21]. However, our results differ from previous research not only in the US but also in Latin America [
22] that claimed that density brings destinations closer together and encourages non-motorised travel. In the AMBA, the increase in density is not sufficient to encourage travel on foot or by bicycle, probably because it does not necessarily mean that the workplace is close to the place of residence.
A high self-containment capacity, as a result of a good functional mix between work and residence enables a locality to retain its workforce, thus meaning that the distance between home and work is likely to be shorter. Previous studies in the US claimed that the direct consequence of this is a greater use of non-motorised mode [
5], while some research in Latin America shows that functional diversity does not affect walking/cycling trips [
22,
24]. Moreover, the effect of this variable on motorised transport is not so clear either. In some situations, suburban dispersion of employment can reduce the average commuting distance, although it tends to increase total per-capita vehicle travel [
6]. Other authors find that functional mix have a small and negative effect of vehicle travel but a positive one in public transport [
4,
5]. In the AMBA, this factor does not affect private transport trips while it is highly relevant to non-motorised and public transport journeys. The results show that a higher diversity in a locality encourages people to walk or cycle to go to work but reduces the use of public transport. The latter result could imply that workers use public transport for longer journeys outside their locality but commute on foot or by bike when they work and live in the same locality.
The urban design variables are, in general, highly significant to explain modal split. Street intersection density as a key factor for connectivity is one of the most relevant variables for modal split, affecting both motorised and non-motorised modes [
5,
20,
48]. The average number of intersections in the AMBA is 167 and UN-Habitat considers that a score of 100 intersections means that a city is more suitable for non-motorised trips [
65]. The observed positive impact on public transport and walking/cycling can be explained by the fact that more intersections means more crossing points, thus shortening distances. By contrast, the negative impact of this variable on private transport may be due to the fact that too many intersections can lead to congestion, which cannot be avoided by using a system of exclusive lanes such as the one available for rapid bus transit [
6]. Finally, previous research does not provide a conclusive result on the relationship between street density and the mode of transport used. The reason is that different street systems with the same network density may have different effects [
66]. Street density does not reflect whether streets are station-oriented, pedestrian friendly streets, or highways. In our study, the SD variable includes all types of roads, from footpaths, which are specifically for non-motorised mobility, to motorways, which require the use of vehicles. This may explain why street density in the AMBA has a substantially positive impact on private transport, in line with previous studies [
10,
16]. The negative impact of street density on non-motorised trips may be explained by the fact that walking and cycling conditions are also affected by other urban design characteristics, such as the quantity and quality of pavements, pedestrian crossings and paths, and the safety and attractiveness of routes [
67,
68]. Therefore, a higher street density does not necessarily mean more journeys on foot or by bike.
As far as the variables related to public transport are concerned, distance to transit and destination accessibility are considered relevant factors in public transport. In fact, previous studies show that they are the most important factors for motorised transport [
5] and that there is a significant decrease in the use of public transport as walking distance to bus stops or stations increases [
10,
69]. By contrast, our results for these two variables are not significant. This conclusion may be partly due to the fact that many travellers have no alternative but to use public transport even if it is not very accessible, since they do not have a car and cannot reach their workplace on foot or by bicycle, but another part may come from the design of these variables in the AMBA. As mentioned above, given the available information, these variables only measure the distance travelled by public transport users both to public transport stops and to their destination. In other words, they only include distances below a certain threshold that ensure a degree of accessibility to public transport. They do not include longer distances that make public transport not so accessible, and that may favour non-motorised travel if the home-work distance can be covered walking or cycling, or car travel if that is not the case.
As expected the public transport supply variable has a substantial impact on the use of motorised transport in the AMBA: The broader the range of public transport, the more public transport and the less private vehicles are used. This is in line with the results of other studies which indicate that car travel tends to decline when there is a strong, competitive transit system [
6]. Concerning non-motorised trips, even though the sign of the coefficient is negative as expected, the variable is not significant. Therefore, we may conclude that in the AMBA, a wider variety of public transport does not significantly affect walking or cycling journeys, while, for instance, Guerra et al. [
21] obtains that in Mexico commuters are more likely to use this mode in areas with a broader public transport supply. Our result may be due to the fact that commuters tend to use public transport only when the distance is too great to travel by bicycle or on foot, and yet prefer to walk or cycle if the distances are short enough, whether or not public transport is available. According to ENMODO data, the average distance travelled by public transport is 7.1 km, while walking or cycling trips cover around 1.8 km. It should be also taken into account the negative reputation of certain public transport, mainly the conventional rail service (delays, cancellations, overcrowding, etc.) and the low purchasing power of the population.
The results obtained show that the socio-economic profile of a locality has a high impact on modal split. As expected, the proportion of public transport and non-motorised travel is lower the better the socio-economic situation of a locality. Note that the use of public transport in the high-level cluster is 11 points below its use in the least privileged localities. By contrast, localities with the lowest socio-economic profiles use private transport in a much lower proportion compared to other profiles: 17 points below the localities belonging to the “high-level” cluster and 10 points below the localities with a “medium” profile [
10]. Our conclusions confirm the results obtained in previous research on this topic that show that socio-economic factors not only are highly relevant to explain modal split but sometimes they are even more important than urban development factors [
21,
22,
24].
Finally, as explained in
Section 3.2, the presence of the spatial lag generates feedback effects in the models for private and public transport. Thus, the total impact of a change in a variable
for a given locality can be divided into direct effects, which remain in the locality, and indirect effects or spillovers that affect other localities (see Equation (
3)). The analysis of the results shown in
Table 7 gives an idea of the strength of these feedback effects. Firstly, it may observed that the direct impact of each variable is very similar to the coefficient estimate. This means that the locality where the change occurs does not receive much feedback from other localities. Secondly, the indirect impacts of each factor are not very significant with the exception of the results obtained for private transport. This conclusion may be due to the fact that spillover effects come from the inclusion of a spatial lag, which is a combination of all the explanatory variables, and it does not seem easy to isolate and assign them to a specific factor. In the case of private transport, indirect impacts of population density, street intersection density, public transport supply, and socio-economic profiles are significant. This means that a locality surrounded by others with low levels of population density, connectivity, and public transport supply will tend to use private vehicles more. Private transport becomes the main option when population density is not sufficient to provide efficient public transport in the area.
4.2. Effectiveness of Public Transport Policies
As pointed out above, the supply of public transport affects motorised mobility, increasing the use of public transport and reducing the use of private vehicles. This suggests that policies aimed at improving the public transport system would help to redirect mobility towards a more sustainable model. However, the results also show that, all things being equal, the use of the private transport is significantly higher in localities with high and medium socio-economic profiles than in those with a lower profile, while the use of the public transport is significantly lower. These differences are so large that they make us wonder whether all localities would respond similarly to policies aimed at promoting public transport, regardless of their socio-economic profiles. As a first approach to test this hypothesis, we include a set of interaction terms in models (
4) and (
5) combining socio-economic profiles (clusters) and public transport variables in order to allow the impact of these variables to differ from one cluster to another. Note that the low socio-economic profile, Cluster3, is the reference group chosen.
The results in
Table 8 might help to refine our earlier conclusions. Firstly, we found that the impact of changes in public transport supply have no significant effect in localities with a low socio-economic profile (Cluster3) for both private and public transport. This result can be partly explained by the fact that less than 26% of households own a vehicle, so private transport use is low and limited, probably intended for long journeys not covered by public transport (secondary roads, sparsely populated areas, etc.). Even if transit supply were increased in these localities, it is unlikely that public transport would cover these routes and the use of private vehicles would still be necessary. Secondly, the difference in the impact of public transport supply between localities in Cluster3 and the well-off localities in Cluster1 is not statistically significant. This result suggests that in the latter, where 46% of households own a vehicle, a broader public transport supply is unlikely to lead to any significant change in mobility patterns. Note that these are the only localities that have underground services, which are used even by families that have their own vehicles [
41]. However, getting these people to switch from cars to buses or trains seems to be no easy task, given the negative reputation of public transport, particularly trains [
49]. To achieve this goal, it would be necessary to change their perception of public transport by improving the safety and security, comfort, and the objective aspects of journey time and frequency of the modes of transport concerned [
70]. Finally, in those localities with a medium socio-economic profile (Cluster2), the effect of improvements in public transport supply is significant and leads to increases in public transport use and decreases in private car use. Around 36% of the households in these localities have access to private vehicles, but it seems that some would be willing to switch to public transport for commuting if the offer were improved.
The previous result indicating that destination accessibility (DA) do not influence average mobility patterns is slightly modified by the results shown in
Table 8. Once again, greater distances from the public transport arrival stop to the workplace result in less use of public transport only in localities with a medium socio-economic profile. In short, it could be concluded that measures to promote the diversity of the public transport supply and to improve its accessibility could increase demand, particularly in localities with a medium socio-economic profile. It seems that, on average, the commuters in these localities have more access to private transport than Cluster3 on the one hand and less resistance to switching to public transport than Cluster1 on the other.