Legislative Term Limits and Government Spending: Theory and Evidence from the United States

2.1 Settings

Suppose that the legislature consists of odd n≥3 legislators. Each legislator is elected from district i with i∈{1,...,n}. The legislators differ in their levels of seniority. We consider three levels of seniority: senior, intermediate, and junior. The length of service is assumed to be senior > intermediate > junior.

The elected legislators negotiate over the allocation of distributive projects to their districts. The distributive projects in our model are continuous units with several possible projects for each district. The amount of distributions for each project is defined as di, which equals the benefit district i receives from this project. Only district i is eligible to receive the benefit from the project. If h distributive projects are allocated to district i, then the total amount of distributions this district receives is hdi. If only the proportion 1−q of h distributive projects is implemented, then the total amount is equal to (1−q)hdi. Without loss of generality, we suppose h=1. ^[4] That is, we focus on the proportion of projects, 1−q, and ignore the number of projects, (1−q)h. The costs of the projects, c(di)=di2, are spread evenly across all districts, and each district pays di2/n. The legislators prefer a higher payoff for their district.

The legislators determine the amount of the distributions to their districts in the following way. ^[5] One of the n legislators is chosen as an agenda setter, who proposes the amount of distributions to each district. If the majority of legislators (i. e. (n−1)/2 legislators and the agenda setter) approve the proposal, then the projects will be implemented. Otherwise, no project is allocated. Hence, di=0 and the payoff is also zero for all i. ^[6]

The selection of an agenda setter is determined by seniority. As discussed by McKelvey and Riezman (1992), legislators with a higher level of seniority are more likely to be an agenda setter because their lengthy career allows them to accumulate legislative skills and receive better committee assignments. ^[7] Our model assumes that the legislator with the highest level of seniority in the legislature becomes the agenda setter. If there are multiple legislators with the highest level of seniority, then each of those legislators has the same probability of being an agenda setter.

To maximize the payoff, the agenda setter proposes di, which allocates the projects to only (n−1)/2 legislators. We call the group of legislators who receive the positive amount of distributions the majority coalition. We assume that di is a product of the negotiations among the members in the majority coalition, including the agenda setter, who consider the benefits for their districts and the amount of the costs incurred by all members in the majority coalition. ^[8] Formally,

This means that the members of the majority coalition circumvent a common pool problem, in that all members maximize their own district’s payoff without considering the costs of the projects imposed on other districts. ^[9]

Furthermore, the agenda setter maximizes the payoffs by manipulating 1−q, the proportion of projects allocated to the districts of the members in the majority coalition. We assume that the agenda setter has the power to control the behavior of other members if his or her level of seniority is higher than that of others. If the agenda setter is senior, then this agenda setter can control a proportion p2 of junior members in the majority coalition, while he or she can control a proportion p1 of intermediate members. Similarly, if the agenda setter is intermediate, he or she can control a proportion p1 of junior members. The agenda setter can control a proportion p0 of members with the same level of seniority. We assume that p2>p1>p0. To simplify, suppose p2=1, p1=p∈(0,1), and p0=0.

This assumption is justifiable because party or committee leaders, who are the typical agenda setters in the federal and state legislatures, possess the formal and informal power to control the behavior of other members by promising future benefits in exchange for loyalty or by disciplining (Clucas 2001; Cox and McCubbins 1993, 2005; Squire and Moncrief 2010). A legislature with a seniority system has a hierarchical structure, which means that the bargaining process is centralized such that a senior leader can control other members in a variety of ways. In contrast, if all legislators have similar levels of seniority and the legislature has no hierarchical structure, then the power that party leaders and committee chairs can exercise against other members is weak. As a result, the bargaining process is likely to be decentralized.

The senior agenda setter, who is typically a party leader, does not allow members with lower levels of seniority in the majority coalition to spend as much as they want for their reelections because the agenda setter is expected to keep the value of party brand names. According to the literature on political parties and legislatures in political science, party members including junior legislators try to keep party labels as informative as possible by allowing party leaders to coordinate legislative activities (Cox and McCubbins 1993; Primo and Snyder 2010). Party labels are a low-cost method to inform constituents about their issue positions. Thus, those labels help (especially junior) legislators secure support from voters who share similar political views. If party leaders let junior legislators to spend as much as possible, it would damage party’s brand names and result in a bad reputation. In short, agenda setters and party leaders allocate some distributions to junior legislators, but the amount is limited so that junior legislators are benefited from both distributive projects and party brand names for future elections.

If the agenda setter can exercise influences on some members of the majority coalition with lower levels of seniority, we call those members who are subject to the control of the agenda setter followers. We suppose that the agenda setter allocates the proportions of projects (i. e. 1−q) to the districts of the followers, and that he or she has no choice but to allocate all of the projects to the other members; thus, the projects with q=0 go to the districts of the other members in the coalition. The agenda setter sets q to satisfy two conditions. First, he or she seeks to reduce the total amount of costs imposed on his or her district. Second, she ensures that the followers have an incentive to approve the proposal by allocating at least some amount of distributions.

2.2 Distributions

Denote f∈[0,(n−1)/2] as the number of followers. As discussed in the previous section, the agenda setter can decide the proportion of projects, 1−q, to be allocated. Thus, the followers’ payoff is

The second term is the costs of implementing projects in the districts of followers, while the third term is the costs of implementing projects in the districts of the other members in the majority coalition (including the agenda setter). Note that these costs are paid by all n districts. The followers reject the proposal if their payoff is lower than zero and approve otherwise because the payoff will be zero if the proposal is rejected.

The agenda setter sets q so that she maximizes her payoff while motivating the followers to approve the proposal. This occurs if the payoff for the followers equals exactly zero. After some calculations, [2] is zero when q=q∗(f) such that

For all f∈[0,(n−1)/2], q∗(f)∈[1/2,(n+1)/(n+3)]. Substitute q in the second term of eq. [2] with q∗(f), then the total costs paid by each district are

In eq. [3], q∗(f) increases with f because, as f increases, C(f) decreases. Put differently, as f decreases, the agenda setter has to allocate a higher proportion of projects (i. e. smaller q) to followers in order to satisfy that the payoff will become zero. This means that as f increases, the total amount of the costs increases.

The payoff of the agenda setter and the other members in the majority coalition is n/(n+1)−C(f), which is positive because it is higher than the follower’s payoff. Therefore, all members of the majority coalition approve the proposal.

The total amount of distributions allocated to the majority coalition is

A higher f (and higher q∗(f)) reduces the total amount of distributions. Thus, if legislators with lower levels of seniority, as compared to the agenda setter, occupy the larger share of the majority coalition, then D(f) decreases. If the agenda setter is senior and all other members are junior in the majority coalition, then D(f) is minimized, where D((n−1)/2)=2n/(n+3). If the agenda setter has the same level of seniority as all of the other members in the coalition, it is maximized, where D(0)=n/2. Therefore, the agenda setter seeks to include legislators with lower levels of seniority in the majority coalition to reduce the total costs paid by her district.

2.3 Elections and Term Limits

Drawing on the above discussion, we examine the electoral process in period (i). Suppose that each district has a representative voter (e. g. the median voter in the Downsian model) who prefers a higher payoff for the district. Consider two types of term limits: moderate and strict term limits. The latter forces legislators to retire earlier than the former does. The type of term limits imposed on the legislature constrains the choice of the representative voter in the election in the following way:

1. In the absence of term limits, the representative voter has a choice of candidates who are either senior, intermediate, or junior.

2. With the adoption of moderate term limits, the representative voter has a choice of candidates who are either intermediate or junior.

3. With the adoption of strict term limits, the representative voter can only elect a junior candidate.

We derive the following equilibria regarding the types of legislators elected in the absence and presence of strict and moderate term limits:

This proposition can be described by the following intuition. Consider a legislature in the absence of term limits. First, suppose that the legislature includes some junior legislators. They always have a nonpositive expected payoff because they receive zero payoff even if they become a member of the majority coalition. On the other hand, senior and intermediate legislators have a positive expected payoff. ^[10] Accordingly, the representative voter who chooses a junior candidate will deviate to choose a senior or intermediate legislator. Thus, in the absence of term limits, the legislature includes no junior legislators in the equilibrium. If no senior legislator is present in the legislature, then the representative voter will deviate to choose a senior candidate who will be an agenda setter with certainty.

Second, suppose that the legislature consists of all intermediate legislators. Then, the representative voter will deviate to choose a senior candidate because the voter will enjoy the comparative advantages in legislative bargaining by electing a legislator with a higher level of seniority. Hence, senior legislators are elected from at least some districts. As the number of senior legislators increases, the expected payoffs for senior legislators decreases because the probability of being an agenda setter decreases. On the other hand, the expected payoff for intermediate legislators increases because the probability of being included in the majority coalition increases. ^[11] When the number of senior legislators equals n‾, the expected payoff of senior and intermediate legislators become equivalent if p is not very high. ^[12] Then, no voter has an incentive to deviate. Thus, Equilibrium I exists. However, if p is very high (i. e. p≥(n2−1)/(n2+1)), then the expected payoff for intermediate legislators always falls below that of senior legislators because most intermediate legislators in the majority coalition are subject to the influence of the senior agenda setter and thus receive only zero payoff. In this case, Equilibrium I does not exist.

Third, suppose that all legislators are senior. They have no comparative advantage, and (n−1)/2 senior legislators are not included in the majority coalition. Thus, the representative voter may deviate to choose an intermediate candidate who will be included in the majority coalition with certainty. However, if a senior agenda setter can exercise an influence on many intermediate legislators (i. e. p≥(n−1)/n) in the majority coalition, the representative voter will not deviate because the intermediate legislator is likely to produce zero payoff. If this is true, the representative voter will not deviate and Equilibrium II exists. Note that, since (n−1)/n<(n2−1)/(n2+1), at least either Equilibria I or II exists for all p∈(0,1).

The same equilibrium outcomes are extended to legislatures with moderate term limits. In that legislature, intermediate legislators play the same role as senior legislators in the absence of term limits.

Note that our model assumes that, in the absence of term limits, voters always have a choice of all three types of candidates who differ in their level of seniority. However, voters seldom participate in electoral contests between a senior and an intermediate candidate. In reality, they often make a choice between a senior or intermediate incumbent and a junior challenger or between two junior candidates, which is not always consistent with our discussion. However, the most important implication of the equilibria is that the representative voter has no incentive to choose a junior candidate, which then means that a legislature is unlikely to include many junior legislators in the absence of term limits. To be precise, Proposition 1 indicates that all other situations beyond the equilibria are unlikely to be stable. For example, suppose that the legislature without term limits contains two junior candidates. In this case, even though voters cannot deviate to choose either an intermediate or senior legislator immediately, voters may wait until this elected junior legislator will be an intermediate (or a senior) legislator in the long run. Thus, even though the components of the legislatures are not exactly same as the components in the equilibria, they will change and approximate the components in Equilibrium I (if p is not very high) or II (if p is very high). Thus, these equilibria can be interpreted as an approximation of the reality.

Note also that our model considers only legislators of one chamber, but there are two possible other political actors who may have a power to approve the bill; a governor and the other chamber. In some states, a governor has a line-item veto, so even if the majority of legislators approve a bill proposed by an agenda setter, a governor may reject it. Moreover, if the state employs bicameralism, another chamber may also be able to reject a bill. In this case, these political actors (governor and/or the majority of the other chamber) become members of the majority coalition of an agenda setter. That is, an agenda setter needs to provide some distributions to policies favored by a governor and/or the majority members of the other chamber. However, the theoretical implications (Proposition 1) do not change so much since some players are just added as the coalition members, so our model does not include their effects explicitly.

2.4 Predicting the Effect of Term Limits

Building on Proposition 1, we now derive two hypotheses regarding the effects of term limits on government spending. A serious challenge for an empirical test is that we cannot directly observe the type of equilibrium (i. e. either equilibrium I or II, or either equilibrium I′ or II′) in the legislature. However, if Equilibrium II changes to II′ or III, both the variance in the level of seniority and the total amount of government spending will show no change. Similarly, if Equilibrium I changes to Equilibrium I′, both the variance and the amount of spending will show no change. In contrast, if Equilibrium I changes to II′ or III, the variance in the level of seniority decreases, and the amount of government spending increases. On the other hand, if Equilibrium II changes to I′, the variance in the level of seniority increases, and the amount of government spending decreases. We summarize these relationships in Table 1.

As a result, we obtain the following hypothesis for an empirical test:

H1: The adoption of term limits that results in a larger reduction in the variance of seniority will increase the amount of government spending.

Note that for all changes in the equilibria, the average level of seniority decreases in the legislature. This means that the change in the average level of seniority is not an important determinant of the total amount of distributions. Thus, the empirical test for H1 requires us to examine how the change in the variance of seniority affects the amount of government spending. ^[13]

For the second hypothesis, we exploit the types of term limits. When moderate term limits are adopted, the amount of government spending will not change because the equilibrium outcomes are identical between the legislatures without term limits and with moderate term limits. In contrast, when strict term limits are adopted, the amount of government spending can increase from D‾ to n/2. Thus, only the adoption of strict term limits will increase the total amount of government spending. Accordingly, we obtain the following empirical hypothesis:

H2: The adoption of stricter term limits will increase the amount of government spending, while the adoption of moderate term limits will show no change in the amount.

3.1 Data and Methods

We use the following model:

where [Spending]jt denotes the measures of total government expenditures in state i in year t. [Term]jt−1 is a measure of the term limits imposed on a legislature in state j at year t−1. wjt−1 includes all time-varying political variables that may have an impact on government expenditures and the adoption of term limits. We take the lag of these variables to address the gap between the budget year and the election year. ^[14]xjt includes all time-varying socioeconomic variables. ρj denotes a state-fixed effect which captures all time-invariant characteristics of state j. ϕt denotes a year-fixed effect which captures any time-specific shock at the national level. Finally, εjt is a state–year-specific error term.

All time-invariant characteristics of state j are captured by the state-fixed effect, ρj. Time-invariant characteristics include stable institutional designs (e. g. budget cycles, budget powers of executive and legislative branches, and election systems) and potentially unobservable cultural norms and ideologies that could be related to the adoption of term limits and government expenditures. Thus, our estimation results are not affected by state characteristics that do not vary over time.

Any time-specific shock is captured by the year-fixed effect, ϕt. Year-fixed effects capture the effects of election years, national economic conditions, and any other major events that occurred in a particular year that might be associated with government expenditures.

The outcome variable, [Spending]jt, is measured in two different ways. First, we use total state government expenditures per capita in dollars. Second, we use the percent of total government expenditures in state personal income. We assume that the size of spending for distributive projects is correlated strongly with the amount of total government spending because the allocation of distributive benefits is determined independently from other necessary expenditures. The government expenditures per capita are reported in constant 1982 dollars. The data on the government expenditures and personal income come from “State Government Finances” compiled by the US Census Bureau. ^[15]

In eq. [8], [Term]jt−1 denotes our two alternative measures of term limits, which exploit the differences in the types of term limits adopted by the states. For the first hypothesis, we measure the size of reduction in the variance of seniority within the legislature before and after the adoption of term limits. We expect that term limits that cause a large reduction in the variance of seniority increase total expenditures because the large reduction means that the adoption of term limits replaces an agenda setter with the high level of seniority with one who has a low level. In contrast, term limits that caused only a small reduction in the variance will have a smaller impact on total expenditures because the small reduction means that the distribution of seniority within the legislature do not change a lot.

The size of the reduction in the variance of seniority within the legislature is measured by comparing the variance before and after the adoption of term limits. More specifically, for each state with term limits, we compute the average variance in the level of seniority before and after the adoption of term limits. To compute the average variance before the adoption of term limits, we include the years after 1990, so that we could compare the change in the variance of seniority just before the adoption of term limits. For Maine where term limits were adopted in 1996, for example, we compute the average variance in seniority between 1990 and 1995 and the average variance between 1996 and 2010, and then took a difference in the averages. The changes in the average variance of seniority before and after the adoption of term limits are reported in Table 2. ^[16] All states but Oregon showed a positive value, which means that the variance of seniority in the house and the senate is larger before the adoption of term limits than after that.

We use the size of the reduction of seniority, as shown in Table 2, to test whether the adoption of term limits that results in a larger reduction in the variance of seniority tends to increase the amount of government spending. We create a continuous scale of the reduction in the variance of seniority for the states that adopted term limits. The positive values of the scale indicate the larger reduction in the variance after term limits became effective. The remaining states (including the states without term limits) and years were coded zero. In the case of Maine, for example, [Term]jt−1 is coded as 0 before 1996 and 0.75 after 1997. As the model includes the state-fixed effect term, we test whether there was a statistically meaningful difference in the amount of government spending before and after term limits came into effect and whether the difference becomes larger as the size of the reduction of seniority increases.

For the second hypothesis which focuses on the effects of strict and moderate term limits, we categorize states that adopted term limits by using the difference in the maximum years of service. Of the 14 states that adopted term limits for the house before 2010, four states (AR, CA, MI, and OR) set the limit at 6 years, while eight states (AZ, CO, FL, ME, MO, MT, OH, and SD) set the limit at 8 years. ^[17] The remaining two states (LA and OK) set the limit at 12 years. Similarly, of the 13 states that adopted term limits for the senate before 2010, eleven states (AR, AZ, CA, CO, FL, ME, MI, MO, MT, OH, OR, SD) set the limit at 8 years, while two states (LA and OK) set the limit at 12 years. No state adopted a 6-year limit for the senate.

We assume that term limits allowing legislators to stay for less time in the legislature (i. e. equal to 6 years) were stricter than those allowing them to stay longer (i. e. 8–12 years). For estimation, we create three indicator variables for the 6-year, 8-year, and 12-year term limits for the house. while we create two indicator variables for the 8-year or 12-year term limits for the senate. They equal 1 in the subsequent years after term limits became effective in state j and zero otherwise. Other states and years are coded zero. Our analysis exploited the states without adopting term limits as a control group. The data on term limits are obtained from the Website for the National Conference of State Legislatures. ^[18]

Importantly, our measure of term limits improves upon the existing binary variable that equals 1 if term limits are adopted and 0 otherwise. Similarly to Sarbaugh-Thompson’s (2010) study, we argue that the binary measure is not sufficient to capture the characteristics of different term limit rules across the states and thus able to examine their impacts on political and economic outcomes.

The vectors wjt−1 in eq. [8] contain control variables for other time-varying political and socioeconomic characteristics of states. We include the measure of state legislative professionalism (Squire 2007) that is likely to be correlated with the level of seniority. The measure is available only for 1979, 1986, 1996, and 2003, and we thus linearly interpolate the values for other years. State political characteristics are measured by the percentage of Democratic legislators in the chamber, an indicator variable for a Democratic governor, and an indicator variable for divided government that takes a value of 1 unless the same party controls the governor’s office and both chambers. The data come from Klarner (2011). In addition, we take into account the presence of the executive term limit. The indicator variable equals 1 if the term limit on the governor was effective and 0 otherwise. The data are obtained from List and Sturm (2006).

The vectors xjt include socioeconomic characteristics. They are captured by the unemployment rate, personal income per capita, the population size, and the percentage of the population under 15 years old and over 65 years old. All monetary variables are reported in constant 1982 dollars. We took the natural log of personal income per capita and population size. All of the socioeconomic data came from the Statistical Abstract of the United States. Summary statistics are presented in Table 3.

3.2 Results

We estimate eq. [8] to find numerical estimates of the effects of term limits. Table 4 presents the estimation results when using the amount of government expenditures per capita as an outcome variable. To test H1, we use the size of the changes in the variance of seniority within the legislatures as a predictor of the per capita spending. Column (1) of Table 4 presents the estimation results for houses. As predicted, larger reductions in the variance of seniority has a positive and statistically significant effect on the amount of per-capita spending. The coefficient indicates that the amount of spending increases by $56 as term limits reduce the variance of seniority by 1. Note that Table 2 indicates that the size of the reduction in the variance varies from 0.67 to 3.10. In contrast, column (2) shows that the same variable for the senate has little effect on the amount of spending.

Next, we test H2 by exploiting the different types of term limits. Column (3) of Table 4 shows that the coefficient associated with the 6-year limit for the house was positive and statistically significant. The coefficient indicates that the amount of government expenditures per capita increased by $176 after states adopted 6-year (i. e. stricter) term limits. The coefficients associated with 8- and 12-year term limits indicate that their adoption had no statistically significant effect on government expenditures. Column (4) shows that coefficients associated with the 8-year and 12-year term limits for senates had a positive but statistically insignificant association with the outcome variable.

Some control variables in Table 4 show consistent patterns to explain the amount of spending. Among the political variables, divided government was the only predictor with a significant coefficient. When the government was controlled by two parties, the amount of expenditures tended to increase by $30–$40. The demographic variables showed that the amount of expenditures increased as the percent of the population who were unemployed or under 15 years old, and per capita personal income increased, while it decreased as the log size of population size increased.

The results in Table 4 are replicated by using the share of government expenditures in state personal income as an outcome variable. Table 5 reports the estimation results. For the houses, larger drops in the variance of seniority as a result of term limits in the house in column (1) increase the share of state government expenditures. Similarly, the adoption of 6-year term limits in column (3) increases the share by 1.1 percentage points. On the other hand, the 8-year and 12-year term limits in the house have no statistically significant relationship with the outcome variable. Similarly, the adoption of term limits in the senate has no significant effect on the share of government spending. With respect to the control variables, the share of government expenditures increased in the presence of divided government and as the population size decreased.

Tables 4 and 5 show the consistent results that the effects of term limits for the house are larger than the senate. We speculate that this is because the adoption of term limits resulted in a smaller change in the variance of seniority in the senate than in the house across the states. The average reduction in the variance in the senate is by 0.83, while the average reduction in the house is 1.45. Thus, the variance of seniority changes to the larger extent in the house as a result of term limits, resulting in the increase in government spending.

Taken together, according to our results, term limits that cause a large change in the distribution of seniority have a larger positive effect on the amount of government expenditures than those that cause a small change. These findings offer evidence that more restrictive term limits that greatly change the distribution of seniority within the legislature increase the amount of spending, while less restrictive term limits that slightly change the distribution of seniority have little effect on the amount of spending. One may argue that our estimation results could be potentially biased because of a potential reverse causation between the adoption of legislative term limits and government spending. In other words, stricter term limits might be more likely to be adopted by the states with the higher levels of spending. However, Erler (2007) and Mooney (2009) suggest that major political and demographic characteristics of states have no strong relationship with the adoption of legislative term limits. Thus, this implies that the reverse causality is unlikely to have a significant influence on our estimation results.