WIXLIB

information expense and investing in the stock market. For any ratio of cost to income abovethe line, the household will not wish to undertake the information expense, and will havehigher expected utility by investing only in riskless assets. The ratio of the information costsufficient to deter investment in stocks relative to first period income is highest for collegeeducated households. Although these households have thehighest first period income, theutility gains to stock market investment for these households are sufficiently large that itwould take a large information expense todeter their participation. At each level ofeducation, the income cost is a decreasing function of the degree of risk aversion. For anydegree of risk aversion, the introduction of (uncorrelated) income risk lowers the informationcost sufficient to deter investment. The addition of a bequest motive, however, increases theinformation cost. Households that care not only about their own future consumption but planto leave a bequest as well have a greater incentive to participate in the stock market and takeadvantage of the expected equity premium, and consequently the information cost required todeter their pa icipationi shigher.3. Specification of the model of stock market participationIn the standard Consumption Capital Asset Pricing Model (CCAPM), demands foreach asset are determined jointly, and the specification of asset demands should considersimultaneously the portfolio choice across all possible asset combinations. When we allowfor the possibility of incomplete portfolios where not all assets are held, the number ofpossible asset combinations is 2“-1, which becomes very large as the number of assets (n)rises. Studies which have tried to estimate the simultaneous decision have resorted to veryaggregated portfolio combinations to make the estimation feasible IUhler and Cragg (1971),King and Leape (1984), Perraudin and Sorensen (1991)]. However, if the aim is to estimatethe probability of ownership of a particular asset, then a reduced-form equation which does7

not includeother asset demandscan reestimated separately. This approachis also used byIoannides(1992) and Agell and Sorensen(1990).As in the standardCCAPM,householdsare assumedto maximizeutility of.consumption, and for estimation, we assume that the household’s indirect utility function canbe written as a linear function of household characteristics plus an error term ui. Letbe the indirect utility function when households invest in the stock market, and letu2i ‘*j ‘2[be the indirect utility function when households do not invest in the stock market.The Xi’sare observable variables pertaining to household i’s characteristics. These includecurrent wealth and income,and measuresof the parametersof the utility function,such asdegreeof risk aversion,rate of time preference,and the intentionof leavinga bequest. TheXI’Salso includevariablesthat may help explain apparentunderinvestmentin stocks ifinformationcosts are importantfor the stockholdingdecisio such as level of eductiion,access to financial market information, and age. The level of education may also beindicative of future income prospects, which in turn may influence portfolio decisions, asdemonstrated in Bertaut and Haiiassos (1996). The ui error terms include unobservedhousehold-specific factors that may be important for the stockholding decision. In practice,the indirect utility function is not observable, and we instead observe a dummy variable ofparticipation or nonparticipation in the stock market:.8

Di 1lfUzi C ‘ii ;that is, the household’s utility when holding stocks is higher than when not participating in thestock market, andDi Ootherwise.Then the probability thatDi 1 P( Uzi ‘Ii) P(X2,B2 MziC x,i ] ‘]i) P(u2i - U,i ‘lip j - ‘ i ) P(ci xi.3 ).With the assumption that the error terms Eiare normally distributed, this dichotomous choicemodel can be estimated by probit maximum likelihood.Since we have access to panel data with household portfolio choices in two years, wecan specify a bivariate model and allow for unobserved heterogeneity resulting fromhousehold-specific factors with correlation p between the disturbances for each household i:eit pi vi,; e,, - N((), 62)If the errors are jointly normally distributed, the equations for 1983 and 1989 can beestimated by bivariate probit. The variance of ei ,62, is normalized to 1 because only theratio of /(s can be identified by probit maximum likelihood.

The probabilities that enter the likelihood function are then given by the bivariate normalcumulative distribution function:.W’,: ho,,Prob (Y1 yil, Y2 yi2) @ (Wil,wi2, i*) JJQ(Zi] 9Zi 9pi*) ‘Zi,dziz-*whereWil (Zyi, -1 )p ,’xil}\?i, (Zyi, -1 )p2’xi2P*- 1)(2yi - l) l’x,] ’xi pf (Zyi,and-1 /2 (M.,: w.,; - 2p,*w,, w,,) /(1 - P,*)2 (wi],wi2, Pi*) e2Z(1 - pi )l’zis the bivariate normal density.4. Data from the 1983/1989 Panel Survey of Consumer FinancesThe data set used for estimation is the 1983 and 1989 panel of the Survey ofConsumer Finances. This panel data set contains detailed information on householdportfolios and incomes, and respondent attitudes towards financial risk, liquidity, use of credit,and masons for saving at two points in time. The sample used for this paper categorizes asnonstockholders in 1983 the 68 households who owned stocks only in the company in whicha household member was employed, and 75 such households as nonstockholders in 1989.Because costs associated with acquiring such stocks are minimal compamd with holding otherstocks, and returns to such stocks may have different covariance properties with householdlabor income, the behavior of these households may distort conclusions about variables,,10*

relevant for the decision to acquire stocks. Since we use several variables on respondentattitudes, the sample is restricted to households where the respondent is the same in bothyears. The resulting sample contains 1,368 households.Because observations are available for two periods in time, we can observe thestockholding decisions of households who hold stocks in both 1983 and 1989, those whoabstain from marketparticipationin both years, and those who changestockholderstatus.The percentage of the sample holding stocks in 1983 (34.7 percent) is about the same as in1989 (34.3 percent), but the composition of 1989 stockholders is not the same as in 1983.Nearly 25 percent of the 1983 stockholders in the sample did not hold stocks in 1989, aboutthe same percentage of 1989 stockholders who were nonstockholders in 1983 (see Table 1).5Specifying the model as a bivariate probit allows calculation not only of the probabilities ofthe four alternate stock market participation categories, but also conditional probabilities ofcontinued participation or nonparticipation.Table 1. Stockholder Composition in 1983 and or estimation of the 1983 and 1989 stockholding decisions, the set of respondentcharacteristics include dummy variables for sex, race, and marital status in 1983. Respondentage is included through a set of 10-year age-range dummies to investigate whether the abilityto overcome inertial reasons for nonparticipation--for example, through increased exposure toinformation--increases with age, or alternatively whether older households are deterred from11

stockholdingby the shorter investmenthorizonsthey may face. Dummyvariablesfocretirementin 1983,remainingretired in 1989,or becomingretired between 1983 and 1989tieincludedto capture these life-cyclechanges.Dummyvariablesare includedfor whetherthe householdhead had less than highschooleducationand if the head had a high schooldegree but no collegedegree, withobtaininga college degree as the omitted dummy variable. This ciistinction is important forthe extent to which higher education leads to greater financial sophistication and ability toacquire information necessary for market participation, but may also reflect long-run incomeprospects. More-educated households may have both a greater ability and incentive toovercome inetiial behavior by undertaking costs associated with stock market participation, asindicated in Chart 1. Chart 1 also shows that for any level of education, the addition of.income risk lowers the cost sufficient to deter stock market participation. To capture laborincome risk, dummy variables are included in the bivariate probit for having an occupationwith above or below average unemployment risk (see Data Appendix). A dummy variable isalso included for having a managerial occupation, since managers may have smallwopportunity costs to finding out about investment opportunities from being involved in relatedprofessional activities.Current household wealth is divided into financial and nonfinancial assets, to capturedifferences in portfolio decisions that may arise from liquidity considerations. Financial networth includes relatively liquid riskless assets, directly held stocks and shares in mutualfunds, bonds, trusts and other managed accounts, IRAs, and the cash value of life insurance,minus the total outstandingon consumerloans. Nonfinancialnet worth includesinvestment12

in real estate, automobiles and other consumer durables, net of any loans outstanding for theseassets. For both financial and nonfinancial net worth, the probit regression uses the log of thevalue. Current labor income is estimated as wage and salary income, income derived from aprofessional business or practice, unemployment or worker’s compensations payments, andincome from Social Security and other pensions, and is also included as the log of its value.Dummy variables are included for self-expressed risk aversion, measured by theindividual’s responses to questions in each year of how much financial risk he was willing totake for commensurate financial return. Households are coded as “willing to take aboveaverage or substantial risk to receive above average or substantial expected return,” “averagerisk to receive average return,” and the omitted dummy is “not willing to take any financialrisk.” As the calibrations in Chart 1 show, the utility gain from stock market participation issmaller for agents with greater risk aversion, so these agents can be deterred fromparticipation by relatively low costs of acquiring information. In a non-expected utilityframework that allows for “first-order risk aversion,” households can be deterred from stockmarket participation even if information is costless to them because their utility functions arenot differentiable at zero stock investment (Yaari (1987), Segal and Spivak (1990), andHaliassos and Bertaut (1995)). Dummy variables are also included for whether the householdconsidered itself to be credit constrained. Although Haliassos and Bertaut (1995) find that arestriction on borrowing against future income alone is not sufficient to deter marketparticipation,b relatively small information costs are required to eliminate utility gains fromparticipation by individuals for whom the constraint is binding. (Bertaut, 1993). FollowingCox and Jappelli (1993), a household was coded as being credit constrained if the respondent13

indicated that a request for credit had been denied, or less credit than requested had beengranted, and the household didnotsubsequentlyreceive thedesired amountof credit -elsewhere, or if the household did not apply for credit because it was expected that such arequest would be denied.For both years, a dummy variable is included for whether or not the respondent statedthat credit card balances were always paid in full. Paying offcredit card balances couldreflect either an absence of liquidity constraints or may be a measure of financial astuteness.’Dummy variablesare includedfor whetherthe majorityof wealth was inherited(in 1983)orwhethera sizable bequestwas received(in 1989),and whetherthe householdplannedto leavea bequest(in 1989). Stocks may be especiallydesirableassets for householdswith bequestmotives,becausecapital gains on bequeathedstocks escape taxation, As illustratedin Chart1, adding a bequest motive increasesthe informationcost necessa to deter stock marketinvestment. Additionally,if an inheritancereceivedwas partly in the form of stocks,thehouseholdneed not overcomeinitial inertialbehaviorto invest, and may continuetoparticipatein the market with little additionalexpendi m on information.Althoughthe holding perioddoes not matter for the standardCCAPM,a typical ruleof thumb is that stocks are considereda superiorinvestmentto bonds for householdswithinvestmenthorizons of five years or more. Dummy variables are included for whether therespondent indicated willingness to tie up funds to achieve above average financial return (in1983) and for having a main financial planning period of five or more years (in 1989).8Use of the panel data set allows for observation of life-style changes that may affectportfolio decisions. In addition to the variables on becoming or remaining retired, dummy14

variables are included for acquiring a new prima job or losing the 1983 prima job.Dummy variables are also included for changes in marital status: remaining married to thesame spouse as in 1983; becoming married between 1983 and 1989; and becoming divorced,widowed, or separated between the survey years. The omitted dummy is remaining single,Respondents who became married could increase stock ownership from marrying someonewho is already a stockholder, or from an increased interest in planning for the future. Wealso include variables measuring the (log of) the amount spent on financing educationbetween 1983 and 1989 and for whether the household bought a house during that period.These major expenses could be reasons for exiting the stock market.Details of all the explanatory variables are included in the Data Appendix. Table 2gives the characteristics of three composite respondents. A median or “typical” respondenthas all variables taken at their population median values.g This “typical” respondent has ahigh school education, is willing to take average financial risklO,has a low unemployment-riskoccupation, does not pay off credit card balances, did not inherit wealth, and does not intendto leave a bequest. The typical respondent is contrasted with composite individuals who havethe characteristics at the 25th percentile and the 75th percentile of the population. In contrastto the median household, the 25th percentile household is younger, is not willing to take anyfinancial risk, has an average unemployment-risk occupation, and lower financial andnonfinancial wealth and labor income. The composite household at the 75th percentile isolder, has college education, pays off credit card balances, plans to leave a bequest, and hashigher labor income and wealth.

5. Results from the 1983 and 1989 bivariate probit5.1. Coeflcient estimatesTable 3 presents the parameter estimates from the 1983-1989 bivariate probitregression. In both years, the parameterestimatesfor both financialand nonfinancialnet.worth are positive and significant. Non-financialwealth has a largercoefficientthan financialwealth in 1983 and a smaller coefficient 1989, suggesting no special role for the liquidity ofassets in predicting stock market participation. Household labor income is not significant ineither year. Results for variables capturing unemployment risk provide limited support for thenotion that labor income risk can influence market participation. Having an occupation withbelow average unemployment risk has a positive coefficient significant at the 9 percent levelin 1983, but an insignificant coefficient in 1989. The coefficient on having an occupationwith above average unemployment risk is not significant in either year. Having a managerialoccupation has an additional significant positive coefficient in 1983 but not 1989, suggestingperhaps that the information benefits of a managerial job were no longer present by 1989.All the dummy variables for education are highly significant, with higher levels ofeducation increasing the predicted probability of ever being a stockholder. The strength ofthese results even after controlling for wealth, current income, and unemployment risk,suggests that education captures a measure of the ability to process information about themarket and investment opportunities. However, because the level of education may also serveas a proxy for future expected resources, it is not possible to disentangle these two effects.The coefficients for both the dummy variable indicators of risk aversion (willingnessto take above average or significant risk for above average or significant financial return, and

willingness to take average risk for average return) are significantly different from the omitteddummy (not willing to take any financial risk) and from each other in both years. Cautionshould be used in interpreting these results, however, because there is considerable evidencethat self-expressed classifications of risk aversion are imperfect indicators of the respondent’sactual degree of risk aversion. Individuals may be describing their “ideal” portfolio, ratherthan portfolio choices they actually face.l 1 In addition, these respondent attitudes were notalways consistent over time. Many fewer sample households declared they were willing totake above average or substantial tisk in 1989, with nearly half of those giving this answer in1983 switching to average risk in 1989. Additionally, about 34 percent of those willing totake average risk in 1983 were not willing to take any risk in 1989. Although some of thesehouseholds that switched risk aversion category may have had adverse stock returnrealizations from the 1987 stock market drop, the majority of those switching from averagerisk to not willing to take risk were nonstockholders in both years, while the majority ofhouseholds that switched from above average risk to average risk were stockholders in bothyears.12The coefficients on the age-range dummies in

Survey of Consumer Finances to study the determinants of not holding certain assets in a particular year, and then estimate assetdemands. Age]] and Edin (1990) perform a similar exercise using data on Swedish households.3 King and Leape (1987) attribute the limited incidence of stockholding to the exogenous and random arrival of information .