ࡱ> q` \bjbjqPqP 4::H` :,>>>R'''82(4f(RV:6)6)6)6)6)***9999999$B;h=J9>/**//9>>6)6):e4e4e4/>6)>6)9e4/9e4e4>>e46)*) ׽k'1le46&:0V:e4=3=e4=>e4*du+e4s,?-f***994d***V:////RRR$v!RRRv!RRR>>>>>>  The Persistence of Tax Refunds: Evidence from Panel Data Sara LaLumia 206 South Academic Building Department of Economics Williams College Williamstown, MA 01267 Phone: 413-597-4886 Fax: 413-597-4045 Email: Sara.LaLumia@williams.edu Session Title: Public Economics and the Low-Income Population (CSWEP Session) Session Chair: Hilary Hoynes Discussants: Bill Evans, Hilary Hoynes, Thomas Buchmueller, Kosali Simon The Persistence of Tax Refunds: Evidence from Panel Data Sara LaLumia* Each year approximately three quarters of tax returns generate refunds. For these returns, the money withheld from paychecks during the year (or remitted in the form of estimated tax payments) exceeds the tax liability computed at the time the tax return is filed. Do tax refunds represent mistakes on the part of taxpayers uncertain about their annual tax liability at the time they choose their withholding? Or do taxpayers intentionally overwithhold to ensure that they will receive refunds? A standard intertemporal model implies that refund receipt is a mistake. In this model an individual prefers receiving a dollar in the current period to receiving a dollar in the future. A taxpayer who receives a refund is acting in exactly the opposite way, choosing to postpone the receipt of some income to a future period without receiving any compensating interest payment. Richard H. Thaler and H. M. Shefrin (1981) propose a model of self-control that can explain intentional refunds. An individual who would like to save, but who has difficulty doing so because of self-control problems, will value commitment devices that limit short run consumption. Overwithholding can function as a commitment device in two ways. Overwithholding reduces the amount of income immediately available for consumption, and it delivers that income later in the form of a large lump sum, out of which consumers tend to have a lower marginal propensity to consume. In this paper I use a 12-year panel of tax return data to investigate patterns of refund receipt over time. I argue that if refund receipt is primarily unintentional, then the average taxpayer should learn from each years tax position and adjust his subsequent withholding. In particular, receiving a refund in one year should be associated with reduced withholding in the following year. I find that receiving a large refund, one greater than five percent of income, is associated with a 0.45 percentage point decrease in the following years withholding as a share of wage income. Receiving a smaller refund, one between one and five percent of income, is also associated with a significant but smaller reduction in subsequent withholding. This pattern suggests that, on average, refund receipt is not intentional. At the same time, I find that there is a substantial group of taxpayers who receive refunds in all years of the panel. These taxpayers, with lower average incomes and wealth, may particularly value overwithholding as a commitment device and may be intentionally choosing to receive refunds. Previous research has used survey data to study withholding decisions. Matthew D. Shapiro and Joel Slemrod (1995) investigate taxpayer responses to a 1992 change in withholding rates that reduced refund amounts without changing tax liability. A taxpayer who wanted to preserve his original refund amount could do so by filing a new withholding certificate (W-4) with his employer. Fewer than 10 percent of survey respondents with wage income reported that they had done so. This could indicate that most taxpayers are not intentionally choosing to overwithhold. Alternatively, it could indicate a tendency to adhere to the default level of withholding. Michael S. Barr and Jane K. Dokko (2006) describe a survey of low- and middle-income Detroit residents that asked respondents whether they would prefer greater withholding and a larger refund, no change in withholding or refund amount, or lower withholding and a smaller refund. Here there is substantial evidence that refund receipt is intentional. Nearly half of the respondents would maintain their current withholding, while slightly more than one third would increase withholding. I. Data Description The tax panel used in this paper was compiled by the Office of Tax Policy Research (OTPR) at the University of Michigan, and spans tax years 1979 to 1990. The tax filing units included in the panel are a subset of the units included in the annual cross sectional data released by the Statistics of Income (SOI) division of the IRS. Like the cross-sectional files, each year of panel data is representative of all returns filed during a particular year, including late returns being filed for earlier years. I focus only on timely returns. While the cross-sectional SOI files are stratified by income, inclusion in the panel is random. Selection is based on the last four digits of the primary filers Social Security number. The number of filing units present in all twelve years of the panel is 4982. The dataset includes much of the information reported on a tax return and accompanying schedules. Hence, there is a great deal of detail on income components and amounts, and reliable information on refund receipt. On the other hand, there is little demographic information. State of residence is reported if adjusted gross income (AGI) is below $200,000. Filing status can be used to proxy for marital status, and number of exemptions for household size. The sex of the primary taxpayer is reported in 1979 and 1980. Taxpayers who are over 65 (or blind) can claim a higher standard deduction. I focus on non-elderly households, which I define as filing units claiming no age or blindness exemptions in 1990, the last year of the panel. This reduces my sample size to 4031, and results in the oldest individuals in my sample being age 53 at the start of the panel. I use information on interest and dividend income to impute Figure 1: Years of Refund Receipt  measures of taxable wealth for each tax filing unit, following the approach of Daniel Feenberg and Jonathan Skinner (1989). Figure 1 describes the number of years of refund receipt for the filing units in my sample. Approximately 29 percent of filing units (1185 out of 4031) receive refunds in all 12 years of the panel. This is substantially higher than the 3 percent of filing units expected to receive 12 consecutive refunds if 75 percent of returns generated a refund in each year and refund receipt in one year was completely independent of refund receipt in other years. There is no spike at the other end of the distribution. Only about 1 percent of filing units in my sample never receive a refund across the 12-year span. This is particularly striking because it indicates that extremely few taxpayers behave in the way predicted by the standard intertemporal model, and that refund receipt may be an intentional choice for close to one third of filing units. Table 1 shows how persistent refund receipt varies with taxpayer characteristics. I divide the sample into men who were unmarried in 1979, women who were unmarried in 1979, and those who filed jointly in 1979. The modal number of refunds is 12 for all three groups. Table 1: Refund Persistence by Demographic Group NMean Years with RefundPercent with 12 RefundsA. Filing Status in 1979 Single Women53810.349.8 Single Men11539.733.5 Joint Filers23408.922.7B. Average AGI First Quartile10089.842.5 Second Quartile10089.734.8 Third Quartile10079.527.0 Fourth Quartile10088.213.3 However, the share with 12 refunds is 23 percent for the married group, 33 percent for the single men, and 50 percent for the single women. I also divide the sample into quartiles based on 12-year average AGI. Those in the lowest quartile are most likely to receive refunds. They are more than three times as likely to receive refunds in all 12 years of the panel than are filers with average AGI in the highest quartile. Grouping filing units on the basis of imputed wealth rather than AGI produces a similar result. Persistent refund receipt is substantially more common among low-wealth filers than among high-wealth filers. The patterns suggested by these simple descriptive statistics are also apparent in a regression framework. I have estimated probit regressions in which the dependent variable is an indicator for having received a refund in all years of the panel. Even controlling for filing status, the receipt of self-employment income and capital gains, the use of a paid preparer, and whether a filing unit itemized deductions, lower levels of income and wealth are associated with significantly higher probabilities of persistent refund receipt. II. Learning In this section I look for evidence that taxpayers learn about their tax liability over time. If the process of completing and filing a return in one year imparts information to the taxpayer that he can use to finetune his withholding for the following year, prior refunds may be associated with reductions in withholding and cases of prior balance due may be associated with increases in withholding. Such a pattern suggests that refund receipt is not intentional. When a taxpayer finds himself receiving a refund, he takes steps to reduce the chance of receiving a refund in the following year. Unfortunately I do not have any information from W-4 forms which would directly indicate changes in withholding. Instead, I construct a variable equal to total withholding of income tax divided by wage income. The presence of a few extreme outliers distorts the mean of this variable, and in what follows I drop observations for which the withholding share is greater than or equal to 2, as well as filing units with fewer than 12 years of non-zero wage income. After applying these restrictions, the mean of my constructed withholding share is 0.143, indicating that filing units withhold an average of 14.3 percent of wage income. Making use of the panel structure of my data, I estimate a fixed effects specification in which the withholding share in year t depends on year t income and demographics as well as on the year t 1 refund amount. I use information on lagged refund amount to group observations by whether they received a refund or owed a balance, and then into categories based on dollar amounts relative to a filing units average AGI: less than 1 percent of a filing units 12-year average AGI, between 1 and 5 percent of average AGI, and greater than 5 percent of average AGI. I include dummies for these categories, with the smallest refunds as the omitted category. I control for withholding share in year t 1, the number of exemptions claimed, whether filing status is joint, and year t AGI. I also include an indicator for returns filed after tax year 1986, because the Tax Reform Act of 1986 changed marginal tax rates significantly. Coefficients from the set of refund dummies are shown in Table 2, with standard errors in parentheses. The pattern of coefficients is consistent with taxpayer learning. Those who receive a large refund have lower subsequent withholding shares, and those who owe a large balance have higher subsequent withholding shares. Receiving a refund greater than 5 percent of ones average AGI is associated with a 0.45 percentage point decline in the following years withholding share. Receiving a refund between 1 and 5 percent of ones average AGI is associated with a significant but smaller decline (0.26 percentage points) in subsequent withholding. Owing a balance of more than 5 percent of ones average AGI is associated with a 0.92 percentage point increase in subsequent withholding share, and owing a balance between 1 and 5 percent of average AGI is associated with a 0.26 percentage point increase in withholding. The pattern that large refunds are associated with subsequent declines in withholding while large balances due are associated with subsequent increases in withholding is robust to measuring refund size in dollar terms rather than as a share of average AGI. Although not shown in the table, the other regressors predict withholding share in sensible ways. The single best predictor of year t withholding share is the withholding share in year t 1. As expected under a progressive tax, withholding share increases with AGI. Withholding shares are significantly lower in years after 1986, consistent with the reduction in tax rates experienced by many filers as a result of the Tax Reform Act of 1986. Both an Table 2: Effect of Previous Refund or Balance Due on Withholding Share Baseline (1)Effect Differs with Mobility(2)DummiesInteractionsLarge Refund-0.0045** (0.0008)-0.0032** (0.0009)-0.0057** (0.0018)Medium Refund-0.0026** (0.0006)-0.0016* (0.0007)-0.0053** (0.0015)Small Balance Due-0.0003 (0.0009)0.0004 (0.0010)-0.0009 (0.0023)Medium Balance Due0.0026** (0.0009)0.0036** (0.0010)-0.0009 (0.0021)Large Balance Due0.0092** (0.0013)0.0132** (0.0015)-0.0111** (0.0029) additional personal exemption and joint filing status reduce average tax liability, and in the regression both are associated with a lower withholding share. While these results indicate that the average taxpayer responds to a large refund by reducing his subsequent withholding, the number of filing units receiving refunds in all twelve years of the panel suggests that the average may mask very different responses among different groups of taxpayers. To investigate this possibility, I interact the set of lagged refund dummies with a variety of taxpayer characteristics. I then estimate fixed effects regressions including both the original set of refund dummies, the set of interaction terms, and all of the income and demographic variables included in the baseline withholding regression. Motivated by the observation that lower-income individuals are more likely to receive refunds in all years of the panel, I first investigate whether learning differs by income. I interact the set of refund dummies with an indicator for having ever had AGI less than twice the poverty line. The coefficient on the dummy represents the withholding response of higher-income individuals, while the sum of the coefficients on the dummy and corresponding interaction term represents the withholding response of low-income individuals. Most of the coefficients on the interaction terms are close to zero and insignificant. Low-income filing units do not respond to large or medium refunds any differently than do higher-income filing units. The only significant difference is in response to a large balance due. Owing a balance of more than 5 percent of average AGI is associated with a 1.44 percentage point increase in subsequent withholding for higher-income filers and with only a 0.57 percentage point increase for low-income filers. I find no evidence that learning differs across taxpayers who have high and low levels of wealth, across taxpayers who use a paid preparer and those who do not, or across single men, single women, and married taxpayers. The one characteristic that I have found to be an important determinant of responsiveness to previous refund receipt is whether a filing unit has ever moved from one state to another. As shown in the second specification in Table 2, cross-state movers respond to a large or medium refund by reducing their withholding by a larger amount. Receiving a refund greater than 5 percent of average AGI is associated with a 0.32 percentage point reduction in withholding for most taxpayers, but with a 0.89 percentage point reduction for cross-state movers. On the other hand, cross-state movers adjust their withholding less than non-movers in response to a balance due. Owing a balance of more than 5 percent of AGI is associated with a significant 1.3 percentage point increase in withholding share among non-movers, but an insignificant and near-zero change in withholding among cross-state movers. David A. Jaeger et al. (2007) use survey data from Germany to document that individuals who are more willing to take risks are substantially more likely to migrate within Germany. If the same is true in the U.S., my results suggest that the reason for refund receipt may differ with ones attitude towards risk. Those most tolerant of risk show the most evidence of learning from prior refund receipt, suggesting that among this group refund receipt is largely unintentional. While other proxies for risk tolerance are available in my tax return data, such as whether an individual is involved in self-employment, cross-state mobility has the advantage of being largely uncorrelated with federal tax liability. III. Discussion The enduring popularity of tax refunds is puzzling in a standard intertemporal model. In this paper I show that about 30 percent of filing units received refunds in twelve consecutive years, and that persistent refund receipt is concentrated among those with low levels of income and wealth. This pattern suggests that refund receipt may be intentional for some taxpayers, particularly those with little saving who may benefit from the forced saving associated with postponing the receipt of some income until the tax refund season. At the same time, there is evidence that the average taxpayer learns about his tax liability over time and adjusts his withholding to move closer to a zero-balance position. Large refunds in one year are associated with lower withholding in the following year. References Barr, Michael S. and Jane K. Dokko. 2006. Tax Filing Experiences and Withholding Preferences of Low- and Moderate-Income Households: Preliminary Evidence from a New Survey. IRS Research Conference Proceedings. Feenberg, Daniel and Jonathan Skinner. 1989. Source of IRA Saving. Tax Policy and the Economy. 3: 25-46. Jaeger, David A., Holger Bonin, Thomas Dohmen, Armin Falk, David Huffman, and Uwe Sunde. 2007. Direct Evidence on Risk Attitudes and Migration. IZA Discussion Paper 2655. Jones, Damon. 2008. Inertia and Overwithholding: Explaining the Prevalence of Income Tax Refunds. Working Paper. Shapiro, Matthew D. and Joel Slemrod. 1995. Consumer Responses to the Timing of Income: Evidence from a Change in Tax Withholding. Review. 85(1): 274-283. Thaler, Richard H. and H. M. Shefrin. 1981. An Economic Theory of Self-Control. Journal of Political Economy. 89(2): 392-406. * Department of Economics, Williams College. 206 South Academic Building, Williamstown MA 01267. Email: Sara.LaLumia@williams.edu. I am thankful to seminar participants at Williams College, Colby College, and the National Tax Association for helpful suggestions, and to Maddie Jones for excellent research assistance. All errors are my own.  In a recent working paper, Damon Jones (2008) finds evidence of strong adherence to the default level of withholding. Using the same panel data that I use here, as well as SOI cross-sectional data, he finds that withholding levels respond very little to the 1992 federal change in withholding rates, to the gain or loss of a dependent exemption, or to the EITC expansions of the 1990s.  Although a single year of the panel is designed to be representative of the tax filing population, the set of filing units present in all years of the panel is not representative. Relative to those who exit, filing units appearing in all years are more likely to file jointly, have higher incomes, and are more likely to include someone over age 65. Some events that are commonly associated with dropping out of other longitudinal datasets are not an issue here. There is no reason to expect that movers would exit from the panel, as cross-sectional files are linked by Social Security number rather than by address. Changes in marital status do not automatically lead to exclusion from the panel. Suppose a couple is married in 1979 and divorces in 1980. Information from the 1979 joint return will be linked to subsequent single returns filed by the person who was listed as the primary taxpayer on the 1979 return.  Restricting attention to only those filing units reporting wage income in all years of the panel, as I do in subsequent analysis, produces a very similarly shaped histogram of years of refund receipt. About 34 percent receive refunds in all years and less than 1 percent receive zero refunds.  Ordering households on the basis of average imputed wealth rather than average AGI produces a substantially different sorting. Only about one-third of households fall into the same wealth and income quartiles.  In a longer version of this paper I use duration analysis to model the probability of exiting from a spell of refund receipt as a function of income and demographic variables. The results are qualitatively similar. Higher levels of income and wealth are associated with shorter spells of refund receipt.  While taxpayers may elect to have a third party withhold tax out of other types of income such as interest payments, the large majority of withholding is done by employers, out of wage income.  If I consider only years before 1987, the pattern that a large refund is associated with significantly lower subsequent withholding while a large balance due is associated with significantly higher subsequent withholding is still apparent, although the coefficients on the refund dummies are somewhat smaller.  This result is robust to defining the low-income group as those in the lowest quartile of average AGI or as those who had AGI less than twice the poverty line in six or more years.      CERSb     + t   LM(cUVW˿׻׷׷׷˷˷˳ӯӯӫh hObh~hdhNh~?hFhShDhr h:hmh9~h: h-h; hc;p how"0Jhow"CJ aJ how"hACJ aJ how"hAhj;  CDERSTUVWXYZ[\]^_`a$a$gdow"gdNzP\\ab~  U r   dgd vW d`gd vW d`gd; d`gdQf$a$gdow"gdow"gdNzERe ";234p   q !!f"#f#g#j#ƾ´ưΨΨΔΔjh"0JUhhh0Rh"jhUhh?h%h[jh* 0JUh 2h* hh+Ih hXh-h vWhL9h9BhRbjhgQX0JUh~?h hOb3q j#w$$$$$$$$ $$Ifa$gdx $Ifgdxgd edgd edgd" $da$gd$a$gddgdQfj#k##($>$v$%&&)&+&3&N&V&n&&&&''''W(X(Z(>)?))*%*F*r*s*t*u*v*****+9+{++++,S,T,U,,..../V/Y/r//ਞh$31hMjh 0JUhhehLDhmIh hNzh jh 0JUh9Rh7jh e0JUhfnh0h%hj/hfhfh/%h e;$$$$$$^UIII $$Ifa$gdx $Ifgdxkd$$Ifs\3 v b t0644 sa$$% %%%^UIII $$Ifa$gdx $Ifgdxkd$$Ifs\3 v b t0644 sa%%#%(%,%1%^UIII $$Ifa$gdx $Ifgdxkd $$Ifs\3 v b t0644 sa1%2%@%E%I%N%^UIII $$Ifa$gdx $Ifgdxkd!$$Ifs\3 v b t0644 saN%O%^%_%`%a%^UIII $$Ifa$gdx $Ifgdxkdp"$$Ifs\3 v b t0644 saa%b%r%w%{%%^UIII $$Ifa$gdx $Ifgdxkd7#$$Ifs\3 v b t0644 sa%%%%%%^UIII $$Ifa$gdx $Ifgdxkd#$$Ifs\3 v b t0644 sa%%%%%%^UIII $$Ifa$gdx $Ifgdxkd$$$Ifs\3 v b t0644 sa%%%%%%^UIII $$Ifa$gdx $Ifgdxkd%$$Ifs\3 v b t0644 sa%%%?)u*v**^VVNFFdgd%dgd7dgd ekda&$$Ifs\3 v b t0644 sa*,Y/27d9e999999 $$Ifa$gdx $Ifgdxgdh d`gd'Y d`gd  d`gdQf / 0-000p001.19111223<3a3>445666777788=999999999;;;<<z<<=|>?@@@@@@:A[A\A̶̺̮hGRh'hh}$vh;BhOhNzhB0hhjh%0JUh%h)KhU\h'YhZFhXoh hMhhfhth/h 'AhGheh$31799999pg[[ $$Ifa$gdx $IfgdxkdD'$$Ifl4FW8Z i  t06    44 la999999pgg[[ $$Ifa$gdx $Ifgdxkd($$Ifl4FW8Z i t06    44 la99:::!:*:4:=:]THHHHHH $$Ifa$gdx $Ifgdxkd)$$Ifl4\W8Z iN t0644 la=:>:L:V:_:h:q:{::^UIIIIII $$Ifa$gdx $Ifgdxkd)$$Ifl\W8Z iN t0644 la:::::::::^UIIIIII $$Ifa$gdx $Ifgdxkd*$$Ifl\W8Z iN t0644 la::::::; ;;^UIIIIII $$Ifa$gdx $Ifgdxkdp+$$Ifl\W8Z iN t0644 la;;%;.;7;@;I;S;\;^UIIIIII $$Ifa$gdx $Ifgdxkd.,$$Ifl\W8Z iN t0644 la\;];^;;BIII^RJRRBBdgddgdh d`gd'Ykd,$$Ifl\W8Z iN t0644 la\A]AABBBBBBB*CeCfCCC!DFFGHHHIIJKLLLLLLLLMMMM N&N1N2NNTOUOOOWPsPPPPPɹ h vW0Jh\h-h)h)6hXh98h986h98hohc;phgGh)h6ho~h>+zhhB0huhOh`h%hk,hOh;*h'hjhk,0JU4ILLLLMM2N3NNNTOUOPPPPQdSW)XX3ZZ.\\\gd"dgd)gdNzdgdPQQdSeSWW)X*XXY3Z4ZZZ.\/\\\\\\\\\\\\h\jhc5Uhc5jh vW0JUh vW\\\\\\\\\\\gdNz 21h:pQf/ =!"#$% DdNN\  C 8A RefundHistYoung2^QBz.hc_0:D`!2QBz.hc_0 PL7x ՙOWW,A[@D>P4fmmq^j[LܙLfǙ{Yd_Itܛɘ$.lD:wN}BgZ̟s9꫊Bq8{pp[UDL--Q8L8- 6v]tO/%.[UKQ]P%9_ZT46oA*cՒ&o<׈VZYg#1-uF!viF5_+Dq-$ƜSnG%V{, MX` "iڰ|~ExSw3fgd}}|U9E5.iR+*\4۱ճn?ֳsȁl籨*Xܯ﩯wԏ?"-6ܑ:؄:5;<.IK1KCgZִ gSjZ,H4GV|{cE޺ZR s!~}L&MbWh)Oz&*m̡_#Z@_C{7U.lDb 'n1m^ ?q^f uqO8 3IN@+b!t Nj&ʨrWUj56t*Oks`uNR~} 3S3_Q/5.K**~Υ = ZwM'PssJ٦>%;nnqV0S0Os\U 4x+] J zT$337)m-d+T>S3wYo*>znOcslza]ɕP!sN6v u.qkH6i?H{_*}͝1FS YMͬMie~s1;_$sUbhd.`.6tX?5SzvU3U7AѤN$p>u;|4F0GnuQ=ݦMݧL :ushLP z:p;C"}'y|zk(@2Yދ@˭dAEݍt]ވĽX';f^_ȮҹG\IC?୭?B蹛f=h t$MlqqrwX˺*jDk,Ngbe47 ^gŠ|:TKL2%Ѻ1;Me*kcbM"vlowvM{j|Sv(veejmwMpc7ʽ_8,OWn_؋esy}uߕP ]kF1s5{Ψш/y]ì_5)II$ IEG} &mFdx=,= r,I=MM2*4QQR1J/,4ɮH* \we|HXc$f\5FSG[F%S_ƪ}q]Pk>4?G;Pg嫿3Ҵʴʾt:Zζ}<ޑe8Yٹ`:rwus1v38̺{2;3uawbq檌Ysx>Ί/ԱmȎ2Qߤ;:왲Ñ+D3++4#מ7U&ٲ.ⲲT՚ȓi1ֳҞ܄ѱ,ٹԾ3ќ793'/Ӝ3N=//4Ωԙvb,7;˞k9$'ӛ[/řϜ۱o0wo=ʋ톹8qȣ#ҭȤ7Jߑjd)&9Ѱ_j鳾p$!ԝ0o`̲Ydܞ`JXٱ.t,ܘyڴڳϏ۳vوF!{<4^mzq??6='(xKtWg{ev)&K1+3-JFvȳE~w, ߶3"f"fjM%f,If+i_Y'(_/)ez.ahd2NL ^ XvŘ.W/=+ILf{e'׆ze\:,stwӍ=8tIevqg7߈4տ%nJYY/r :W.l=-PTUje0,/PT_0_ bc5T?W**5cWRaV~ }T?W**5Wld,5T?W**5?? VCSa¬RsIVO J 5JYL:'˦RaVa}RaV~?-kw OuQfeۣwoƫ.Ll~/T\ND\̝7:;sz1W$A_LO9K(g/ -~eʹC -' ~, ɅHq=ğt#ߏ#r9?/F^E/x{EuʌRq\qWTgby*.pX_x @-rMx{@\{<?O3igp j\bmMxpWvE -^j^B KU|W213Ǜ <܀m>!`|@\A{S>ԱoHKl7X'2 X@- ՠ/^_匫d|5qT|%,ieT,qm8#ȇlp!!u:QcDBiδ>:*M)ɟ;۰ՊσU"ىPZ,?98ٰ'?'w8ؠ|<)!ȇM (~V"V`x;a?o [p#ov.r-ZՊ>8383ůkA_K? t/h@^D܀`#1Z$!{A5JD7oߛops`?<{nb;m[93࿅lv[z+\op  #s5~A}~{w߁}~ɱ\;ȹܻ:r+x_q#li=u[GրU`ϩ~?× _>/ǟ釟Wݰ긊v5ݳ>*ܴܳҽ֡_˸݁֯2^!ɽע}ד́Wʰ],hh8N{ QϩsjY}\r 0Cx9C: OQt<3v hY(G3n OO>-Ga]=Xz//ޡo#;O>Ux6߁wa8P~'b?XUn&^B'\nǶ ·;[ve h/m+7߿TBbND?Eb Zw߉}'wcj!^}B'(s dlMTGaC(d'p!x54RA3J@3>i׀,#C 먠†}$~X Ú KTxF&"P 9j:D|2xWf^JݦSTp/r3nƛ;nM7kn M jToy8݁'F0?]<wCn7sgy]Puz }+]-5q`Fezͺ7>)eL?ګ]yp}>вͮlԐm W#H?|^5W Q$Ʊ88N!nOgۊd0JsT8vq1l '=6%8nX[O>'5'&O&0v!x G0y Njm'>nGm ƀ`G/G\sG?a{@ﻌ{B~|#ڃCkM_a,rcX5A|fos ﰂe5U>!tGmSן w%YAFX)8N ,V:A~θv^|ɓq1aO9ނgyZig 9N:Icb͌[`/Ik34ҴuWeIz9-mX+k}3Vٓޝ5mlrCZ|-mܔCnM&wm6ie Xݝ}3U5-țo¾A8+/g9+^}þXiḼr'9{KnB<5==j25J|˲S7YvA=f0IJWB;"o;6-7c, p3l䭖W]KYe>&<вr|㻜q+ȷRގۖ`-xy=_ fye+!`3a}l%V+1[卖"u6. lImpTc. OaX+v[>ܱʳE`3^rGΩ+dIr.'rpJb5t`'>S ໗{[ȷ1i"oF kmbfoqz)w;`v,X%[`y'N%XB6j쑯۬m|w=f3]"^`}M~˺Q` oFRh].!ao> a¸!պ y m k P>iXc`_:ftb o`9 A~(5y갮¯fƭu؍܂mɺ߽̅ tC߂Pasѫ8=}3'1O1*1<*B*phB'B ow"Comment ReferenceCJaJ<"< ow" Comment TextCJaJ@j!"@ ow"Comment Subject5\HBH ow" Balloon TextCJOJQJ^JaJ>@R> ow" Footnote TextCJaJ@&@a@ ow"Footnote ReferenceH*fW r"&X1\9TV{y p _ b T CDERSTUVWXYZ[\]^_`ab~Ur qjw #(,12@EINO^_`abrw{?!u"v""$Y'*/d1e1111111111111111222!2*242=2>2L2V2_2h2q2{2222222222222223 333%3.373@3I3S3\3]3^33:AAADDDDEE2F3FFFTGUGHHHHIdKO)PP3RR.TTTTTTTTTTTT0000000000000000000000000000000000000000000000000000@000000@0 @0 @0@0 @0@0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 @0 000000000000@0 00 @0 0 0 0 0 0 0 0 0 0 0 @0 @0@0 @0@0 @0@0 @0 @0 @0@0 @0@0 @0@0 @0 @0 @0@0 @0@0 @0@0 @0 @0 @0@0 @0@0 @0@0 @0 @0 @0@0 @0@0 @0@0 @0 000000000000000000000@0@0@0@0@0@0@0@0@00_ @0I0@0I0@0I0@0I0I0 12NOab$Y'*1111=2>2222233@3I3S3\3]3DHIT0@000000 00 00 00 0 0 0  0 0  0 00 00 0 0 00 0000000 00 00 00 0 0 0  0 0-.0-0-0  0 00@0 0 j#/\AP\/35AJLa$$%1%N%a%%%%%*999=:::;\;I\\0246789:;<=>?@BCDEFGHIKM\18@0(  B S  ? TK$<[#dSJd$#o#$)T# |[# ( QTQ  !""#$$ *%#?????HHH2I;IDIDIJIT      #?????HHH:IBIIIQIQIT  B *urn:schemas-microsoft-com:office:smarttagscountry-region>*urn:schemas-microsoft-com:office:smarttags PostalCode9*urn:schemas-microsoft-com:office:smarttagsState8*urn:schemas-microsoft-com:office:smarttagsCity=*urn:schemas-microsoft-com:office:smarttags PlaceName=*urn:schemas-microsoft-com:office:smarttags PlaceType9*urn:schemas-microsoft-com:office:smarttagsplace X  JQkq s z 6B V#^#//BBEEEEEFKFZF`FFFFFGGrGyGH H"H)HHIITTTTTTTTTTTTTT?!D!11^3h3I5Q5550787DEEEEE F'F3FFFFEGSGUGzGGGH*HWHtHHHHHHPPTTTTTTTTTTTTT3333333333333333333333q11111^3>?AADD(F3FEGUGGHuHHHTTTTTTTTTTTTTHTTTTTTTTTTTTT&S;*U\s:L9 6* ,W%Nfno)K'YXoM}v-; ow"%'k,.j/B0$31 2Ev2c5`8i:N<3x=?~? 'AB}-CmRCLDZFgG+ImI0R9RGR* W vWgQX@Z( b!bObRb e fQfHmf g7.gh2jFkxjkfnc;pGrfht}$vxAx|x>+zNz9~o~'[Yej&su`j7jefve]`F  W@X-$JP?/:D1a)098Of/%m/7;B`a- 9Bu7Ui~O\G2L2_2q2222222222333%373I3\3]3T@*~TP@UnknownGz Times New Roman5Symbol3& z Arial5& zaTahoma"1hR&#fGkˆd =% =%!4daHaH2HX ?0R2IOIT Sara LaLumiaOh+'0  0 < H T`hpxIOITNormalSara LaLumia100Microsoft Office Word@@E@~EC@ k =՜.+,0 hp  williams college%aH  I Title  !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNPQRSTUVWXYZ[\]^_`abcdeghijklmnopqrstuvwxyz{|}~Root Entry F4ӶkData O-1Tablef=WordDocument4SummaryInformation(DocumentSummaryInformation8CompObjq  FMicrosoft Office Word Document MSWordDocWord.Document.89q