Scottish Government

DEVELOPING A METHODOLOGY TO CAPTURE LAND VALUE UPLIFT AROUND TRANSPORT FACILITIES

APPENDIX 2: LUTI MODELLING WORK

1. INTRODUCTION

This appendix reports on the additional work concerned with land-use/transport interaction (LUTI) modelling in the context of land value capture (LVC). It has been prepared in response to the instruction from the Scottish Executive in December 2003.

The purposes of the additional work were:

• to explore the potential applicability of LUTI models in the LVC context, including both strengths and weaknesses, and considering their relationship with the GIS-implemented Geographically Weighted Regression (GWR) approach; and then
• if appropriate, to establish requirements for future model development, including, if possible, the scale and feasibility of the work which would be needed to realise those possibilities.

The note is structured as follows:

• Section 2 reviews the case for considering that LUTI models are potentially applicable to land value (LV) forecasting.
• Section 3 summarises the processes in key LUTI models which actually forecast rents.
• Section 4 compares these features of LUTI models with hedonic pricing and GWR methods.

2. POTENTIAL APPLICABILITY OF LUTI MODELS TO LAND VALUE FORECASTING

Most, if not all, LUTI models in current use incorporate property rent or price variables, for different types of property and by zone. The rents are estimated within the model systems, as part of a process of reconciling the demand for each kind of property in each zone with the available supply at a particular point in time.

The details of the rent estimation processes vary considerably, as shown in the next section, but the rent forecasts will reflect other characteristics of the LUTI models. These include:

• sophisticated treatments of the transport system and of accessibility;
• since the models consider location choice, they consider different locations as competitors, and hence have the potential for negative impacts away from the schemes under consideration;
• in the more sophisticated models, major transport schemes can generate both local redistribution effects and possible net expansion or contraction of the local economy in question; and
• taking account of both existing and new building stocks, including the possibility that transport changes will (through their rent impacts) influence both the total quantity and (subject to planning policies) the distribution of new development within an area.

The treatment of the transport system depends on the exact details of the transport model used, but in general, it takes account of the infrastructure and services available, fares and tolls, competition/complementarity between different modes, and the relief (or not) of congestion on roads. Some models also consider changes in the availability of parking (due to changes in supply or changes in demand).

The treatment of accessibility varies considerably between different approaches to LUTI modelling. However, the result in all cases is that rents are affected by all the modelled factors in the transport system for all journeys, with consideration of the different aspects of accessibility which may influence actual/potential occupiers of land.

LUTI models in general, therefore, represent a category of rent-forecasting methods which take account of a very wide range of factors and responses, many of which are difficult to deal with in regression-based methods which forecast a single variable. They therefore appear to be potentially useful in considering the possible land value impacts of proposed schemes. There are, however, some major issues to consider. In particular:

• are the mechanisms by which these models adjust property rents (and the focus on rents) acceptable in a property/valuation context?
• are these models useful with the scale of zone sizes for which they are currently implemented?

The following two sections summarise how existing models forecast rents, and the zone sizes at which they operate. Section 5 then compares LUTI and non-LUTI methods. Section 6 discusses the material, and draws conclusions leading to suggestions for further work. These suggestions aim to allow the analysis of LVC issues to capitalise on the investment being made in LUTI models in Scotland.

3. RENT CALCULATIONS IN LUTI MODELS

Models considered

In Appendix A we review in detail the exact ways in which rents are calculated in two groups of models or modelling packages:

• those which are in current use in the UK; and
• those which, from previous reviews or current awareness, we (and others in the UK) recognize as significant advances in modelling methodology.

The packages in current use are TRANUS, MEPLAN, MENTOR and DELTA. Of these, the first and last are in current use in Scotland: TRANUS in Inverness and DELTA for a number of applications in the Central Belt plus the ongoing development of TELMoS.

The other models considered are:

• the residential component of the IRPUD model of Dortmund (Germany) - a one-off system which has been an important influence on a range of other models;
• the UrbanSim package developed in the United States - various applications mostly under development;
• the residential component of the TLUMIP model of Oregon being developed for the Oregon Department of Transportation; and
• the MUSSA model of Santiago de Chile - again, a one-off model rather than a package.

This section is a summary and discussion of what we found in Appendix A.

Overall approaches

TRANUS, MEPLAN, MENTOR and DELTA are all aggregate models: that is, they work by forecasting what proportions of total numbers of households or jobs will locate in each zone. TRANUS, MEPLAN and MENTOR are, for the present purpose, very similar, and we will concentrate on TRANUS unless the need arises to comment specifically on MEPLAN or MENTOR (which in turn are identical for the purposes of this note - MENTOR being a version of MEPLAN which can be integrated with other transport models, whereas MEPLAN (like TRANUS) incorporates its own transport model).

UrbanSim and the residential components of IRPUD and TLUMIP are microsimulation models: in effect, they operate on a list of individual households and allocate each household to a specific zone using a Monte Carlo simulation method. Monte Carlo simulation means that for each individual decision to be represented, e.g. one household's choice of where to live:

• a choice model (often of the same kind as in TRANUS or DELTA) is used to calculate the probability of the household choosing each outcome;
• one particular outcome is then chosen for that particular household by a random process reflecting those probabilities.

It would be possible simply to reproduce the TRANUS or DELTA models in microsimulation form, but to do so would be highly inefficient and indeed pointless. The advantages of microsimulation arise from extending traditional model designs so as to introduce one or more of:

• more information about the decision-makers (e.g. a sample of real household with detailed differences in composition, age of household members, instead of a category such as "older professional couples");
• more information about the choices (e.g. a sample of individual properties rather than a choice between zones);
• more sophisticated decision-making processes (e.g. variation in preferences between apparently similar households; consideration of the information available to the decision-makers; rule-based decisions rather than trade-off decisions).

All of the above are impossible or impractical to introduce into aggregate models in any substantial way, and all of them are widely recognized as desirable improvements in land-use (and indeed transport) modelling practice. On the other hand, microsimulation:

• can involve greater data requirements;
• raises new problems in calibration, rather than relying on standard statistical methods;
• most critically, because of the random element, each run of the model represents one possible outcome rather than the unique outcome; repeated runs of the model with the same inputs will produce somewhat different outputs. Where very large samples and few choices are involved, the differences between the results of different runs may be insignificant. Where it is necessary to look at the results from a small part of the sample - e.g. the outcome for a small set of properties rather than for the whole city - and especially where the focus is on the differences due to policy measures such as public transport change, rather than the absolute results - the implications of the model using random sampling may be quite significant. Dealing with this remains a major issue in the use of microsimulation models.

(Note that DSC, in collaboration with MVA and the University of Leeds, has recently been commissioned by Department of Transport to develop a microsimulation model of household location. This will involve integrating microsimulation sub-models of household change into the DELTA framework, working on relatively fine zones (English wards). This project is due to report towards the end of 2005. The treatment of rents or prices within this model, and the treatment of properties within zones, will need to be debated in the coming months.)

MUSSA exists in both aggregate and disaggregate forms, though this is not always obvious from the published papers which tend to describe one or other without being explicit about which is being discussed.

For completeness, we note that the non-residential parts of IRPUD and TLUMIP are aggregate models more similar to either TRANUS or DELTA. We have not considered those components.

TRANUS and DELTA

TRANUS and DELTA all work by adjusting rents within one modelled time period until the demand for floorspace matches the available supply. In TRANUS, the "modelled time period" is the situation at the end of a five-year period; most activities are assumed to locate, and all floorspace is assumed to be available, in each period. In DELTA. the "modelled time period" is a one-year period; only a minority of activities locate or relocate, and only a minority of existing floorspace is available to new occupiers, within any one year.

In both cases, the demand for floorspace of one type in one zone is the number of activity units seeking to locate there, multiplied by the amount of floorspace demanded per unit. The number of units seeking to locate is influence by the total availability of floorspace, the accessibility of the zone, the cost or utility of locating there and various other factors such as quality or environment. The ways in which accessibility, cost or utility and other factors are introduced into the model vary between TRANUS and DELTA, and indeed between different applications of each package, but in both cases the key points are that:

• the cost or utility is affected by the rent of floorspace in the zone, and
• the number of units trying to locate in each zone will be affected by rent there and in every other zone.

The amount of floorspace demand per unit (i.e. square metres per household or per employee) is also variable and affected by the rent. Each run of either TRANUS or DELTA therefore involves an iterative adjustment of the rent values modifying both the location choices and the floorspace/unit choices, until the total floorspace used matches the available floorspace, for each type of floorspace in every zone.

Various other complications also affect this process. In both TRANUS and DELTA, different activities compete for each floorspace type (e.g. different household types for housing, different business sectors for offices). In at least some applications of TRANUS, activities also have a choice between several types of floorspace, e.g. households may choose between houses and flats. In DELTA, floorspace will be left vacant if rents fall, relative to the previous year, or vacant floorspace will be brought back into use if they increase. This contrasts with TRANUS, which converges on solutions where all floorspace is always occupied.

TRANUS does not model any choices by landlords - they are assumed to accept whatever rents result from the competition (or lack of it) between potential occupiers. In DELTA, the vacancy response implies that a proportion of landlords will be unwilling to accept falling rents and will choose to keep their property vacant. In neither model is there any explicit choice by landlords between alternative tenants.

It is also worth noting that in both TRANUS and DELTA residential rents are in effect constrained by the fact that they are related to household incomes - the overall level of residential rents cannot go too far out of line with income growth. Non-residential rents are not constrained in any equivalent way and therefore could go to extreme values if the supply of floorspace was too far out of line with demand.

IRPUD, UrbanSim, TLUMIP

IRPUD, UrbanSim and TLUMIP all work by allocating individual units of activity to zones, cells or individual dwellings in the course of a one-year (or shorter) period, with rent values which are fixed at the beginning of that period. As in TRANUS and DELTA, the choice of location by households or businesses is influenced by a range of factors including accessibility, cost (based on rent) and various quality terms. In the IRPUD case, the model explicitly models households' choices between available dwellings, and the dwellings may be occupied by a household or not; there is therefore no possibility that the number of households located in a zone will exceed the availability of space. In UrbanSim and TLUMIP there are presumably similar mechanisms, though they are not altogether clear from the available documentation.

The models then compare the amount of space occupied with the amount available, and use this ratio to adjust the rent to be used in the next period.

These models are therefore broadly similar to "error correction" models in that the rent in period 2 reflects whether the vacancy rate in period 1 was higher or lower than an "expected" or "normal" value. Note that IRPUD works in two-year periods, UrbanSim in one-year periods and TLUMIP in four-month periods.

IRPUD works in terms of specific lettings/sales of individual units and involves an accept/decline choice by the landlord - but no adjustment of the rent. Landlords are assumed to invest in their housing stock if by doing so they can expect to raise their profits. The proportion of dwellings upgraded in each period is calculated for each dwelling type in each zone as a function of the expected rent increase in that submarket after improvement. As the eventual rent increase is not known at this point in time, the landlords employ a simple rent expectation model based on vacancy rates at the beginning of the simulation period.

In the development component of UrbanSim/TLUMIP, the models simulate developer choices to convert vacant or developed land to urban uses, including the type of improvements and density, based on their profitability expectations and subject to constraints imposed by governmental policies such as zoning and infrastructure availability. These profitability expectations are influenced by prior prices and revealed demand in the location and building type preferences of businesses and households.

The model simulates land market clearing by adjusting prices to reconcile the competing demands for locations and structures among households and businesses against the supply of space in each zone. The ratio of demand to supply in each zone for each type of space (housing and commercial structures by type) induces proportional price adjustments for these structures. The adjusted prices produce new market signals to demanders in the subsequent year, thereby influencing preferences for zones and building types.

MUSSA

MUSSA has been developed to forecast the expected location of agents, residents and firms, in the urban area. The location problem assumes that properties are allocated to the highest bidder by auctions and market equilibrium is attained by the condition that all agents are located somewhere, therefore, supply satisfies demand. This process produces rents for each property unit and levels of satisfaction (benefits) to located agents. The whole model is therefore specified in terms of discrete units, including dwellings by type and households by categories (like IRPUD and UrbanSim, and unlike TRANUS or DELTA, where floorspace and households are treated as continuous quantities). Consumer agents, ie households and firms, are rational and their idiosyncratic differences are modelled by a stochastic behaviour.

The critical feature of MUSSA is that it incorporates an explicit view of property owners as decision-makers in the location of households or businesses, in that they choose whether or not to accept a particular bid for a unit of property. The model therefore involves predicting the bids which households or businesses will make, based on their willingness to pay for alternative locations. As in all the other models considered, the consideration of alternative locations is influenced by accessibility and a range of other factors.

A significant difference with other land use models is that in MUSSA the interaction between consumer agents - households and firms - is explicitly described in the equilibrium. These interactions makes each agents location choice dependant on all other agents choices, making the calculation of equilibrium a highly complex mathematical problem solved in MUSSA by ad-hoc algorithms. It is worth noting the tremendous dynamic in the land use pattern introduced by location externalities, because each choice affects all other choices and, in theory, the whole location pattern.

4. ZONES AND ZONE SIZES

LUTI models have traditionally worked with relatively large zones compared with traditional, network-oriented transport and traffic modelling. They therefore represent the average conditions (and rents) within those large zones. This was seen as a necessary trade-off to allow computing time and memory for the extra complexity of the land-use modelling.

In practice, much of the extra computation is to do with using more sophisticated transport models, and the same trade-off has been made in "strategic" transport models such as START, TRAM, STM and APRIL which involve more complex choice processes. The extra requirement of LUTI modelling arises mainly from the need to work forward through a sequence of periods, involving a sequence of transport model runs for different years rather than modelling just the traditional "horizon year".

Recently there have been moves to build LUTI models with larger number of zones directly matching the zone systems of more traditional transport models. This was done in Scotland for CSTCS (where the LUTI model matched the most relevant part of the CSTM3A zone system) and is being extended for TELMoS (which fully matches TMfS). The approximate numbers of zones in these and other current Scottish models for the largest cities in each model are shown in Table A below. Whilst it is difficult to compare Edinburgh and Glasgow with Inverness, the figure for Inverness suggests that the TRANUS application there has been built with "traditional" rather than "strategic" zones.

Table B

Source: DSC documentation of DELTA applications; DSC notes from TRANUS model seminar, Victoria Quay, 11 December 2003.

n/a = not applicable, ie the city is peripheral to the model and therefore not fully modelled in full detail, or the city is outside the model altogether.

As noted above, the microsimulation models such as IRPUD, UrbanSim and TLUMIP are moving away from simple zones to consideration of what happens to individual properties within zones. However, this does not necessarily mean that they consider the exact locations of individual properties. UrbanSim and the latest versions of IRPUD get close to this, by working with very fine cells. The original version of IRPUD did not - it processed a list of properties within each (large) zone, without considering spatial variations within the zone. TLUMIP is somewhere in between, in that most of the modelling takes place at a fairly fine zone level, with some of the results (but not rents) then converted down to a fine cell level.

The ability of any of these models to consider the impacts of transport on property values at a fine scale is of course also dependent on the associated transport model working at the required level of spatial detail.

The "more detailed" microsimulation models may therefore be capable of considering the property value impacts of transport in much greater spatial detail than traditional LUTI models, but this is not automatically the case.

5. COMPARISON WITH NON-LUTI METHODS

Table B attempts to summarise the comparison of the above models with hedonic pricing and geographically weighted regression (GWR) techniques ( see Appendix B).

Table B

The comparison highlights the following potential merits of LUTI modelling:

• more sophisticated accessibility treatment (though such measures could be used in GWR etc);
• consideration of demand responses on the part of occupiers - notably, it takes account of the fact that there is a finite demand for space at any one time, which will limit the scope for land value increases - the GWR approach will not do this;
• forecasting ability - will produce a stream of forecasts, taking account of developments over time; and
• potentially, to forecast the ways in which the LVC process itself will modify the outcome (and the achievement of other government objectives), by incorporating the LVC process itself into the modelling (e.g. by modifying the costs of occupying property, or the returns to development).

More pragmatic arguments in favour of the possible use of LUTI-type models are:

• to capitalise on the investment that is already made or committed to such models (including TELMoS); and
• to ensure consistency between the different aspects of appraisal (including EALIs) and the consideration of LVC in the potential funding of schemes - bearing in mind that the funding method itself may have to come within the consideration of the appraisal (as is already the case for road user charging) if the funding method is expected to change either the actual outcome (in terms of development or use of land) or the distribution of benefits between different groups in society.

The demerits of LUTI modelling relative to HP/GWR are

• spatial aggregation;
• it works on broad generalisations of floorspace type without taking account of individual property characteristics (except in some microsimulation applications);
• microsimulation applications are not deterministic, so more difficult to use or interpret in planning/financial situations where clear-cut answers have to be reached;
• greater complexity - can be argued that by trying to take into account all the factors they are less clear about the role of any one factor.

It must of course be kept in mind that HP/GWR are not in any sense "perfect"; they form the basis for consideration of LUTI models only because they are (or at least HP is) the more conventional approach. We note for example that hedonic pricing models for housing have frequently left out variables having to do with features associated with the property's location. Too many problems arise regarding fulfilment of the assumptions underlying the multiple regression technique and it is easier the leave these variables out.

Heteroscedasticity is another problem that often arises with the use of location variables. It arises in the context of hedonic pricing models for housing when separate models are not used for neighbourhoods. A wider variance for the error term results for higher priced properties, showing that higher priced properties tend to sell over a broader range of variables (square footage, number of bedrooms, etc.) than lower priced properties.

6. DISCUSSION AND CONCLUSIONS

The potential uses of LUTI modelling in Scotland were considered recently in WSP's scoping study (WSP, 2003a) investigating the development of a land-use and transport model for Scotland. They emphasised that "it is important for the model to predict the relative levels of prices/rents between areas as these may have a major impact on the location of households and industries" (WSP, 2003a, para 5.48).

They suggested (ibidem, para 5.44 ) that LUTI models should deal with medium-term changes in rents so as to avoid the short-term impact of macroeconomic variables on prices (housing prices, in particular), though elsewhere WSP have suggested that these could be brought into modelling via a link "to the financial instruments that influence such cycles" (2003b, para 7.4) In general, however, WSP seem to have followed the view that rents or prices are important to include for realism of other model outputs rather than of interest in themselves.

They do, however, note as potential uses of LUTI models both "more complete cost-benefit analysis of transport policies" and "cost-benefit analysis of land-use policies" (WSP, 2003a, paras 3.7-8; WSP, 2003b, paras 5.7-8). They do not actually mention rents in this context, though they do mention the need to obtain "a complete and consistent set of economic responses". It seems essential that such a set of responses should include rents or prices - for consideration of the distribution of impacts, even if it was argued that these did not affect total benefits.

WSP refer to the potential use of LUTI models to improve the estimation of the impact of transport policies on land-use patterns, and vice versa (2003a, paras 3.4-5; 2003b, paras 5.4-5). They do not particularly refer to the integration of land-use planning and transport planning in this context, but clearly the achievement of better integration, with all the components working in the desired directions, would be one of the underlying objectives of better forecasting. This seems to follow the same approach as that underlying the development of DELTA, which consciously followed the conclusions reached by Still (1997). He found that planning professionals felt that models which included a representation of the property market were likely to be more realistic than those which used other, more artificial means to balance supply and demand, but that they were not particularly interested in the price or rent outputs themselves.

One particular aspect of integration between land-use planning and transport planning to which enhanced LUTI models should be relevant is the assessment of proposals for Transport Development Areas (TDAs). The definition of a TDA put forward by RICS (Hine et al, 2000) included the proposal that: "As an economic concept, TDAs are also the focus for … arrangements whereby public transport operators received additional funding based on the transfer, where appropriate, of part of the higher financial returns to development which might be achievable in such areas". The appraisal of such proposals would need to take into account a number of factors which would probably be better dealt with through LUTI models (including the patronage and revenue of public transport to/from TDAs, and the impact on other corridors/development areas elsewhere) than in the more limited, land-value specific methods.

But having said all that, exactly how useful are LUTI models for rent forecasting in their present form, or forms? This question comes back to:

• the rent calculations themselves, and
• the zone size problem.

The rent calculations themselves do not seem to be a problem. In some cases, the rent calculations are incremental (estimating changes from a base situation), and so the initial values are built into the input database. Where the models calculate absolute rents, they are likely to use variables similar to those in hedonic price models (and indeed may incorporate such models). In so far as the rent calculations in LUTI models put more emphasis on the impact of supply:demend balances, this is incorporating a set of important effects which are missing in hedonic analyses and GWR.

The zone size problem is more of a sticking point for the use of LUTI in relation to LVC issues, especially for commercial property. This is particularly so for traditional "town centre" retail property which is very highly sensitive to the details of location, and where indeed rents are often different within different parts of one property. For residential property the zone size issue may be somewhat less critical in terms of sensitivity to details of location, but there is another problem of the range of differences between individual properties and between the qualities of different areas at a very fine level.

We conclude that it is not possible to move directly to using existing LUTI models in the kinds of analysis that would be needed to provide the full demonstration of LVC possibilities, and certainly not in the kinds of analysis that would be needed to "enforce" LVC by demonstrating to owners (or if necessary to the appropriate tribunals) the land value changes accruing to specific properties from specific transport changes.

However, there are other stages in the planning of possible LVC schemes to consider, as well as possibilities for the enhancement of existing methods (both LUTI and HP/GWR). We believe it would be worthwhile:

(d) to consider how well existing models with relatively fine zones (e.g. TELMoS) can contribute to the earlier stages in appraisal of LVC proposals - especially having regard to their ability to consider how LVC schemes may affect occupiers' decisions in ways which affect the achievement of other local and national government objectives;

(e) to consider whether it is possible to modify or extend LUTI models to work at much finer spatial levels and possibly on samples of individual properties;

(f) to consider, either as a way of achieving the above or as an alternative to it, whether it is possible to use the existing or finer zonal outputs of LUTI models within HP or GWR as a way of incorporating those effects which are better or only handled in LUTI.

Our recommendations are as follows.

Possibility (a) should be followed up by looking at a sample of TELMoS results, showing the rent impacts of transport schemes to asses their usefulness for consideration of LVC. This small piece of work could take place later in 2004, and should probably involve:

• testing a (probably hypothetical) scheme whose impacts are fairly clearly expected in the light of the present study; then
• (ii) examining the TELMoS results in the light of those expectations.

The examination should consider whether the TELMoS results conform with the expectations; if so, whether TELMoS adds value to what can be informally predicted; and if not, whether TELMoS provides a convincing account of why the simple expectations may not be correct. If the findings are encouraging, consideration should be given to the possibility of incorporating LVC methods into TELMoS in order to assess the impacts of LVC itself.

Possibility (b) should be kept in mind as a possible major enhancement of a model such as TELMoS, or as a new research study;

Possibility (c) should be included - as a possibility for further investigation - in the specification for any future work using HP or GWR (possibly as a follow-up to the present project in general).