The Effect of Unfunded Pension Liabilities on State Borrowing Costs Part 2

Chapter 4. Empirical Strategy We turn now to the empirical model of this study. This chapter delineates first the basic empirical strategy underlying my study and then explains the specification issues of the model, specifically endogeneity and nonlinearity. 4.1 The Basic Empirical Model The aim of this study is to estimate the credit supply curve with a particular focus on the effect that unfunded pension liabilities have on credit supply. The following is the basic regression for this study: ∗ The dependent variable, ∗ ∗ ∗ , is the spread of the state bond yield over the Treasury yield, the assumed risk-less rate, for a given state and year. This spread will capture the default spread since, as explained in Chapter 3, excluding the issue of taxes, any spread over the risk-free rate is due to default risk. The variable is the measurement of individual states’ unfunded pension liabilities divided by GSP. As state liabilities, these unfunded pension liabilities increase a state’s probability of default and therefore, should increase the state’s default spread. is state debt divided by GSP and clearly affects the default spread, although the complications of this relationship are discussed below in Section 4.2. is the unemployment rate and controls for business cycle variation in the supply curve. The last variable, , is a dummy variable for fiscal institutions that affect states’ default risk as well, such as limitations to states’ debt or deficits. 14 4.2 Specification Issues A major specification issue in the above regression is the endogeneity between the debt variable and the yield. This endogeneity occurs because it is difficult to separate the supply of credit from the demand for credit. The demand for credit, that is, state’s demand for funds, has the opposite relationship between debt levels and yields from the supply of credit. States will demand more credit if there is a lower yield while investors will supply more credit with a higher yield. Therefore, the above regression will be endogenous so that the debt variable in the equation is correlated with the error term. Most regressions in this study use state fixed-effects to partially account for this simultaneity. State fixed-effects control for state-specific characteristics that are hard to capture quantitatively. These state-specific characteristics can lead to shifts in the supply curve instead of movements along the supply curve. To estimate the supply curve, the regressions in this study must try to capture the variation due to the movements along the supply curve and shifts in the demand curve. Any variation due to shifts in the supply curve would estimate the demand for credit instead of the supply of credit. Not controlling for state characteristics leads to a downward bias in the debt variable estimates. If a state characteristic causes investors to be more willing to supply credit at a lower yield, the lower yield will incentivize the state to increase their debt levels. For example, the state of Delaware, due to its small size, does not have very developed municipal governments, which means that the state has a much larger taxing power, given its population and income, than the typical state. This state characteristic allows Delaware to have a higher debt to GSP ratio than the average state for any given yield, and therefore, this characteristic leads to a rightward shift in Delaware’s supply 15 curve.4 It is important to note as well that fixed effects also capture different taxtreatments between states. As long as states maintain the same tax-treatment, fixed effects effectively account for the tax-treatment complication described in Section 3.1. Although fixed-effects solve to some degree the issue of an endogenous debt variable, fixed effects do not completely resolve the problem of bias. Fixed-effects account for shifts in the supply curve between states, but not the shifts that happen over time. In other words, fixed effects do not ameliorate the problem of time-series simultaneity. For example, political factors that vary over time, such as the state of California’s fiscal gridlock during the past few years, may cause supply curve shifts that are not captured by state dummy variables. To address time-series endogeneity, this study uses instrumental variables. These instruments must affect the demand for state borrowing but not the supply, which is extremely difficult since most variables affect both the supply and the demand for credit. The instrumental variables used in this study are the residuals of the linear trend of each state’s general fund revenues, general fund expenditures, total expenditures, and general fund deficits. The logic behind these instruments is that these temporary deviations from these linear trends will lead to shifts in the demand curve. Unfortunately, as explained in Section 6.2, there is not enough correlation between these variables and the debt variable to make them valid instruments. The last specification issue is nonlinearity. As explained in Section 3.2, the credit supply curve is nonlinear due to the rationing of credit. To account for nonlinearity my study uses squared terms for both the debt and unfunded pension liability variables. 4 Example taken from Goldstein and Woglom (1991). 16 4.3 Specifications in Previous Studies Most of the studies mentioned in Section 3.3 that estimate the credit supply curve have had to grapple with the endogenous debt variable. Goldstein and Woglom (1991) use state-fixed effects and also, to account for the time-series endogeneity, use demographic factors and lagged debt variables to instrument for debt. Based on a study by Metcalf (1993) that regresses debt on yields, Capeci (1994), Eichengreen and Bayoumi (1994), and Bayoumi et. al. (1995) instrument with percent of population under 18 and over 65, average number of people per household, and the level and growth of the population. These two studies also use the severity of fiscal restrictions as an instrument, although, as Poterba and Reuben (1999) explain, this instrument is highly questionable due to the fiscal restrictions’ own effect on bond yields. Overall, these instruments have had varying degrees of success, but in most cases, it seems possible that these instruments are affecting bond yields not only through their effect on debt. Estimating the nonlinearity of the credit supply curve is another concern running through these studies. Goldstein and Woglom (1991) use quadratic terms for the debt variable to account for the nonlinearity. Eichengreen and Bayoumi (1994) and Bayoumi et al. (1995) use similar methods of accounting for nonlinearity through a divisor term that attempts to account for the backward bend in the curve. As mentioned above, my study uses squared terms for both debt and unfunded liabilities to test for nonlinearity. Since there are no data points in the backward-bending section of the supply curve, the technique using a divisor-term seemed unnecessary. 17 Chapter 5. Data and Summary Statistics 5.1 Data My study uses state GO bond yields to test the effect of state liabilities on yields. A general obligation bond is guaranteed by the full faith and credit of the state and retains the highest level of security among state bonds. Revenue bonds, on the other hand, are guaranteed by a specific revenue source, such as an airport or toll road. I use GO bonds because these bonds reflect the fiscal health of the state government, which is the point of focus in this study. For bond yield data, my study uses Bloomberg’s 10-year state bond yield indexes. I use indexes instead of the active bond yield for both feasibility and comparability reasons. Many state bond yields are not actively traded and so it is difficult to find yields for all states on the same day. Furthermore, it is difficult to compare across different bond issuances because of the various characteristics, such as callability, insured status, and size of issuance, that differ between state bond issuances. The indexes created by Bloomberg address both of these issues. These indexes are calculated daily based on non-callable, uninsured bonds of the same tenor. These yields are built using the optionadjusted spread model and is based on pricing from the Municipal Securities Rulemaking Board, new issues calendars, and proprietary contributed prices.5 Bloomberg creates bond yield indexes for 19 states and so my study analyzes these 19 states. The bond yield spreads used in this study are the difference between these indexes and the constant maturity Treasury rates, which the Treasury calculates using the daily yield curve. 5 Bloomberg was unable to give more information on the construction of these indexes. Their 2012 release concerning the details the constructed municipal curves has not come out yet. 18 Data for the unfunded pension liabilities was taken from the Pew Center for the States. The Pew Center calculates these liabilities using information from states’ Comprehensive Annual Financial Reports (CAFRs) as well as the CAFRs of some pension plans. The Center also uses actuarial valuations, which are documents that detail the current and future cost of the pension plans. The Center’s unfunded pension liabilities calculations only include pension plans listed in the state CAFRs and exclude any pension plans in which the state does not have a financial interest. In some cases, the state allows local governments to participate in their pension plan, and so in these cases, municipal and local employees’ pension plans are part of the state’s financial interests and are included in the calculation for states’ unfunded pension liabilities. The Pew Center’s calculation is limited to only defined-benefit plans, in which participants are promised a defined benefit usually the rest of their lives, and cash balance plans, in which participants must give an annual contribution and are also guaranteed a set benefit, and does not account for defined-contribution plans, which are only a small minority of the total pension plans. The Center does not recalculate the liabilities, but keeps the states’ chosen discount rates. Despite any drawbacks, the Pew Center’s calculation of the unfunded liabilities is very comprehensive and systematic in its approach. The data for the debt variable is taken from Moody’s reports of states’ net taxsupported debt. Net tax-supported debt is any state debt “secured by state operating resources which could otherwise be used for state operations.” (Kimberly, 2010, p. 2) Both pension and debt data are taken from the states year-ends, which are usually June 19 30th of each year.6 June 30th unemployment figures are taken from the Bureau of Labor Statistics. The variable for fiscal constraints is a dummy variable for “no carryover restrictions,” hereafter referred to as NOCA laws. NOCA laws, a term coined by Lowry and Alt (2001), restrict the ability of states to carry a general fund deficit into the next fiscal year by requiring next year’s projected revenues to pay for the previous year’s general fund deficit. The laws generally mean that provisions and expenditures are nullified if they surpass the fiscal year’s revenues, although some NOCA law states, such as Georgia, do allow borrowing if the legislature increases taxes so that future revenues cover the additional expenditure within a given time period. (Alt and Lowry, 1994, p. 817) Although these laws are rarely legally enforced, they act as signals to investors of states’ future actions.7 Other states have weaker anti-deficit laws, but the NOCA law states, as identified by Lowry and Alt as well as the Advisor Commission of Intergovernmental Relations, are the states with strongest anti-deficit laws. The last data, which consists of state revenues and expenditures, comes from the National Association of State Budget Officers (NASBO). This organization releases annual expenditure reports that detail state spending areas and revenue sources. My study uses the general fund expenditures, general fund revenues, and total expenditures given in these reports between 1997 and 2009. It is important to note that the general fund revenues calculations only include taxes collected by the state for the general fund, and do not include any federal fund or state between-fund transfers, although states in 6 The only three states with different year-ends are Michigan, whose fiscal year ends on September 30th, New York, whose ends on March 31st, and Texas, whose ends on August 31st. 7 Alt and Lowry (1994) show that NOCA laws lead to faster responses by state governments after an economic shock and Crain (2000) shows that NOCA laws increase expenditure volatility. 20 their CAFRs may differ in their categorizations. This data set is not used in the main regression, but only in the IV regressions and the financial crisis analysis in Section 6.3. 5.2 Summary Statistics Table 5.1 presents the summary statistics for main data set. Although state bond yields have default risk and Treasuries do not, GO bond yields are on average about 60 to 80 basis points lower than Treasuries due to preferential in-state tax treatment. Unfunded pension liabilities and net tax-supported debt are each divided by the state GSP to compare the liability numbers across states. Unfunded pension liabilities are on average lower than debt as a percent of GSP, although unfunded pension liabilities have a greater variance. Looking at the pre- and post-crisis periods in Table 5.2, we can see that these two different periods had very different characteristics. As would be expected, the unemployment rate shot up and the variance between states did as well. Bond yield spreads changed from being negative pre-crisis to being positive post-crisis. This large change is presented in the graph in Figure 5.1. During recessionary times, such as 2002 and 2003, the spread narrows, but it is only in 2008 that the spread is positive. Figure 5.1: Average State Bond Yield, Treasuries, and Spreads 700 600 500 Basis Points 400 300 200 100 0 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 -100 -200 -300 Average 10-year State Bond Yield Average 10-year Treasury Spread 21 Table 5.1: General Summary Statistics GO Bond Yield* (%) Treaury Yield* (%) GO Bond Spread (%) Unfunded Pension Liabilities/GSP (%) Net Tax-Supported Debt/GSP (%) NOCA Unemployment (%) Mean Median St. Dev. 25 percentile 75 percentile 4.30 4.28 0.49 4.08 4.54 4.91 5.03 0.95 3.99 5.44 -0.78 -0.61 0.53 -0.93 -0.24 1.68 1.15 2.94 -0.36 3.37 3.22 2.46 1.95 1.91 4.10 0.32 0.00 0.47 0.00 1.00 5.22 4.90 1.69 4.30 5.90 Tables 5.2: Pre- and Post-Crisis Summary Statistics Pre-Crisis Post-Crisis Mean Median St. Dev. 25 percentile 75 percentile Mean Median St. Dev. 25 percentile 75 percentile GO Bond Spread (%) -0.77 -0.84 0.39 -0.95 -0.52 0.24 0.18 0.29 0.10 0.29 Unfunded Pension Liabilities/GSP (%) 1.24 0.93 2.63 -0.51 2.83 4.07 3.62 3.38 1.15 6.90 Net Tax-Supported Debt/GSP (%) 3.16 2.32 1.95 1.85 4.02 3.57 2.87 1.97 2.26 4.41 Unemployment (%) 4.80 4.70 1.10 4.10 5.50 7.55 7.15 2.40 5.30 9.30 Notes: The pre-crisis period is designated as June 30th, 1997 to June 30th, 2007 and the post-crisis period is designated as June 30th, 2008 and June 30th, 2009. 22 The pension and debt numbers also show a drastic change between the two periods. This change is especially noticeable for the unfunded pension liabilities number. The mean for unfunded pension liabilities jumps almost three percentages points whereas the mean for debt increases by less than a half a percentage point. While in the pre-crisis period the average unfunded pension liabilities is about 1.5 percentage points lower than average debt level, in the post-crisis period the unfunded pension liabilities becomes higher than the debt by half a percentage point for the mean and three quarters a percentage point for the median. Figure 5.2 shows the average unfunded pension liabilities and the average debt level side by side. Debt levels slowly rise between 1997 and 2009, but this seems less to do with the 2008 financial crisis and more to do with an increase in debt around 2004. Unfunded pension liabilities have also been increasing before the financial crisis, but at a much faster rate than debt levels. This upward trend appears to have intensified with the financial crisis. However, it is important to note, that increasing unfunded pension liabilities is not merely a result of the financial crisis. It is also useful to compare unfunded pension liability trends in states with low and high unfunded pension liabilities. Figure 5.3 divides states into states with average unfunded liabilities less than the mean (“Low Unfunded Pension Liability States”) and states with average unfunded liabilities greater than the mean (“High Unfunded Pension Liability States”). Although the both lines follow a similar trend, the high pension liability states trend is steeper, and by 2009, the difference between the two unfunded pension liability averages is almost six percentage points, up from less than two percentage points in 1997. 23 Figure 5.2: Average Unfunded Pension Liability and Average Debt 5 Average Unfunded Pension Liabilities Average Debt 4 Liability as % of GSP 3 2 1 0 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 -1 Figure 5.3: High and Low Unfunded Pension Liability States 10 Low Unfunded Pension Liability States High Unfunded Pension Liability States Difference Between High and Low Unfunded Pension LIability as % of GSP 8 6 4 2 0 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 -2 -4 Notes: Low unfunded pension liability states are states in the data set with an average unfunded pension liability of less than the mean average unfunded pension liability. High unfunded pension liability states’ average unfunded pension liabilities are above the mean. 24