An Experimental Economics Investigation of the Land Value Tax: Efficiency, Acceptability, and Positional Goods

Joshua M. Dukea,*, TianHang Gaoa

September 18, 2017 aDepartment of Applied Economics and Statistics, University of Delaware, 531 S College Ave, Newark DE 19716 *Corresponding author. E-mail addresses: [email protected], [email protected]

Acknowledgements The authors are grateful for funding from the Lincoln Institute of Land Policy and support from its fellows and staff members, including Joan Youngman, Semida Munteanu, and Mark Skidmore. The authors appreciate the helpful feedback from the participants at the 2016 David C. Lincoln Fellowship Symposium. The authors thank Emerson Paradee, Xingguo Wang, and Greg Vitz for assistance with administering the experiments.

Abstract This research offers the first economic experiment investigating the land tax, where landowners invest under different property tax regimes. A voting treatment assesses the relative acceptability of land value taxation. Results show a land tax produced greater overall welfare in only 37.5% of the experiment sessions. Systematic over-investment arises from the positional- good characteristic of residential land investment, but this effect vanishes when the positional- good indicator is removed. The experiments show that the participants unexpectedly voted in favor of the land tax, suggesting that the efficiency and acceptability of the land tax may be more complex than in non-behavioral economics modeling.

JEL Classification: C91; H21; H31; H71

C91, Design of Experiments--Laboratory, Individual Behavior;

H21, Taxation, Subsidies, and Revenue-- Efficiency • Optimal Taxation;

H31, Fiscal Policies and Behavior of Economic Agents-- household;

H71; State and Local Taxation, Subsidies, and Revenue

Keywords: Land Use, Behavioral Economics, Taxation

1. Introduction

Economists famously led by Henry George have argued in favor of a pure land value tax (LVT) or its close cousin, a split-rate tax (SRT), in part because they encourage land investment and raise revenue for public goods without distortions. As a fixed resource, land will capitalize locational benefits collectively supplied. Targeting these rents with a land tax will not distort the incentive to invest in parcel improvements, as does the ubiquitous uniform property tax (UPT) that taxes improvements and land equally. Most of the recent work on George focuses on theoretical and simulation models of optimal city size (c.f., Behrens et al. 2015). This paper seeks to complement those studies with the first economic lab experiment on how property taxation affects household investment. The data explore three dimensions of the relative performance of the land tax, including the investment incentives, wealth created, and political acceptability. The experiment also explores a new possibility in land value tax research; land value taxation can exacerbate the positional goods characteristics of housing and produces lower aggregate welfare through overinvestment.

The efficiency of land taxation has been assessed using simulated general equilibrium models of open and closed cities and a handful of empirical studies. Some of the results show that LVT generates more intensive capital investments in land (Pollack and Shoup 1977;

Plassmann and Tideman 2000; Banzhaf and Lavery 2010). Other land tax benefits, including attenuated tax distortions and increased investment, have also been explored under a system of inaccurate assessments with positive results found for LVT (Chapman, Johnston, and Tyrrell

2009). Several studies examine whether the ostensibly progressive land tax may be regressive through a tax regime change in real-world settings (England and Zhao 2005; Bowman and Bell

2008; Choi and Sjoquist 2015). Prior results show that LVT can provide sufficient revenue

1 (DiMasi 1987) and can increase density in cities (Brueckner and Kim 2003; Banzhaf and Lavery

2010; Choi and Sjoquist 2015).

In reaction to this wealth of theory and evidence, Fischel (2015, p. 15) asked: “If economists like (the land tax) so much, why does it seem so rare in practice?” The land tax must be politically unacceptable or practically un-implementable. Research on Pennsylvania’s efforts suggests although SRT led to greater efficiency, few municipalities have adopted the tax and, surprisingly, some LVT adoptees are returning to UPT (Fischel 2015, pp. 16-17; Banzhaf and

Lavery 2010). Youngman (2016, p. 18) writes that land taxes have suffered from both

“administrative failures” and a “lack of political support.” Bourassa (2009, p. 195-6) identified five objections to the land tax; three reasons involve planners’ professional challenges such as difficulty to set rates, adjusting zoning to prevent too much density, and educating the public about land taxation. Bourassa also identified two economic challenges from land taxation, which will be addressed in this paper: (1) it is a tax on unrealized capital gains, that is, wealth rather than cash flow; and (2) it is a policy change that creates winners and losers, and losers will object. That LVT taxes unrealized capital gains may lead some to object on ethical or economic grounds—especially as it contrasts with most taxes that are levied on transactions. The fact that most of the United States currently raises local revenue with UPT means that a switch to LVT would be a policy change, which unavoidably would create policy losers; Plummer (2009) explored these horizontal and vertical inequities in tax incidence and LVT acceptability with both qualitative arguments and numerical examples. In sum, these studies suggest that acceptability depends on more than arguments about social efficiency or tax progressivity.

This paper extends these behavioral arguments about LVT acceptability by introducing another challenge to land taxation: positionality. Housing has long been recognized as a

2 positional good (Veblen 1899; Frank 1985a,b; Hirsch 1977). When a good has positionality, consumer spending will be affected by interpersonal comparisons (Frank 1985b). Recent work by Patacchini and Venanzoni (2014) finds a significant effect from social comparison and a conformity peer effect in shaping the demand for housing quality. One implication of these findings is that one’s housing improvements can cause a negative externality for neighbors, causing neighbors to overinvest to re-attain their prior level of housing utility. Housing utility is derived in the context of one’s neighborhood, and this can lead to an expenditure cascade: “a process whereby increased expenditure by some people leads others just below them on the income scale to spend more as well, in turn leading others just below the second group to spend more, and so on” (Frank, Levine, and Dijk 2014, 57).

This paper argues that LVT efficiency and acceptability ought to be complicated by positionality. The equity of LVT depends on (1) whether those similarly situated in society are treated “the same” or “the same enough” to satisfy norms of fairness; and (2) whether those differently situated accept their differential burdens. Tax equity associated with cash flow, or income, is easier for the general public to observe and understand. Land capital inequities introduce a second “wealth” dimension upon which a tax policy is evaluated, and imperfectly observed. The public will likely be skeptical of LVT because it is more difficult to observe land capital differences and yet the tax rates will be “the same” and perceived to be vertically inequitable, on the face. As Fischel (2015, p. 17) explained, Pittsburgh rolled back its SRT system when people rejected the idea that owners in the same neighborhood pay the same taxes even when their houses are of different sizes. In other words, LVT does not seem vertically equitable because “people in larger houses should pay proportionately more” (Fischel 2015, p.

17), but they do not. LVT acceptability therefore conflates with the positional-goods aspect of

3 housing, and this produces an obstacle for acceptance for which behavioral economics are well positioned to explore.

The contribution of this work arises from using behavioral economics to analyze the efficiency of LVT, the positional-goods characteristics of housing, and the acceptability of LVT.

To our knowledge, this is the first economic experiment on LVT. Another innovation in this experiment is that it makes endogenous the process by which land values appreciate; therefore, capturing the fundamental rationale behind a Georgian land tax—to capture the rent imbued by location. The treatments test (1) the social efficiency of LVT and SRT relative to UPT, or status quo, baseline; (2) the effect of a positional-goods signal on investment behavior; and (3) the public acceptability of the LVT and SRT relative to UPT using a voting mechanism. Experiment participants will take on the roles of different types of landowners making decisions about investing in their property (termed “improvements” in this experiment). Experimental economics techniques are ideally suited for studying LVT because there are limited instances of its adoption, the few examples of adoption of LVT is often incomplete (as SRT), actual LVT policies do not vary as much as might be desired in empirical analysis, implementation is complicated by inaccurate assessments, and any one location only has one policy treatment. Lab experiments with induced values overcome all of these shortcomings, providing the researcher with control and allowing analyses that test causal hypotheses. Furthermore, compared with simulations, the experiment also allows for behavior to deviate from the simulated optimal, which enables the researchers to observe system dynamics arising from the propagation of suboptimal decisions made by humans. Experiments also have limitations, especially in that they are simplifications of real-world complexity. However, they offer evidence unavailable from empirical or theoretical research and thus can help planners and policy makers better

4 understand the relative effectiveness of different policy options.

This experiment allows participants to make land investment decisions under LVT, SRT, and UPT tax regimes. Moreover, the choice of tax regime becomes endogenous in the experiment through a simple voting mechanism treatment. The constraints of laboratory experiments limit the voting mechanisms, though there is a more complex theoretical literature on public support for different innovative taxes (c.f., De Borger and Glazer 2017). Our voting treatment simplifies the voting process as a simple referendum on the tax to be employed in one treatment. Using heterogeneous induced values, the experiment examines how different groups behave in terms of landed wealth and income wealth. The experiment will reveal what types of groups win from LVT and what groups lose. It also reveals whether winners and losers vote in their best interest or with respect to other concerns, such as equity. Thus, the research helps understand the acceptability of LVT and SRT relative to UPT in a controlled setting.

2. Theory

This section develops a theoretical model that supports an experiment comparing different land and improvement tax institutions. The experiment is contextually framed because it includes a positionality treatment. There are two complex, dynamic interdependencies among landowner decisions and a third positional-good aspect that warrant the use of behavioral economics.

Induced values derive from a real-world setting, which is described along with the economic model and which trades off improvement and consumption. Real world data also inform the derivation of revenue neutral tax treatments. A subsection proposes an optimal improvement strategy for each treatment, which is implemented in a simulation to derive predictions.

2.1. Economic Model

5 The parameterization begins with a real world setting, that is, experiment participants make decisions and face incentives in “experimental dollars” in the role of a representative property owner in Harrisburg, Pennsylvania. Each utility maximizing owner i, where i ϵ {1, 2, 3,

4, 5}, allocates period t income to investment in property improvements, xit, and a normalized consumption good, yit. The consumption good reflects all other uses of income, including savings and is assumed to be net of taxes and a present value.1 For simplicity, prices are normalized to $1 for both items. The Cobb-Douglas utility function represents a change in monetized utility:

a 1-a (1) Uit = xit yit

In the experiment, Uit was termed “utility payoff” or “earnings” for the participants, so we follow that terminology. Assigning five values of an improvement-preference parameter, a=(0.2, 0.25,

0.3, 0.7, 0.8), to each of the five members in a neighborhood induces the first level of heterogeneity. The parameterization was selected to achieve the desired voting predictions described below.

Another level of heterogeneity is introduced by the experiment’s three neighborhoods

(k=L, M, H indicating low, middle, or high incomes and housing values). The two different heterogeneity levels result in 3x5=15 different types of participants in each experiment session.

k k k k Census data parameterizes income, Iit , and property value, PVit = LVit + IVit . The components

k k of property values are land value, LVit , and improvement value, IVit , which were apportioned

k k k using Harrisburg data. Observed data thereby define three initial parameters (Ii0 , LVi0 , IVi0 ), which are different for each neighborhood k. Harrisburg uses SRT with a 6.0 ratio, where tax on land is 28.67 mills and on improvements is 4.78 mills (Bourassa 2009). Table 1 describes income and housing data. Model dynamics are simplified and do not involve explicit

6 discounting so that experiment participants could better understand their choice environment.

The variable xit represents a decision in period t that delivers benefits into the future. For instance, if an owner adds a bedroom, then that bedroom provides a stream of benefits into the future. This benefit stream has a present value to the owner represented by the improvement choice raised to the a parameter.2 Improvements made in any given period are accounted for using a measure of an owner’s improvement value, which mirrors housing quality or size:

k k k (2) IVit = IVit-1 + xit

Because choices (xit, yit) exhaust net income, the choice of xit determines the other choice yit via period t’s budget. This approach isolates the improvement choice for the experiment participant.

2.2. Interdependencies

The simple dynamics of the model allow participants to focus on the behavioral economics interdependencies of interest.3 The two mechanisms of interdependency are a pecuniary capitalization externality, g, described here and a second interdependency of tax redistribution, described in the next subsection. First, the internal effect of improvement choice (equation 2) also triggers an external neighborhood impact. Improvements to one property capitalize in neighbors’ land values because the neighborhood is now “nicer”:

k k k (3) LVit = LVit-1 + g∑ixit

Within the experiment, this effect was termed the “nicer neighborhood effect”—an effect in real estate (“location, location, location”) that is well known but was not found in our review of the existing LVT literature. Similarly, no estimates for g were found in a review of hedonic literature. The positive pecuniary externality parameter, g, is a small number (this experiment used g=0.05). Because the model has no salvage value, this unearned increase in one’s land value cannot be recouped at the time of sale—a reasonable assumption when one buys in the

7 same market into which one sells. In sum, the “nicer neighborhood” results in a negative externality because of increased land taxes.

The internal and external processes (equations 2 and 3) of land improvement provide for a complex interdependency. One participant improving property provides that participant with a

k benefit (via the utility function), but the neighbors only bear a cost (via the tax on LVit ) from that

k decision. Of course, the decision maker also bears some of the increased tax cost via LVit and

k IVit . Because all the neighbors follow this process, equations (1), (2), and (3) have the characteristic of a multilateral externality. Dynamically, the externality in the form of the tax on

k LVit continues to accrue in each period, mimicking the artifact of taxing unrealized capital gains noted in Bourassa (2009). Obviously, these externalities are expected to accrue differently under different tax treatments. LVT ought to provide a greater degree of interdependency because all

k the tax will fall on the LVit measure rather than only a portion of the tax under UPT. The impact of SRT ought to fall between the extremes of LVT and UPT.

2.3. Tax Institutions

The hypotheses test whether improvements change under different tax institutions. Let the UPT

UPT rate be τ0 on both land and improvement. This means that UPT tax revenue will be Tt =

SRT ∑k∑iτ0(LVit+ IVit). SRT taxes τL on land and τI on improvement, so Tt = ∑k∑i(τLLVit + τIIVit).

LVT 4 LVT only taxes land, τLL, so Tt = ∑k∑i(τLLLVit). Revenue neutrality, which is set at t=0, allows for a controlled test of tax institutions that is comparable across all periods: ∑k∑iτ0(LVi0+

IVi0) = ∑k∑i(τLLVi0+ τIIVi0) = ∑k∑i(τLLLVi0), where the initial values are reported in table 1. To solve, assume β*LVi0=IVi0, which simplifies to the following revenue neutrality condition:

(4) �!(1 + �) = (�! + ��!) = �!!

Equation 4 is uniquely solved using Harrisburg’s SRT rates, τL =0.02867, τI =0.00478 (Bourassa

8 2009) and β=3.503. The other revenue neutral tax rates are: �!!=0.045414 and τ0=0.010085.

During experimental design, the researchers found that the model took too many periods to produce salient treatment effects, so these revenue-neutral rates were multiplied by 10:

* * * * τL =0.2867, τI =0.0478, τLL =0.45414, τ0 =0.10085.

Tax return must be used to maintain revenue neutrality as improvement values rise because tax revenue also increases. All tax revenue above the initial baseline tax of $149,409

(calculated from the parameter’s initial values) is returned in equal shares to each participant at the end of the same period in which the tax is collected:

z z (5) TRt = (Tt - 149,409)/15, for z = UPT, SRT, LVT.

This equal return of excess revenue is the second major interdependency—and a major assumption—in the model. The redistribution is progressive because the higher-income neighborhoods will have higher levels of improvements and thus pay more tax but only receive the average share in return above the level in equation (5). This interdependency links the decisions of all 15 types in the model in contrast to the externality interdependency, which only affected the 5 neighbors within each neighborhood.

2.4. Approach to Modeling Behavior

Figure 1 explains the comparative statics of myopic optimal choice, (xit, yit) under different treatments. The five preference parameters, a, determine indifference curve shapes. The initial budget constraint’s slope comes from a tax treatment and intercepts from a given income level.

A new tax affects optimal choice in two ways. There is a substitution effect when the budget constraint rotates because the new rate changes the price of improvements and produces an income effect when the tax alters the net-income level. The income and substitution effects can move in different directions; thus optimal improvement may rise or fall under different

9 treatments. For instance, in figure 1, panel a, the LVT treatment produces more income and a lower price on improvements, resulting in an optimal increase improvements and general consumption. Panel b mimics panel a, except for a negative income effect; in this case improvement choices increase but general consumption falls. The income effect could be so great that it overwhelms the change in improvement price; for instance, panel c has a drop in both improvements and general consumption.

Although dynamic versions of the owner’s problem in equation 1 could be modeled in various ways, the researchers choose a simple myopic solution within any period, ṯ, to characterize optimality.5 The constraints of the experiment programming made this into a two- step process. In the first step, the objective function for the UPT treatment is:

a 1-a k k k (6) max Uiṯ=xiṯ yiṯ s.t. yiṯ=Iiṯ -τ0(LViṯ +IViṯ )-xiṯ !!!

k k k k k The related constraints for SRT and LVT are yiṯ=Iiṯ - (τLLViṯ +τIIViṯ )-xiṯ and yiṯ=Iiṯ - τLLLViṯ -xiṯ, respectively. Substitution reveals an unconstrained maximization problem, which solves as the following optimal choices under the three treatments:

UPT* ! k k * k (7) xiṯ = [Iiṯ -τ0(LViṯ-1 +gΣi x-iṯ +IViṯ-1 )] (!!!!)

SRT* ! k k * k (8) xiṯ = [Iiṯ -τL(LViṯ-1 +gΣi x-iṯ )- τIIViṯ-1 ] (!!!!)

LVT* k k * (9) xiṯ =a[Iiṯ -τLL(LViṯ-1 +gΣi x-iṯ )]

A second step occurs once myopic decisions have been made. The z-Tree program (Fischbacher,

z 2007) then calculates the tax return under each regime, z, TRiṯ as in equation (5) within any period, ṯ, but after that period’s improvement choices were made. Then, tax return is added into earnings as a consumption good, in effect augmenting earnings by augmenting consumption.6

The approach means that the “optimal” choices in equations (7-9) are not ex post efficient because they result in more-than-optimal consumption when tax return is positive, resulting in an

10 efficiency criterion that is approximately myopically optimal (herein, “optimal”). It is important to note that this approach derives the simulated optimal but does not constrain the choices of experiment participants.7

2.5. Simulated Optimal Data

The researchers used z-Tree (Fischbacher, 2007) to design and simulate the experiment, repeatedly making adjustments to ensure salient treatment effects. The final simulation of optimal choices provides baseline predictions for the final version of the experiment. Simulation

k,z* analysis reveals the optimal improvement choice in each period, type, and treatment: xiṯ , for i=1,…,5; k=L, M, H; t=1,…,5; z=UPT, SRT, LVT.

Figure 3 displays the optimal choices over five periods for the UPT and LVT treatments, not displaying SRT because it falls in between the other treatments. Figure 3 shows that optimal investment decreases under both tax regimes, though the slopes differ. Unsurprisingly, owners with higher preferences (a=0.7, 0.8) choose higher improvement than owners with lower-a values (a=0.2, 0.25, 0.3). Under LVT, all owners initially increase improvement levels, but then owners with higher-a values reduce improvement levels faster than owners with lower-a values.

This finding holds across the different neighborhoods. In addition, the three lower-a types reduce improvement levels more under LVT than under UPT—eventually, this level would be lower in LVT than UPT for these types. However, the two higher-a types consistently improve more under LVT than UPT. Overall, the change to the LVT treatment has a much bigger effect on owners with higher a value.

k,z* The simulation also reveals the evolution of property values, PViṯ , and period-specific

k,z* individual earnings, Uiṯ . This earnings measure can be aggregated into a welfare measure by

k,z* k,z* type over five-period treatments (Ui =ΣṯUiṯ ) and the community-level welfare

11 z* k,z* (U =ΣkΣiΣṯUiṯ ). Figure 4 shows that the earnings are higher for owners with higher-a values in LVT than UPT, while the opposite is true for owners with lower-a values. This figure also shows that the earnings of some low-income types are rising because of redistribution. Under

LVT, the low-income types are almost uniformly better off over time. The mid-income range is stable in terms of welfare. It is the high-income types that almost always get worse off over time because of the progressive tax redistribution.

The simulation results support the land-tax acceptability, i.e., voting, expectations. The experiment was, first, designed be a potential Pareto Improvement (U-UPT* > U UPT*). This required that the winners from the land tax (the two high-improvement-preference types) win more than the losers lose (three low-improvement-preference types). In optimal simulations, the researchers found community earnings were 1 percent greater under LVT than UPT and 0.5 percent greater under SRT than UPT. Second, the experiment was not designed to be a Pareto

Improvement. Indeed, three of five types in each neighborhood would have higher earnings in

k,UPT* k,-UPT* k,-UPT* k, UPT* UPT: UPiṯ > UPiṯ , for a=0.2, 0.25, 0.3, for all t; UPiṯ > UPiṯ , for a=0.7, 0.8, for all t. If owners vote in line with their earnings, then these conditions predict a vote in the first period of a land tax treatment to be nine to six in favor of abandoning a land tax in favor of UPT.

3. Experiment Data and Hypotheses

This section derives an experimental design from the model, explains the data, and offers hypotheses. Fifteen tablet computers were linked to an administrator computer using z-Tree software (Fischbacher 2007) at the University of Delaware Center for Experimental and Applied

Economics. In each period, experiment participants made one improvement choice. The researchers presented a calculation aid and market information on the participant’s decision

12 screen because this decision was complicated by the mathematics of solving a nonlinear objective function, interdependencies with other participants’ decisions, expectations about a tax return, a lack of information about other participants’ contemporaneous decisions, and the challenge of optimizing dynamically. The calculation aid was a table of 15 possible improvement choices, each of which would result in a corresponding general consumption decision and a level of earnings (i.e., cash earnings in experimental dollars).8

Student participants were largely undergraduate business and economics majors, though some engineering and environmental social science majors were recruited when sessions were difficult to fill. The University of Delaware Institutional Review Board approved the protocol and informed consent was obtained. Participants completed informed consent, read paper instructions, watched a recorded instructional presentation, asked questions, and then were trained over a two-period UPT practice session. After training, participants completed three different treatments, making 15 improvement choices and up to 4 voting choices. New instructions were distributed with every treatment. At the end of the experiment, participants completed a paper survey and were paid. Each session lasted 1.5-2 hours. Average earnings for one participant were $17.53-$18.82, but the individual earnings varied from $14.00 to $20.50.

Earning potential varied significantly with the induced values, so participants rotated through types systematically to produce approximately equal expected earnings—also, this is the reason why the researchers could not pay a random round. The experimental dollars were converted to

U.S. dollars at a rate of 7,900:$1.

All 120 participants in eight sessions started with a UPT treatment—termed “tax plan

1”—over five periods because this matches the status quo for almost all U.S. property taxation.

Then, all participants had a land tax treatment (either SRT or LVT) termed “tax plan 2.” A

13 between-subjects design was used because a within-subjects, full-factorial design lasted too long

(well over two hours) during pretesting. Six sessions had LVT for five periods, while the two other sessions had SRT treatments. Both land tax treatments started with one period of UPT, and then the experiment administrator announced that the tax regime was changing from tax plan 1 to

2. Choices under a land tax were then made for four more periods.

The third treatment repeated the second, except after every improvement choice under a land tax (periods 2-5), the participants would vote whether to continue under a land tax or revert to UPT (voting treatment: Vote=1). To aid in this decision, participants were told what their current tax was under the land tax and what it would have been in that period under the alternate tax institution, UPT. This allowed participants to make an informed comparison of the political economy of the land tax and UPT treatments. At the end of each period, votes were tallied and announced. If eight or more voted against the land tax, then the remaining periods would be conducted under UPT and further voting would be suspended. Voting rules were explained to participants with written instructions and an oral presentation before beginning the treatment.

The researchers anticipated two key drivers of this voting decision. First, participants ought to vote with self-interest, selecting the tax institution that provides them with the highest earnings. The second driver was more challenging to design; some participants may believe the land taxes cause them to be positional-good losers or are somehow unfair. The positional good objection relates to prior findings (Bourassa, 2009, p. 195-6) that land taxes are rejected because homeowners believe should not pay the same or near the same property tax as a neighbor, who has a larger house or more improvements. The researchers sought to include this effect, that is, a trigger for positional-good jealously or sanction on deviation from norms. The final experiment designed this element using a graphical representation of housing size (figure 2).

14 Figure 2 graphically displays 15 participant houses. Each house in the same neighborhood started each treatment (figure 2, panel 1) with the same initial improvement (blue bar) and land (red bar) values. Bar height indicates value. The improvement bar increased with each participant’s choice (equation 2), while the land value bar increased with the capitalized neighborhood externality (equation 3)—i.e., was the same value for all five in the neighborhood.

The experiment had one lab assistant whose sole responsibility was, within 20-30 seconds, to use a spreadsheet to process the 15 choices made each period and construct a corresponding graph of the evolved neighborhood; the lab administrator would then indicate the display before starting the next period. The purpose of this graph was to communicate the positional-goods elements of housing. To test the impact of this positional-good graph (PG-Graph), four LVT and two SRT sessions were run with the graph and two LVT sessions were run identically without a graph.

UPT=1 SRT=1 LVT=1

Vote=1 I II III PG-Graph=1 n.a. Vote=1 IV V VI PG-Graph=0 n.a. n.a. Vote=0 VII VIII IX PG-Graph=1 Vote=0 X XI XII PG-Graph=0 n.a.

The experimental design is in the box above1. A fractional design was used—collecting

1For the referees: This paper includes the entire design—i.e., all treatments proposed and examined in the experiment—to be allow for a full understanding of the choices made during the research and to not systematically select the hypothesis tests reported. While running PG-Graph sessions, the researchers realized that the graph was having a significant impact. This was unexpected and warranted additional attention. Moreover, the SRT sessions were less interesting than expected. So, the researchers added more sessions without a PG-Graph to estimate the

15 no data on treatments I, IV, V, and VIII—because of budget limitations, a belief that the LVT treatments would produce stronger effects than SRT, and the incoherence of a voting treatment for UPT given that the voting treatment would be from a land tax to UPT. In addition, initial treatments only used PG-Graph, but additional sessions were run three months later without the

PG-Graph after seminar feedback on an initial presentation of results. Four sessions had the following treatments (VII, IX, III), while two sessions had (VII, VIII, II) and two sessions had

(X, XII, VI). Four treatment combinations are not identified (I, IV, V, and XI).

The structure of the data mirrors the simulated data. Improvement choices (NUPT, PG-Graph

=300, NUPT,- PG-Graph =300) are observed in eight sessions under the initial UPT treatment, with

k,UPT 120 participants making 5 decisions each: xiṯ . Then, the 90 participants in six LVT

LVT k,LVT LVT,PG-Graph LVT,-PG-Graph treatments produce N =360 choices of xit (N =240, N =120), while 30

SRT,PG-Graph k,SRT participants in two SRT treatments produce N =120 choices of xit . To reinforce the effect of tax regime change, all participants in the LVT or SRT sessions start the first voting- treatment decision under UPT and switch to LVT or SRT with an emphasized announcement

k,UPT UPT, PG-Graph from the lab administrator. This setting generates extra observations on xiṯ (N

=90, NUPT,- PG-Graph =30). Finally, the voting treatments produced another NVote=600 plus 480

k,Z possible votes: vit , for i=1,…,5; k=L,M,H; t=2,3,4,5. However, if the votes in period 2, 3, or 4 were in favor of UPT, then any subsequent votes were not observed because the tax regime irrevocably switched back to UPT. The land tax was rejected by voters in two sessions, so the final voting-treatment data had 448 observations and 600 improvement choices (NUPT,Vote,PG-

effect of showing the graph information. To stay within budget, the sessions with SRT treatments were dropped. Hence, there is not a full-factorial design. Although the regressions will explain these treatment effects in a controlled session and the full factorial design is not needed for our data analysis, the researchers want to be explicit with referees about the research process. If the referees have strong feelings about this, the researchers could potentially remove the SRT treatments from the paper/regressions and just clarify the design choices in a footnote.

16 Graph=180, NLVT,Vote,PG-Graph=75, NLVT,Vote,-PG-Graph=240, NSRT,Vote,PG-Graph=105). All (Nz=1800)

k,Z improvement choices above have corresponding earnings measures (Uit ).

The simulation results provide the hypotheses for the experimental data. Table 2 presents the most important hypotheses, along with the experiment results. Summarizing, the hypotheses anticipate that LVT produces 1% greater wealth than UPT, while SRT produces 0.5% greater wealth. In the voting treatment, researchers expect a land tax to be rejected in period 2. The researchers also anticipate that the PG-Graph treatments will have no differential effect because the graph does not affect earnings.

4. Results

Key experiment results diverged from simulated expectations, revealing that this behavioral economics setting—even with student participants—offers insights about challenges to land tax efficiency and acceptability. This section first demonstrates unexpected inefficiencies in land tax treatments at the community level and in terms of property values. Then, the results are examined at the individual behavioral level, including voting treatments, to explain why land tax tended to be less efficient than expected. Positional-good issues are revealed to be a key driver of land tax inefficiency.

4.1. Results on Social Welfare and Community Characteristics

Table 3 reports the community welfare by treatment for all fifteen participants by session (UZ) relative to the simulated optimal as an approximate efficiency frontier. The results on the UPT treatments show that several sessions (6-8) were close to 100% efficient, while the other sessions

(1-5) were slightly less efficient—in the 94.3-95.9% range. The land tax treatment results are surprising in that no LVT session was efficient (between 89.3 and 96.1%) and yet the SRT sessions were much closer to the efficiency frontier. The results on average efficiency suggest

17 that the positional-good graph affected the efficiency of LVT, which averaged 92.6% efficient with the graph and 94.7% efficient without the graph.

The efficiency results can also be compared within sessions, by treatment. Surprisingly,

LVT outperformed UPT in only 33% of sessions, even though LVT was biased toward efficiency because it was designed to be 1 percent more efficient. Similarly, SRT outperformed

UPT in 50% of sessions despite the 0.5% efficiency bias. These experimental data thus do not consistently support the hypothesis on aggregate welfare, which was derived from the simulation, that land taxes are more efficient. The lack of land tax efficiency is an unanticipated result and a principal finding of this paper.

The welfare results also can be examined by averaging over type across sessions. Table 4 presents the simulated and experimental results comparing LVT and UPT (six sessions). Within treatments, the experimental earnings are almost always lower than the simulated earnings; the two exceptions are that type 2 had higher earnings than predicted under UPT, while type 14 earned more than predicted under LVT. The lower earnings levels for the other types match the aggregate results in table 3. That the earnings for almost all types were lower in the experiment suggests that the observed inefficiency of LVT is a robust result.

When the experimental treatment effect (from UPT to LVT) is compared to the predicted effect, one observes several expected patterns. First, the parameters predicted that the low- preference types (1-3, 6-8, 11-13) would have lower earnings under LVT, and this was observed for all types but type 12. Second, the high-preference types (4, 5, 9, 10, 14, 15) were predicted to have higher earnings under LVT and this was observed in the experiment. Although some anticipated patterns were observed, the magnitudes of the earnings changes from LVT in the experiment were much larger for the high-income types. Under LVT, the collective earnings of

18 the three low-preference types (1, 2, 3) was 12.9% lower rather than the anticipated loss of 4.1%.

In contrast, the collective earnings of the high-preference types (4, 5) were 13.9% higher rather than the anticipated 9.3% increase. In the middle-income neighborhood the same pattern exists for the low-preference types (observed lower earning of 8.2% rather than 3.8%), but not for the high-preference types (observed higher earnings of 8.9%, similar to the expected 9.6%). The low-income types with low-preference had a similar result collectively (4.4% observed loss versus 3.5% expected loss), but the pattern was asymmetric with two types (11, 13) with much lower observed earnings and the aforementioned type 12 with higher observed earnings. The high-preference types in the low-income neighborhood earned more (12.9%) collectively in LVT than the expected level (9.9%). In sum, the experiment results suggest that those owners with a low preference for improvements earn considerably less than expected under LVT. Many of the high-preference types earn considerably more under LVT. Below, a regression explores this unexpected phenomenon, which finds systematic overinvestment by the low-preference types.

The final set of community level results (figure 5) shows that property values are higher for 14 of 15 types under LVT than UPT. This is expected due to the lack of distortion in LVT on improvement investments, which encourages more improvements and thus higher property values. In addition, the treatment effect on property values tends to be largest for the high- income neighborhood; that is, the gray bar has the largest positive difference from the corresponding black bar. So, despite the progressive redistribution, LVT caused the high-income owners to invest more in their housing than under UPT. Although LVT results in higher property values (figure 5) for all in the high-income neighborhood, those same owners with low- preferences actually had much lower earnings in the experiment (preceding discussion of table

4). This result arises from overinvestment (see below) and from the institution used to

19 redistribute the extra tax collected under LVT from the extra improvements. LVT exacerbates both these effects, causing greater losses for more of the low-preference types in all neighborhoods but also attenuating this effect on low-income types via the progressivity of redistribution. In consequence, the low-income owners have slightly higher property values but an only slight different earnings effect in absolute value than their corresponding types in the high-income neighborhood.

4.2. Results on Investment Behavior

Figure 6 shows the simulated and the experiment improvement behavior results averaged by type and period under LVT and UPT. As expected, almost all types increased improvements under

LVT (the dashed black line is above the solid black line). Compared across owners within each neighborhood (examining one row at a time), owners with higher a value (a=0.7, a=0.8) chose more improvements compared to owners with lower a value (a=0.2, 0.25, 0.3). Furthermore, the experimental data show some behavioral trends from the simulation. For instance, under LVT, all owners initially increased their improvement choice, but then reduced it in the following periods. However, the experiment results (black lines in figure 6) do not overlap with the simulated expectations (red lines) for all owners. Figure 6 shows that there are deviations, which tend to be largest among the types with lower a values. The results suggest that the experiment produces some positive deviations for the UPT treatment (solid black lines), but the deviations are generally largest under LVT. Indeed, among these nine low-improvement-preference types,

LVT encouraged much more improvements than expected in seven to eight of the types.

Although over-improving was not completely unexpected given that some participants were treated with a positional-good graph, the magnitude of over-improvements was large and there is survey evidence that the graph made a difference.

20 Table 5 presents a set of linear regressions of behavior. The first model is pooled

(Nz=1800) and shows that the induced values explain much of the variation. All else equal, the land tax treatments increase improvements by 1,341 in LVT and by 521 in SRT relative to UPT, which is expected because the tax rates on this decision are lower. Because the voting treatment always was the final treatment, introducing a possible order effect, related regressions were estimated with (NVote=600) and without N-Vote=1,200 voting. The nonvoting treatments show a similar land tax effect as in the pooled model. In addition, a clear pattern emerges with the period variable; relative to the first period, point estimate improvements consistently drop in each successive period. This matches the pattern of black lines in figure 6. Falling average improvements over time likely arises from: (1) the increased taxes of a progressively larger house despite constant income; and (2) the high income types who are incentivized to do the most improvements but also pay the most taxes and suffer the most from the progressive excess tax transfer. The researchers also were concerned that there would be an effect from the first period UPT tax in a land tax treatment and a final period effect, given that the game ends in period 5. As such, models were estimated without period 1 and without periods 1 and 5 (the final columns of table 5). The LVT treatment is robust. The SRT treatment generally consistent, though it is more difficult to discern as the quantity of data drops because of suspected multicollinearity. In sum, the behavioral regressions show that the land tax treatments can deliver greater improvements than UPT.

At the conclusion of the experiment, the researchers administered an anonymous survey

(via paper and unfortunately un-linkable with the decision data). The 90 respondents in sessions with a positional-good graph reported that they paid attention to the data in the myopic optimal table (93 percent). However, many were also considering their relative position. Indeed, 98

21 percent reported that they noticed the positional-good graph and 48 percent claimed that their decisions were affected by information in this graph. Further, 29 percent reported that the graph had “a great deal of influence” on their decision, while 48 percent reported that it had “a little influence.” Only 23 percent reported that it had no influence.

That the positional-good graph had any effect was somewhat surprising in that it was nonbinding and not related to the earnings. The asymmetry in housing size was induced, and any participant that deviated from the optimal choice would earn less money. The overinvestment by those types might be explained by three possible reasons. First, over-improvement suggests that this experiment was picking up some position-good-type utility. When participants observe the different property values or housing-size from the graph, they might chose to over-improve to catch up or become more in-line with the other owners in their neighborhood. This is problematic for analysis because it suggests that the participants are following some unobserved, or meta-, utility function that differs from the function that drives the earnings; of course, this is a valuable insight because it also applies to empirical work on housing investment and research on positionality. Second, overinvestment may be a type of order effect. This experiment always had a first treatment of UPT and then each land tax treatment started with a UPT period; this was done to mimic reality, but it highlighted for participants the change from a treatment where improvements were taxes to one in which they were taxed at a lower rate. The focus on this change may have unduly influenced some participants to overinvest even when it lowered their earnings. Although this is an important qualification, the reality is that almost all owners in the real world would face the same change in tax treatments from UPT to a land tax. Third, overinvestment could be simply a result of mistakes—albeit with a systematic tendency to overinvest among low-improvement preference types. Regardless of the reason, overinvestment

22 should propagate more in LVT than UPT because there is a greater incentive to improve in LVT and, because of capitalization, more deviations will lead to greater neighborhood-wide deviations from the simulated optimal path.

4.3. Acceptability of the Land Tax: Voting Results

The voting treatment results in table 6 show that LVT is more acceptable than expected, but as this subsection argues, this unexpectedly high acceptability is due in part to overall suboptimal performance relative to the simulated optimal behavior. Nine participants were predicted to vote against LVT or SRT in period 2 (the first possible voting period), and then the game would revert to UPT in the third period. However, in the experiment, only two sessions voted against a land tax in any period: (1) One PG-Graph session voted against SRT in period 4; (2) One non-

PG-Graph session voted against LVT in period 2. In six other sessions, the land tax continued despite the prediction that it would lower earnings for a majority of the community. Because of the unexpectedly low votes against SRT and LVT, the researchers recalculated the expected votes in each period by examining whether UPT or the land tax treatment would make each type better off conditional upon any suboptimal deviations in the current and prior periods. In other words, suboptimal deviations (say, from overinvestment) would require an updating of the optimal choice and thus predicted votes in subsequent periods.

The updated rational number of votes (URV) is reported in table 6, and it was no longer necessarily optimal to vote against SRT and LVT. Three sessions had rational voting outcomes, where a land tax was maintained; new optimal paths in sessions 1, 4, and 8 resulted in URV for the community such that it was never optimal for a majority to vote against the land tax—and the experiment participants closely followed this updated expectation. Yet, in LVT session 1 and 4, the optimal paths were far from efficiency—table 3 showed that total welfare was 5-6% lower

23 than its corresponding UPT treatment. The deviation in the SRT session (#8), however, followed a different pattern in which both SRT and UPT sessions were close to the efficiency frontier. In sum, these three sessions had deviations from the optimal path that switched the community’s optimal taxation profile, and the participants voted in line with their updated interests.

At the other extreme, in session 5 the participants voted against LVT in period 2, as originally predicted. This session was the closest to expectations in terms of the voting pattern and it was relatively efficient with LVT only 0.5% less efficient than UPT; however, this session used no positional graph. Voters also rejected SRT as expected in session 7, but this rejection did not occur until period 4. Session 7 was comparatively efficient. Of the remaining three sessions, session 3 and 6 were surprising in that the URV was 9 votes in almost all periods, but the votes remained in the 4 to 7 range. These two sessions, therefore, saw substantial voting against one’s interests.

4.4. Statistical Analysis of “Deviations” from Optimal Behavior

The behavioral deviations suggested by figure 6 are statistically analyzed in this subsection, and it will be shown that even when one accounts for updated optimal behavior in successive periods, the pattern of deviation persists. By updated optimal behavior, the researchers are referring to the same process as used for updating rational voting—i.e., solving for myopic optimality based upon the new property characteristics (as in equations 2 and 3), which resulted from prior-period, potentially suboptimal decisions. Deviations in each period are calculated from this new updated optimal myopic choice. Figure 7 plots the raw deviation data. Positive deviations indicate over- investment, which would be consistent with either a mistake or the positional good explanation of behavior. Underinvestment or negative deviations come from several different possible reasons: (1) some participants may have been embarrassed by having a large “house”; (2)

24 participants may have taken a higher tax bill in the future into consideration and tried to reduce tax bill by reducing improvements; or (3) some participants may have made a mistake.

There were more positive than negative deviations, suggesting the possibility of positional-good effects. The average improvement choice over 1,800 decisions was 7,230, so a deviation of 300 indicates a 4.1% change from the updated optimal. Figure 7 data show that

30% of decisions deviated 300 or more in a positive direction, while only 16% deviated negatively. This pattern holds for smaller deviations; for instance, 38% of decisions deviated

100 or more in a positive direction, while only 22% deviated negatively.

Two regressions (table 7) analyze the deviation data from figure 7. The first column shows the pooled regression (Nz=1800), which has two key results on the LVT treatment. First,

LVT led to a $450 more improvements relative to UPT, all else equal. This is an “over- improvement” because it deviates positively from the updated optimal. Second, when there is a positional goods graph, then LVT led to an additional $530 over-improvement relative to UPT.

The pooled model also shows that participants tend to have less deviation ($459 less) in the last period, suggesting that the final period did indeed have a meaningful impact of behavior. As such, the researchers estimated several additional models to remove the last period impact.

These models also addressed multicollinearity arising from the fractional design and limited data on the SRT treatment. The researchers propose that the purest test would be to compare UPT and LVT treatments, removing first and final periods as is done in many experimental papers to remove period effects (column 2 of table 7). This test shows that the interaction of LVT and PG-

Graph is statistically significant, resulting in a $915 over-improvement relative to UPT and LVT without the positional goods graph. This is the key result of the paper—that LVT allows the positionality of housing to propagate rapidly into an inefficient outcome.

25 The results also show that the low improvement preference types tend to deviate toward over-improvement in all treatments; the researchers estimated a prior model that interacted the preference-types with PG-Graph, however, only those with preference parameter a=0.25 had a significant interaction and the model had multicollinearity. Thus, the researchers have incomplete evidence about whether those participants induced to have the smallest houses were motivated by the positional-goods graph to over-invest.

The researchers also explored the effect of deviation of earnings to better understand the cost of deviating and to check whether the complexity of interactions allowed some participants to somehow outperform the simulated optimal strategy in the complex, interdependent system— perhaps through cooperative or coordinated behavior. Table 8 reports an explanation of participant earnings by deviation and other control variables. The results of the main treatment effects suggest that a deviation will lower earnings. A $1 deviation in improvement choice leads to a $0.07 decrease in participant earnings. The LVT treatment control is not significant, by itself, in any model. When this simple model is augmented with interactions for LVT and SRT and then again with controls on the induced values, a surprising pattern emerges. LVT treatment interacts with deviation such that there are lower earnings for participants who deviate in the

LVT treatment—but not in the UPT treatment. A $1 deviation leads to a $0.41 loss in earnings in LVT relative to UPT. Although the results suggest deviations strictly lower earnings in the

LVT treatment, the aforementioned limited SRT data likely cause multicollinearity and may mask a similar SRT effect. Under LVT, a $1 deviation of over-investment will lower earnings by $0.27-$0.41. Therefore, over-improvements are costly to the participant or, conversely, under-improvements would lead to larger earnings.

When one combines the principal results of tables 5, 7, and 89, one sees that participants

26 respond to the positionality of housing by over-investing, but the effect only shows up under

LVT. Over-investing is most likely to occur among the low preference types, who were induced in the experiment to have small houses. These over-investment decisions lead to lower earnings and therefore indicate a willingness to trade-off cash with a “house” that more closely approximates the neighbors’ housing size.

5. Discussion and Conclusion

The experimental results produced unexpected findings with respect to the simulated optimal predictions of the model. LVT and SRT did not produce the most efficient outcome in half of the treatments, despite a design where the UPT treatments were supposed to generate slightly less welfare than LVT and SRT. The principal driver of this unexpected result is the systematic tendency to over-improve in LVT, among low-improvement types, and when a positional-goods graph is shown. All these effects are processed through the complex system of the z-Tree interface (Fischbacher 2007), where excess taxes are redistributed and there is a capitalized interdependency for improved neighborhoods.

In the voting treatments, participants did not reject LVT and SRT as often as expected.

But this voting behavior was more “rational” than the design predicted because deviations from myopic optimality in early periods altered the expected performance and efficiency of the tax treatments. Obviously, some of the results are artifacts of the model, experiment design, and induced values. Yet the behavioral economics method offers some potentially generalizable insights that warrant a new avenue of research and policy consideration.

One way to think about the tendency for the low-improvement-preference types to over- invest is that improvements are “free” from tax under LVT and “partially free” under SRT, so

27 there is less of a tax burden from improvement to be borne by the decision maker. However, there is more of a burden to the neighbors through the capitalization externality because there will be more improvements. In other words, the externality is exacerbated in the LVT and SRT treatments. Ceteris paribus, the greater capitalization interdependency means that when an owner makes a suboptimal decision in LVT and SRT, it seems to propagate as a mistake or cost to social welfare more than under UPT. The results therefore suggest that one might expect that real-world instances of LVT may not delivering on the expected promises because of this unforeseen error propagation, which LVT exacerbates relative to UPT—a possible answer to

Fischel (2016) about why LVT is not more common. This further may imply that effective use of LVT may pose a special challenge—one in which homeowners must make optimal investment decisions. Optimal decisions require high quality information about homeowner’s own utility/profit function and neighborhood characteristics. The results also suggest that informational challenges to LVT are more than just associated with the assessment process.

The results also showed that LVT might exacerbate the positional-goods characteristic of housing. Indeed, this experiment with undergraduate participants was able to produce positionality by using a relative-position graph treatment. This result has important implications because—if as Frank, Levine, and Dijk (2014) suggest—positional-good improvements cause an expenditure cascade, then economists must rethink whether they really believe a land tax will increase welfare. For, it is the land tax that allows improvements to be made without distortions.

Obviously, not all land improvements are the same. For instance, improvements made to substandard or unsafe housing would not have welfare-lowering, positional-good characteristics.

Or, improvements to commercial land, where profit-maximization guides decisions, would not have positional-good characteristics. As this work is the first experimental economics

28 exploration of LVT, future experiments could seek a more general specification of a utility function—one which could potentially capture positionality.

A common question in LVT research is whether the tax is regressive. This research used a built-in mechanism—equal tax redistribution above a revenue-neutral threshold—to predetermine that LVT would be progressive. So, it is important for policy makers to see how

LVT can be assured to be progressive. But the mechanism used also introduced a new dimension of interdependency. The model and experiment reported here did not test for alternate mechanisms to return excess tax revenue. Future experiments ought to explore other mechanisms and, because of the interdependency, test for interactions with the tax treatments and other characteristics of the design.

One of the motivating aspects of LVT unacceptability was that it was a tax on unrealized capital gains (Bourassa 2009), and that this is very unpopular with many homeowners. Our research shows how LVT exacerbates a fundamental capitalization process—wherein LVT leads to more improvements which leads to nicer neighborhoods which leads to more unrealized capital gains which leads to more LVT acceptability challenges. This paper contributes to the

LVT literature by proposing and then explicitly modeling this process. A possible policy solution to this challenge may involve using the efficiency gains of an LVT system to offset perceived costs of unrealized capital gains. If capital markets are efficient, then there really ought to be no perceived costs; homeowners enjoying appreciation can simply access this new capital to pay for the increased taxes, all of which ought to be perfectly capitalized. However, if capital markets are inefficient or if homeowners do not access them, then there is a real or perceived cost to unrealized capital gains. This cost could potentially be addressed with a side payment or a tax reduction.

29 This research began with a concern that LVT was unacceptable to voters. In the experiment, LVT was induced to burden those with high land values and low improvement values but not those with low improvement preferences (low-a values). When low-a value owners over-invest, these anticipated LVT policy losers instead become winners. The researchers anticipated that some losers could be offered a side payment to switch their votes, as one makes a potential Pareto Improvement into a Pareto Improvement.10 Despite these explanations of updated rationality, the number of votes against observed in the experiment still deviated to some extent from the “updated rational” and the deviation comes from both the policy losers and winners. This remaining deviation could possibly be explained by concerns of fairness.11 It is also possible that the low rate of objection to LVT or SRT is that people are comfortable with the status quo, which was LVT or SRT in this case. Hence, there may be the equivalent of order effect in the experiment. Future research could examine a treatment in which

LVT or SRT was “opt-in” rather than “opt-out.”

30 Table 1: Model Parameters from Harrisburg, Pennsylvania Neighborhood Income Low Mid High Property Valuea $49,200 $78,000 $169,100 Land Valueb $10,925 $17,320 $37,551 Improvement Valueb $38,275 $60,679 $131,549 Incomec $31,468 $49,930 $84,878 Source: Table values constructed by authors using the following data sources. aData come from three Census Tracts (213, 217, 219) in Dauphin County, Pennsylvania, measuring the high/low and middle extremes in the Harrisburg area. The U.S. Census (2016a) measure is median value of owner occupied housing units from the 2005-2009 American Community Survey 5-year estimates. bLand and improvement values were imputed from the Census housing data. Specifically, the City of Harrisburg reports (The Center for the Study of Economics 2016) in their millage rates that the total taxable assessed value, PV, for land is $357,997,500 and the taxable value of improvements and buildings is $1,254,150,100. Thus, the ratio of improvements to land values is β=IV/LV=3.503. These data are used to apportion housing value into land and improvement value. cData on median owner income (2010-14) come from three Census Tracts (213, 217, 219) in Dauphin County, Pennsylvania, as with housing value (U.S. Census 2016b). During the initial design, the researchers found that the extreme differences in income led to too high earnings for participants in the role of high- income types and an inadequately salient treatment effect. The researchers therefore reduced the income values in the table by 50 percent in the final parameterized version of the experiment.

Table 2: Hypotheses for the Experiment Hypotheses for LVT Support in Support in Experiment Simulation LVT increases community investment Yes Yes (measured as property values) relative to UPT LVT increases social welfare relative to Yes No. LVT generated higher social UPT welfare in one third of the experiment sessions LVT increases investment in near term Yes Yes. But “low preference” owners but this impact dissipates over time for over-invested the “low preference” owners Owners vote against (for) LVT when Yes Some support, but some failures they observe higher (lower) tax compared to UPT LVT can generate sufficient tax revenue Yes Yes (tested as positive tax growth) Owner tends to overinvest when they No Yes can observe their relative status in the neighborhood Note: This table omits the hypotheses for split-rate taxation, which are similar to land value taxation (LVT).

31 Table 3: Experiment Results on Aggregate Welfare by Treatment and Session

Change Change PG- UPT LVT SRT UPT to LVT UPT to SRT Graph Welfare Welfare Welfare (%) (%) Simulation Result (Myopic Efficiency n.a. $748,379 $755,991 1.02% $752,637 0.57% Frontier) $715,513 $674,928 Session 1 Yes -5.67% (95.6%) (89.3%)

$713,971 $717,869 Session 2 Yes 0.55% (95.4%) (95.0%)

$705,417 $726,227 Session 3 Yes 2.95% (94.3%) (96.1%)

$717,052 $680,954 Session 4 Yes -5.03% (95.8%) (90.1%)

$717,747 $713,861 Session 5 No -0.54% (95.9%) (94.4%)

$739,305 $717,678 Session 6 No -2.92% (98.8%) (94.9%)

$745,177 $739,355 Session 7 Yes -0.78% (99.6%) (98.2%)

$743,536 $744,278 Session 8 Yes 0.10% (99.4%) (98.9%) Note: Treatments for uniform property tax (UPT), land value taxation (LVT) and split-rate taxation (SRT). PG-Graph indicates whether treatment used a positional-good graph.

32 Table 4: Treatment Effect on Welfare by Type (LVT versus UPT)

Predicted Observed UPT LVT Earnings UPT LVT Earnings Type Simulation Simulation Change in Experiment Experiment Change in LVT LVT $74,691 $65,952 1 $78,059 $73,995 -5.2% -11.7% (95.7%) (89.1%) $72,442 $65,166 2 $72,373 $69,464 -4.0% -10.0% (100.0%) (93.8%) $66,082 $54,669 3 $68,002 $66,091 -2.8% -17.3% (97.2%) (82.7%) $57,474 $64,261 4 $60,574 $65,264 7.7% 11.8% (94.9%) (98.5%) $60,289 $69,882 5 $65,417 $72,501 10.8% 15.9% (92.2%) (96.4%) $52,762 $43,934 6 $54,944 $52,229 -4.9% -16.7% (96.0%) (84.1%) $49,217 $46,095 7 $50,937 $49,014 -3.8% -6.3% (96.6%) (94.0%) $46,477 $46,181 8 $47,851 $46,613 -2.6% -0.6% (97.1%) (99.1%) $41,324 $44,990 9 $42,338 $45,706 8.0% 8.9% (97.6%) (98.4%) $41,798 $45,525 10 $45,490 $50,549 11.1% 8.9% (91.9%) (90.1%) $36,000 $33,453 11 $37,251 $35,562 -4.5% -7.1% (96.6%) (94.1%) $30,711 $30,882 12 $34,514 $33,338 -3.4% 0.6% (89.0%) (92.6%) $32,286 $30,320 13 $32,399 $31,669 -2.3% -6.1% (99.7%) (95.7%) $27,453 $30,600 14 $28,212 $30,535 8.2% 11.5% (97.3%) (100.2%) $29,163 $33,345 15 $30,018 $33,461 11.5% 14.4% (97.2%) (99.7%) Note: Total earnings in experimental dollars for each type after five periods. Data are averaged across sessions. Treatments reported are uniform property tax (UPT) and land value tax (LVT). Because of transfers, experiment earnings for some types can exceed the simulated predictions. Types 1-5 are high income, types 6-10 are middle income, and 11-15 are low income. Types (5, 10, and 15) have the highest preference for improvements.

33 Table 5: OLS Regression of Experimental Data Explaining Improvement Decision with Heteroskedastic-Robust Variance Treatment and Pooled Vote=1 Vote=0 Vote=0; Vote=0; Control Period 2-5 Period 2-4 Variables (N=1,800) (N=600) (N=1,200) (N=960) (N=720) 1,341*** 912** 1,440*** 1,442*** 1,363*** LVT (332) (383) (333) (347) (420) 521** 232 544*** 543*** 482* SRT (205) (552) (204) (209) (253) 241 194 317* 319 217 PG-Graph (177) (472) (193) (224) (276) -10 Vote (254) LVT * PG- 420 525 344 342 610 Graph (414) (588) (416) (434) (530) 63 LVT * Vote (457) 134 SRT * Vote (408) LVT * Vote * -48

PG-Graph (510) -388 501 -676*** Period2 (243) (555) (266) -1,045*** -262 -1,310*** -634** -634** Period3 (224) (471) (257) (277) (276) -1,498*** -1,281*** -1,481*** -805*** -805*** Period4 (210) (443) (244) (269) (269) -2,081*** -1,950*** -2,047*** -1,371*** Period5 (203) (415) (230) (251) -6,876*** -6,322*** -7,153*** -6,669*** -6,795*** Type a=0.2 (236) (437) (275) (308) (379) -5,830*** -4,699*** -6,396*** -5,904*** -6,039*** Type a=0.25 (237) (491) (251) (281) (348) -5,640*** -5,128*** -5,896*** -5,637*** -5,731*** Type a=0.3 (230) (427) (264) (290) (362) -632*** 166 -1,032*** -992*** -958*** Type a=0.7 (222) (470) (229) (250) (308) -3,173*** -2,279*** -3,621*** -3,562*** -3,675*** Neighborhood2 (181) (327) (214) (241) (298) -5,273*** -4,530*** -5,646*** -5,492*** -5,708*** Neighborhood3 (185) (363) (207) (233) (284) 13,960*** 12,714*** 14,559*** 13,556*** 13,776*** Constant (317) (578) (343) (389) (452) R2 0.62 0.54 0.68 0.65 0.64 Notes: Heteroskedastic-robust standard errors in parentheses below coefficient estimate. Baseline indicator variables indicate: UPT=1, a=0.8, Neighborhood1=1. Fractional design does not allow fixed effects estimation. Stars indicate t-tests at *** (1%), ** (5%),* (10%) levels. A prior pooled model without type and neighborhood controls had poor explanatory power (r2=0.04).

34 Table 6: Vote Count against a Land Tax and in Favor of Uniform Property Tax Period 2 Period 3 Period 4 Period 5 PG-

Graph URV EV URV EV URV EV URV EV Simulation Result (Myopic Efficiency n.a. 9 9 n.a. n.a. n.a. n.a. n.a. n.a. Frontier) Session 1 Yes 5 5 5 7 5 7 5 6

Session 2 Yes 8 3 7 3 7 5 7 4

Session 3 Yes 8 5 9 6 9 6 9 5

Session 4 Yes 6 5 6 4 6 4 7 5

Session 5 No 9 8 n.a. n.a. n.a. n.a. n.a. n.a.

Session 6 No 9 4 9 6 9 6 9 7

Session 7 Yes 9 6 9 7 9 8 n.a. n.a.

Session 8 Yes 7 4 7 5 7 7 7 7 Notes: Column heading “EV” indicates experiment data on votes, while “URV” indicates the updated rational number of votes given the evolved neighborhoods and tax incidence. When votes against reach eight, the session abandons LVT or SRT and reverts to UPT, and no additional voting questions are asked (indicated by n.a.). Induced values and optimal behavior in simulation suggested that nine would vote against in the first vote. Updated rational votes are recalculation of optimal response in light of prior deviations from the optimal path. The updated rational vote shows that because of the deviation in the early periods, the number of voting against a land text may not reach eight even if all participants make decisions rationally from that point onwards.

35 Table 7: OLS Regression Explaining Deviations from Updated Optimal with Heteroskedastic-Robust Variance

LVT sessions; Treatment and Control Pooled period 2-5 Variables (N=1,800) (N=810) 450.1* 413.1 LVT (259.6) (341.3) 161.4 SRT (166.7) 196.4 36.7 PG-Graph (152.4) (287.7) -58.7 383.9 Vote (242.8) (295.5) 530.4* 914.5** LVT * PG-Graph (277.1) (402.4) 190.3 -106.1 LVT * Vote (335.5) (411.3) 180.6 SRT * Vote (398.8) 18.4 Period2 (206.5) -205.1 -356.6 Period3 (191.4) (269.6) -253.4 -331.4 Period4 (177.3) (257.5) -459.2** Period5 (178.5) 1,489.3*** 1,863.6*** Type a=0.2 (205.0) (334.4) 1,577.8*** 2,073.1*** Type a=0.25 (195.2) (318.0) 1,367.2*** 1,777.8*** Type a=0.3 (201.6) (331.3) 557.2*** 481.3* Type a=0.7 (180.5) (259.9) -164.0 -313.3 Neighborhood2 (165.1) (282.0) -26.4 -205.0 Neighborhood3 (154.7) (252.2) -776.1*** -915.8** Constant (234.0) (388.7) R2 0.09 0.13 Notes: Standard errors are in parentheses below coefficient estimate. Baseline indicator variables indicate: UPT=1, a=0.8, Neighborhood1=1. Fractional design does not allow fixed effects estimation. Stars indicate t-tests at *** (1%), ** (5%),* (10%) levels.

36

Table 8: Regression Explaining Earnings by Deviation and Experimental Controls with Heteroskedastic-Robust Variance Treatment and Control Pooled (N=1,800) Pooled (N=1,800) Pooled (N=1,800) Variables -.07* .06 .11 Deviation (.04) (.07) (.08) -140.28 - 112.67 -94.19 PG-Graph (180.68) (179.73) (89.67) -236.10 -86.17 -3.29 LVT (170.50) (178.27) (122.56) 59.69 -33.96 -19.90 SRT (239.63) (239.77) (131.68) 154.39 167.13 178.87* Vote (169.33) (167.26) (104.01) -.27*** -.41*** LVT * Deviation (.08) (.10) .14 .05 SRT * Deviation (.12) (.18) 684.45*** Type a=0.2 (148.92) 316.86** Type a=0.25 (128.31) -277.79** Type a=0.3 (122.68) -551.17*** Type a=0.7 (134.43) -3,768.90*** Neighborhood2 (98.24) -6,255.09*** Neighborhood3 (114.06) 9780.93*** 9740.17*** 13,023.50*** Constant (174.01) (172.93) (93.76)

R2 0.0063 0.0224 0.6741

Notes: Standard errors in parentheses below coefficient estimate. Baseline indicator variables indicate: UPT=1, a=0.8, Neighborhood1=1. Fractional design does not allow fixed effects estimation. ** Significant at 5% level, ***

37 Significant at 1% level.

UPT, yṯ * LVT, yṯ *

UPT, LVT, yṯ * LVT utility (a=0.2) yṯ UPT utility (a=0.2) * UPT utility (a=0.2) LVT utility (a=0.2)

LVT budget Other Consumption Other constraint Consumption Other LVT budget UPT budget UPT budget constraint constraint constraint

x UPT,* x LVT,* x UPT,* x LVT,* ṯ ṯ Improvement ṯ ṯ Improvement

Panel a Panel b

UPT, yṯ * UPT utility (a=0.2)

y UPT,* ṯ LVT utility (a=0.2)

Other Consumption Other UPT budget constraint LVT budget constraint

x LVT,* UPT, ṯ xṯ * Improvement

Panel c

Figure 1: Myopic Utility Maximization Under Two Tax Treatments Source: Original work by authors. Note: UPT utility curve and budget constraint are generated from experimental data for a low-income neighborhood and a=0.2. However, the LVT curves are stylized to visualize a larger treatment effect.

38

200000 All households

Improvement Value 150000 Land Value

100000

50000

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 High income Median income Low income HouseholdID

300000 All households Improvement Value

250000 Land Value

200000

150000

100000

50000

0 HouseholdID 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

Figure 2: Positional-Good Graph at Initial Value (panel 1) and After Investment (panel 2)

39

Figure 3: Simulated Optimal Behavior by Period Source: Original work by authors. First row is the high-income neighborhood followed by mid- and low-income.

Figure 4: Simulated Optimal Welfare Results by Period Source: Original work by authors. First row is the high-income neighborhood followed by mid- and low-income.

40 300000

250000 UPT LVT

200000

150000

100000 Property Value ($) Value Property

50000

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 HouseholdID Figure 5: Treatment Effect on Property Value by Type (LVT versus UPT)

Figure 6: Treatment Effect on Behavior by Type (LVT versus UPT) Note: First row is the high-income neighborhood followed by mid- and low-income.

41 20000 15000 10000 5000 0 0 200 400 600 800 1000 1200 1400 1600 1800 2000 -5000 -10000 -15000 -20000

Figure 7: Ordered Deviation from Updated Optimal Note: These data reflect all 1,800 choices in the experiment; 120 participants made 15 decisions. These are deviations from the initial UPT treatment, a second land tax treatment, and a third voting treatment.

42 References

Banzhaf, H. Spencer, and Nathan Lavery. 2010. “Can the Land Tax Help Curb Urban Sprawl?

Evidence from Growth Patterns in Pennsylvania.” Journal of Urban Economics 67 (2):

169-179.

Behrens, Kristian, Yoshitsugu Kanemoto, and Yasusada Murata. 2015. “The Henry George

Theorem in a second-best world.” Journal of Urban Economics 85: 34-51

Bourassa, Steven C. 2009. “The Political Economy of Land Value Taxation.” In Land Value

Taxation: Theory, Evidence, and Practice, ed. Richard F. Dye and Richard W. England.

Cambridge, MA: Lincoln Institute of Land Policy.

Bowman, John H., and Michael E. Bell. 2008. “Distributional Consequences of Converting the

Property Tax to a Land Value Tax: Replication and Extension of England and Zhao.”

National Tax Journal 61 (4): 593-607.

Brueckner, Jan K., and Hyun-A. Kim. 2003. “Urban Sprawl and the Property Tax.” International

Tax and Public Finance 10 (1): 5-23.

Center for the Study of Economics, The. 2016. “City of Harrisburg Two-Tier Tax Rate,”

http://www.urbantoolsconsult.org/upload/City of Harrisburg 2 tier tax rate.pdf.

Chapman, Jeffrey I., Robert J. Johnston, and Timothy J. Tyrrell. 2009. “Implications of a Land

Value Tax with Error in Assessed Values.” Land Economics 85 (4): 576-586.

Choi, Ki-Whan, and David L. Sjoquist. 2015. “Economic and Spatial Effects of Land Value

Taxation in an Urban Area: An Urban Computable General Equilibrium Approach.”

Land Economics 91 (3): 536-555.

De Borger, Bruno, Amihai Glazer. 2017. “Support and opposition to a Pigovian tax: Road

pricing with reference-dependent preferences.” Journal of Urban Economics 99: 31-47

43 DiMasi, Joseph A. 1987. “The Effects of Site Value Taxation in an Urban Area: A General

Equilibrium Computational Approach.” National Tax Journal 40 (4): 577-590.

England, Richard W., and Min Qiang Zhao. 2005. “Assessing the Distributive Impact of a

Revenue—Neutral Shift from a Uniform Property Tax to a Two-Rate Property Tax with a

Uniform Credit.” National Tax Journal 58 (2): 247-260.

Fischbacher, Urs. 2007. “Z-Tree: Zurich Toolbox for Ready-Made Economics Experiments.”

Experimental Economics 10 (2): 171-178.

Fischel, William A. 2015. Zoning Rules! The Economics of Land Use Regulation. Cambridge,

MA: Lincoln Institute of Land Policy.

Frank, Robert H. 1985a. Choosing the Right Pond: Human Behavior and the Quest for Status,

New York: Oxford University Press.

____. 1985b. “The Demand for Unobservable and Other Nonpositional Goods.” The American

Economics Review 75 (1): 101-16.

Frank, Robert H., Adam Seth Levine, and Oege Dijk. 2014. “Expenditure Cascades.” Review of

Behavioral Economics 1: 55-73.

Hirsch, Fred. 1976. Social Limits to Growth. Cambridge: Harvard University Press.

Patacchini, Eleonora, and Giuseppe Venanzoni. 2014. “Peer effects in the demand for housing

quality.” Journal of Urban Economics 83: 6-17

Plassmann, Florenz, and T. Nicolaus Tideman. 2000. “A Markov Chain Monte Carlo Analysis of

the Effect of Two-Rate Property Taxes on Construction.” Journal of Urban Economics

47 (2): 216-247.

Plummer, Elizabeth. 2009. “Fairness and Distributional Issues.” In Land Value Taxation:

Theory, Evidence, and Practice, ed. Richard F. Dye and Richard W. England.

44 Cambridge, MA: Lincoln Institute of Land Policy.

Pollock, Richard L., and Donald C. Shoup. 1977. “The Effect of Shifting the Property Tax Base

from Improvement Value to Land Value: An Empirical Estimate.” Land Economics 53

(1): 67-7.

U.S. Census. 2016a. “Median value of owner-occupied housing units.” factfinder.census.gov.

U.S. Census. 2016b. “Median owner income (in 2014 dollars).” www.census.gov/quickfacts.

Veblen, Thorstein. 1899. The Theory of the Leisure Class. New York: Macmillan.

Youngman, Joan. 2016. A Good Tax: Legal and Policy Issues for the Property Tax in the United

States. Cambridge, MA: Lincoln Institute of Land Policy.

1 Adding a tax on consumption ought not change the results or the treatment effect. The paper concerns how property taxes change affect behavior. A consumption tax ought to work as a constant in the budget constraint, and thus its absence or inclusion should not affect the results. If a consumption tax has a redistribution effect, it will have an income effect. 2 The use of present values simplifies the decision that experiment participants will need to make and prevents the introduction of another form of error—that of mistakes made in dynamic optimization from understanding discounting and from intertemporal risk preference differences. Consider that if someone invests $100 in improvements, then they must expect at least that much monetized utility as a stream of benefits. For instance, if the improvements create $5 in extra benefits per year and the discount rate is 0.05, then the perpetuity is $100=$5/0.05. To simplify this model, we assume that the monetized utility is exactly equal to the cost of the improvements. 3 This model has few elements of dynamic optimization because of a desire to ensure experiment participants understood their incentives and to focus on the variables of interest. The endowment is freely provided, but there is no salvage value. There is no discounting. There is a definite end period announced at the start of each treatment. 4 Note that the model assumes that property tax is calculated and fulfilled at the end of each period after all owners make their improvement decisions. In other words, property taxes are levied on the updated property value at the conclusion of a period. 5 A dynamic solution might be able to deliver even higher welfare; however, the experiment results will be shown to be suboptimal even with respect to the derived myopically optimal solution. So the myopic solution is sufficient to demonstrate the predicted treatment effect. 6 The program was constructed in this way—returning taxes after myopic choices were made—for several reasons. First, in the experiment the researchers wanted to allow participants to make any possible decision, rather than directing them only to the “optimal” choice. By in effect assuming tax return was zero, ex ante, it allowed participants to form their own judgments about what they expected the return to be. It would be possible for savvy decision makers to anticipate a certain level of tax return, which would augment consumption, and therefore shade their optimal improvement choice slightly upwards. Second, the tax return arises in part from the in-period decisions of 15 different participants. It is impossible to predict these decisions in advance. Third, the substantive impact of this assumption is small. The magnitude of TR is small. For the lowest income, the tax return in period 1 averaged 6.0-6.1% and rose to 26.0-27.9% in the final period, depending on treatment. For the highest income, the same measures were from 2.2-2.3% in period 1 and 9.6-10.3% in the final period. Finally, the absolute magnitude of this TR value is not important for isolating treatment effects. Rather, it is the relative magnitude in any period across all treatments. The values simply do not change very much. 7 In sum, the deviations from an intertemporal optimum are four fold. In-period optimization means that there is no consideration of dynamics. There is no endogenous solution in which every type is adapting to a best response;

45 every type just plays a naïve myopic strategy. Tax return increases consumption payoff by 1-a, but comes after decisions are made. So, optimally, someone might form expectations about tax return and shade their investment upwards. Finally, there is no consideration of last-period effects. As described in the experimental design, the game will end after five periods. Because there is no period 6, there is no future negative externality of neighbors’ improvements. 8 The administrator computer systematically selected the values in this table in each period by first identifying the approximate myopic optimal choice and then distributing other values across the support of income. This optimal choice evolved over time because incomes fell with larger tax burdens. The lab administrator did not refer to any choice as “optimal,” obviously. During training, participants were informed that these possible values were only decision aids, and any value from 0 to their income could be assigned to improvements. 9 Heteroscedasticity is detected in the OLS regression. Heteroskedastic-robust regressions are estimated and reported in tables 5, 7, and 8. Regressions with standard error clustered at individual level were conducted but are not reported because they show the same parameter estimations. Very few differences exist between these two modeling approaches, and neither approach is obviously superior. The differences are that a type dummy in table 5 and two type dummies in table 8 fell below the statistical significance threshold; however, the major variables of interest have consistent statistical significance as the heteroskedastic-robust estimation. 10 Post-experiment survey questions were designed to collect preliminary data about the acceptability of mechanisms to increase land tax acceptability. These questions anticipated that experiment results would match simulated predictions, where nine participants would vote against a land tax because it would lower their earnings relative to UPT. As such, a side-payment mechanism was proposed such that the policy winners pay the policy losers to vote in favor of a land tax. The survey asked several questions about the acceptability of this side payment mechanism. All questions were phrased in terms of “tax plan 1” and “tax plan 2.” The first question was, “If you were better off under (LVT/SRT), would you have been willing to give some of your earnings to another participant if they would vote to maintain (LVT/SRT)?” Some respondents (37 percent) reported that they would be willing to make the side payment, while others (39 percent) were unwilling. For the remaining respondents, the question was deemed non- applicable because they were not better off under a land tax. The survey asked a parallel question about the supply side of this side-payment-for-votes mechanism: “If you were worse off under LVT/SRT, would you have been willing to accept some of your earnings to another participant if they would vote to maintain LVT/SRT?” Respondents (48 percent) reported that they would be willing to receive payments to change their vote, while 26 percent were unwilling. These question provides some evidence that a side-payment treatment might be able to generate enough vote-switching to support a land tax. 11 A final survey question was a simple one: “What tax plan is the fairest (i.e., treats everyone equally)?” The researchers had no a-priori beliefs about this question. UPT was deemed the fairest by 57 percent, while 24 percent viewed LVT/SRT to be the fairest. “Equally fair” was chosen by 8 percent, while 11 percent responded, “don’t know.”

46