home computers and educational outcomes: evidence from the nlsy97
TRANSCRIPT
HOME COMPUTERS AND EDUCATIONAL OUTCOMES: EVIDENCEFROM THE NLSY97 AND CPS*
ROBERT W. FAIRLIE, DANIEL O. BELTRAN and KUNTAL K. DAS*
Although computers are universal in the classroom, nearly 20 million children in theUnited States do not have computers in their homes. Surprisingly, only a few previousstudies explore the role of home computers in the educational process. Home computersmight be very useful for completing school assignments, but they might also representa distraction for teenagers. We use several identification strategies and panel data fromthe two main U.S. data sets that include recent information on computer ownershipamong children—the 2000–2003 Current Population Survey (CPS) Computer andInternet Use Supplements matched to the CPS basic monthly files and the NationalLongitudinal Survey of Youth 1997 (NLSY97)—to explore the causal relationshipbetween computer ownership and high school graduation and other educationaloutcomes. Teenagers who have access to home computers are 6–8 percentage pointsmore likely to graduate from high school than teenagers who do not have homecomputers after controlling for individual, parental, and family characteristics. Wegenerally find evidence of positive relationships between home computers andeducational outcomes using several identification strategies, including controlling fortypically unobservable home environment and extracurricular activities in theNLSY97, fixed effects models, instrumental variables, and including futurecomputer ownership and falsification tests. Home computers may increase highschool graduation by reducing nonproductive activities, such as truancy and crime,among children in addition to making it easier to complete schoo
I. INTRODUCTION
The federal government has made the provi-sion of computer and Internet access to school-children a top priority. Spending on the E-rateprogram, which provides discounts to schoolsand libraries for the costs of telecommunicationsservices and equipment, is roughly $2 billion peryear (Puma, Chaplin, and Pape 2000; UniversalServices Administration Company 2005). Re-cently, the U.S. Department of Education re-leased the National Educational Technology
*This research was funded by the William T. GrantFoundation and the Community Technology Foundationof California. The views expressed here are those of theauthors and not necessarily those of the William T. GrantFoundation, the Community Technology Foundation ofCalifornia, or the Board of Governors of the FederalReserve System. We thank seminar participants at Stan-ford University, University of North Carolina, ChapelHill, University of Connecticut, University of Arizona,University of Melbourne, The Federal Reserve Bank ofSan Francisco, the 2007 IZA/SOLE Transatlantic Meet-ings, W.T. Grant Foundation Economics and Technologyworkshop, CJTC workshop, the 2006 UC Regents Meet-ing at the UCSC Silicon Valley Center, and the 2007 Soci-ety of Labor Economists Meetings for helpful commentsand suggestions.
Fairlie: Professor, Department of Economics, Engineering2 Building, University of California, Santa Cruz, CA95064. Phone 1-831-459-3332, Fax 1-831-459-5077,E-mail [email protected]
Beltran: Economist, Division of International Finance,20th and C St. NW, Mail Stop #42, Washington,DC 20551. Phone 1-202-452-2244, Fax 1-202-872-4926, E-mail [email protected]
Das: Lecturer in Economics, Department of Economics,Engineering 2 Building, University of California SantaCruz, CA 95064. Phone 1-831-459-1343, Fax 1-831-459-5077, E-mail [email protected]
771
Economic Inquiry doi:10.1111/j.1465-7295.2009.00218.x(ISSN 0095-2583) Online Early publication September 3, 2009Vol. 48, No. 3, July 2010, 771–792 © 2009 Western Economic Association International
assignments. (JEL I2)l
ABBREVIATIONS2SLS: Two-Stage Least SquaresCIUS: Computer and Internet Use SupplementsCPS: Current Population SurveyGPA: Grade Point AverageNLSY97: National Longitudinal Survey of Youth 1997NELS: National Educational Longitudinal Survey
Plan as part of the No Child Left Behind Policy.The plan calls for increased teacher training in
technology, e-learning opportunities for stu-dents, access to broadband, digital content, andintegrated data systems (U.S. Department ofEducation 2004). Several state and local govern-ments and private programs have also createdone-to-one computing in selected schoolsthrough the provision of laptop computers toschoolchildrenandteachers.1 Inarecentnationalsurvey funded by the U.S. Department of Educa-tion, nearly all principals report that educationaltechnology will be important for increasing stu-dentperformanceinthenextfewyears,andaclearmajorityofteachersreportthattheuseoftechnol-ogy is essential to their teaching practices (SRIInternational 2002). The result is that nearly allinstructional classrooms in U.S. public schoolshave computers with Internet access, with anaverage of 3.5 computers per classroom (U.S.Department of Education 2005).
In contrast to the ubiquity of computers inthe classroom, nearly 20 million children, repre-senting 26% of all children in the United States,do not have computers in their homes. This dis-parity in access to technology at home or the so-called digital divide may have implications foreducational inequality. Surprisingly, however,the role of home computers in the educationalprocess has drawn very little attention in the lit-erature. There is also no clear theoretical predic-tion regarding whether home computers arelikelytohaveanegativeorpositiveeffectonedu-cational outcomes. Home computers are clearlyvery useful for completing school assignmentsand may facilitate learning through researchand educational software. The use of home com-puters may also alter the labor market returns tocompleting high school, ‘‘open doors to learn-ing’’ encouraging some teenagers to stay inschool(Cuban2001;Pecketal.2002),andreducecrime. On the other hand, home computers areoftencriticizedforprovidingadistractiontochil-dren through video games and the Internetor fordisplacing other more active forms of learning(Giacquintaetal.1993;Stoll1995),andtheInter-net makes it substantially easier to plagiarize andfind information from noncredible sources.Therefore, it is an empirical question as to which
Indeed, the few previous studies examiningthe relationship between home computers andeducational outcomes find somewhat mixedresults.2 Using the National Educational Longi-tudinal Study of 1988, Attewell and Battle(1999) find that test scores and grades amongeighth graders are positively related to homecomputer use. Using more recent data fromtheUnitedStates—the2001CurrentPopulationSurvey (CPS)—Fairlie (2005) finds a positivecross-sectional relationship between home com-puters and school enrollment. Schmitt andWadsworth (2006) also find evidence of a posi-tive relationship between home computers andperformance on the British school examinationsfrom analysis of the British Household PanelSurvey between 1991 and 2001. In contrast,Fuchs and Woessmann (2004) find a negativerelationship between home computers andmath and reading test scores using the interna-tional student-level Programme for Interna-tional Student Achievement database. Theconclusions drawn from this literature on therelationship between home computers andedu-cational outcomes are limited; however,because of the mixed results, the primary focusis on test scores instead of other educationaloutcomes and on the lack of a comprehensiveapproach to addressing the potential endoge-neity of computer ownership.3
The answers to whether home computersimprove educational outcomes and whetherthe effects are sizeable are especially importantin light of the large and persistent disparities inaccess to technology across racial, income, andother demographic groups. For example, esti-mates from the 2003 CPS indicate that roughlyone-half of all African American and Latinochildren and less than half of all children livingin families with incomes less than $30,000 haveaccess to home computers. In comparison, 85%
1. See Stevenson (1999), Lowther, Ross, and Morri-son (2001), Rockman et al. (2000), Silvernail and Lane(2004), Mitchell Institute (2004), and Urban-Lurain andZhao (2004), for example, and Keefe, Farag, and Zucker(2003) for a summary of numerous programs.
2. A larger literature examines the classroom impactsof computers. See Kirkpatrick and Cuban (1998) and Nollet al. (2000) for reviews of this literature.
3. Schmitt and Wadsworth (2006) estimate regressionmodels that include future computer ownership and findstatistically insignificant estimates. A finding that futurecomputer ownership has a positive relationship withachievement would raise concerns that computer ownershipsimply proxies for an unobserved factors such as educationalmotivation. Fairlie (2005) addresses the endogeneity issue byestimating instrumental variable models with cross-sectionaldata from the 2001 CPS. Computer ownership is found toincrease school enrollment among teenagers.
ECONOMIC INQUIRY772
of the two opposing forces dominates and themagnitude of any effects.
of white, non-Latino children and 94% of chil-dren in families with incomes greater than$60,000 have access to home computers.If home computers are an important input intothe educational process then disparities inaccess to technology may translate into futuredisparities in educational and labor marketsand other economic outcomes.4 Financial, in-formational, and technical constraints maylimit the optimal level of investment in personalcomputers among some families.
In this study, we contribute to the sparse lit-erature on the educational impacts of homecomputers by using the only two major U.S.panel data sets with recent information on com-puter ownership—the 2000–2003 CPS Com-puter and Internet Use Supplements (CIUS)matched to the CPS basic monthly files andthe National Longitudinal Survey of Youth1997 (NLSY97)—and employing several empir-ical strategies to attempt to identify the causaleffects of home computers on high school grad-uation and other educational outcomes. Thedetailed panel data available in the CPS andNLSY97 allow for the estimation of specifica-tions that include detailed home environmentcontrols, instrumental variables, fixed effects,and future computer ownership. We explorethe relationship between home computer andhigh school graduation, grades, school suspen-sion, and criminal activities and present a simpletheoretical model to shed light on potentialmechanisms. This comprehensive approachhas not been taken in the previous literature.
We find fairly consistent evidence that homecomputers have a strong positive relationshipwith high school graduation and additionaleducational outcomes. The estimated effectsof home computers are generally similar evenafter controlling for detailed, and typicallyunobservable, measures of the home environ-ment and extracurricular activities, instrumen-tal variables, and fixed effects. We also performseveral falsification tests with the data. Specif-ically, we do not find evidence of a strong rela-tionship between educational outcomes andfuture computer ownership, cable television,or the presence of a dictionary at home, whichmay be correlated with unobservables but can-not or are unlikely to have causal effects. The
estimates also suggest that home computersmay increase high school graduation partlyby reducing nonproductive activities, such astruancy and crime, among children.
II. THEORY
Before turning to the empirical results, wefirst present a simple theoretical model of highschool graduation that illustrates the potentialeffects of home computers. A linear randomutility model of the decision to graduate fromhigh school is used. Define Ui0 and Ui1 as theith person’s indirect utilities associated withnot graduating from high school and graduat-ing from high school, respectively. Theseindirect utilities can be expressed as:
Ui0 5 a0 þ b#0Xi þ c0Ci þ k0tðWi;CiÞþ hY0ðZi;CiÞ þ ei0ð1Þ
andUi1 5 a1 þ b#1Xi þ c1Ci þ k1tðWi;CiÞ
þ hY1ðZi;CiÞ þ ei1;ð2Þ
where Xi, Zi, and Wi are individual, parental,family, geographical, and school characteris-tics, Ci is the presence of a home computer,Y0 and Y1 are expected future earnings, andt is the child’s achievement (e.g., test score),and ei is an additive error term. Xi, Zi, andWi do not necessarily include the same charac-teristics. Achievement is determined by thecharacteristics, Wi, and the presence of com-puters is allowed to have different effects onthe utility from the two educational choices.Expected earnings differ between graduatingfrom high school and not graduating fromhigh school and are functions of the character-istics, Zi, and home computers.
In the simple model, there are three majorways in which home computers affect educa-tional outcomes. First, there is a direct effectof having a home computer on the utility ofgraduating from high school, c1. Personal com-puters make it easier to complete homeworkassignments through the use of word processors,spreadsheets, Internet browsers, and other soft-ware, thus increasing the utility from completingschoolwork (Lenhart et al. 2001). Althoughmany students could use computers at schooland libraries, home access represents the highestquality access in terms of availability and
4. See Noll et al. (2000) and Crandall (2000) for anexample of the academic debate over the importance ofthe digital divide, and Servon (2002) for a discussion ofpolicies addressing the digital divide.
FAIRLIE, BELTRAN & DAS: HOME COMPUTERS AND EDUCATIONAL OUTCOMES 773
autonomy, which may provide the most benefitsto the user. Access to a home computer may alsofamiliarize the student with computers increas-ing the returns to computer use in the classroomor increasing preparation for class (MitchellInstitute 2004; Underwood, Billingham, andUnderwood 1994). Estimates reported belowindicate that approximately nine of ten highschool students who have access to a home com-puter use that computer to complete schoolassignments. Furthermore, 46% of teachersreport that lack of student access to technol-ogy/Internet is a barrier to effective use of tech-nology in the classroom (SRI International2002), and results from school laptop programsindicate very high rates of use of these com-puters for homework (Mitchell Institute 2004;Stevenson 1999; Urban-Lurain and Zhao 2004).
Access to home computers may have anadditional effect on the utility of staying inschool beyond making it easier to finish home-work and complete assignments. In particular,the use of home computers may ‘‘open doorsto learning’’ and doing well in school (Cuban2001; Peck, Cuban, and Kirkpatrick 2002),and thus encourage some teenagers to gradu-ate from school. The use of computers at homemay also translate into more positive attitudestoward information technology potentiallyleading to long-term use (Selwyn 1998). Manyteachers report that educational technologyincreases outside class time initiative amongstudents (SRI International 2002).
Personal computers also provide utility fromgames, e-mail, chat rooms, downloading music,and other noneducation uses creating anopportunity cost from doing homework. Thehigher opportunity cost increases the utilityof not graduating from high school. Computersare often criticized for providing a distractionfor children through video games and the Inter-net or for displacing other more active forms oflearning (Giacquinta, Bauer, and Levin 1993;Stoll 1995).5 Fuchs and Woessmann (2004) findinternational evidence of a negative effect ofhome computers on test scores and suggest thatit may be due to the distraction from effectivelearning. On the other hand, the use of com-puters at home, even for these noneducationaluses, keeps children off the street, potentiallyreducing delinquency and criminal activities.
Decreasing these nonproductive activitiesmay decrease the utility of dropping out ofschool. The two opposing factors make it dif-ficult to sign the effect of computers on the util-ity from not graduating from high school, c0.
Another way in which personal computersaffect the high school graduation decision isthrough their effects on academic achieve-ment. Computers could improve academicperformance directly through the use of edu-cational software. By making it more efficientto gather information, revise papers, and per-form calculations, computers also could allowstudents to focus more on the content and sub-stance of school assignments. As noted above,previous research finds that home computersare associated with higher test scores (Attewelland Battle 1999; Schmitt and Wadsworth2006). Computers, however, may displaceother more active forms of learning anddecrease learning by emphasizing presentation(e.g., graphics) over content (Giacquinta,Bauer, and Levin 1993; Stoll 1995). The Inter-net also makes it substantially easier to plagia-rize and find information from noncrediblesources. Therefore, the theoretical effects ofcomputers on academic achievement, dt/dC,and thus on the utility from graduating fromhigh school, k1dt/dC, is ambiguous.
Finally, home computers and the skillsacquired from using them may alter the eco-nomic returns to completing high school. Itis well known that information technologyskills are becoming increasingly important inthe labor market. The share of employmentin information technology industries andoccupations, and the share of employees usingcomputers and the Internet at work have risendramatically over the past decade (Freeman2002). Computer skills may improve employ-ment opportunities and wages, but mainly incombination with a minimal educational cre-dential such as a high school diploma, imply-ing that dY1/dC . dY0/dC.
Focusing on the high school graduationdecision, we assume that the individual gradu-ates from high school if Ui1 . Ui0. The proba-bility of graduating from high school, yi 5 1, is:
Pðyi 5 1Þ5PðUi1.Ui0Þ5F½ða1�a0Þþðb1�b0Þ#Xiþðc1�c0ÞCi
þhðY1ðZi;CiÞ�Y0ðZi;CiÞÞþðk1�k0ÞtðWi;CiÞ�
ð3Þ5. Computers may provide a similar distraction as
television, although Zavodny (2006) does not find evi-dence of a negative effect of television on test scores.
ECONOMIC INQUIRY774
where F is the cumulative distribution func-tion of ei1 � ei0. Following the standard ran-dom utility model framework (McFadden1974), the equation can be estimated witha logit regression by assuming that ei1 � ei0has a type I extreme value distribution. InEquation (2.3), the separate effects of com-puters on the probability of graduating fromhigh school are expressed in relative terms.Home computers have a direct effect on thegraduation probability through relative utilityand indirect effects through improvingachievement and altering relative earnings.
Unfortunately, identification of the sepa-rate parameters is difficult. Relying solely onnonlinearities for identification of the separatefunctions is likely to produce unstable esti-mates, so preferably, we would like Z andW to contain variables that are not includedin X.6 However, instead of making thesestrong structural form and distributionalassumptions and applying tenuous exclu-sion restrictions, we estimate the followingreduced-form model:
Pðyi 5 1Þ 5 F½aþ b#pi þ cCi�;ð4Þ
where p includes all individual, parental, fam-ily, and school characteristics.7 The direct andindirect effects of the variables on the highschool graduation decision are captured, butnot separately identified. In particular,although the more detailed assertions of thetheoretical model cannot be tested, the totaleffect of home computers on high school grad-uation can be estimated using Equation (2.4).The theoretical model does not provide a pre-diction regarding the sign or magnitude of theeffect of home computers on high school grad-uation, and thus we turn to an empiricalanalysis.
III. DATA
The data sets used in the analysis are thematched CIUS and monthly basic files tothe CPS and the NLSY97. These two paneldata sets are the only major U.S. panel datasets with recent information on computerownership and educational outcomes.
The CIUS, conducted by the U.S. CensusBureau and the Bureau of Labor Statistics,is representative of the entire U.S. populationand interviews approximately 50,000 house-holds. It contains a wealth of informationon computer and Internet use, includingdetailed data on types and location of use.To explore the relationship between computerownership and subsequent high school gradu-ation, we link CPS files over time to create lon-gitudinal data. Households in the CPS areinterviewed each month over a 4-mo period.Eight months later, they are reinterviewed ineach month of a second 4-mo period. Therotation pattern of the CPS makes it possibleto match information on individuals in a CIUSwho are in their first 4-mo rotation period(e.g., October 2003) to information from thesame month in their second 4-mo rotationperiod (e.g., October 2004). Thus, a 2-yearpanel can be created for up to half of all ofthe original CIUS respondents. To matchthese data, we use household and personalidentification codes provided in the CPS andremove false matches using age, race, andsex codes.
The NLSY97 is a nationally representativesample of 8,984 young men and women whowere between the ages of 12 and 16 on Decem-ber 31, 1996.8 Survey members were inter-viewed annually from 1997 to 2002. TheNLSY97 contains an over sample of 2,236black and Latino youth in the same age group.The NLSY97 contains information on com-puter ownership and detailed informationon educational outcomes, criminal activities,and individual and family characteristics.
IV. HOME COMPUTERS AND HIGHGRADUATION
Although access to computers in thenation’s schools is universal, access to homecomputers is far from 100% among children.Estimates from the 2003 CPS indicate thatslightly more than one-fourth of all childrenin the United States do not have access toa computer at home. Among children aged16–18 yr who have not graduated from highschool, slightly more than 20% do not haveaccess to a home computer (Table 1). Levelsof access to home technology are substantially6. It is also difficult to find a good measure of achieve-
ment and calculate predicted earnings for both educa-tional choices.
7. We are implicitly assuming, however, that Y(Z,C)and t(W,C) are separable in Z, W, and C.
8. See Center for Human Resource Research (2003)for additional details on the NLSY97 sample.
FAIRLIE, BELTRAN & DAS: HOME COMPUTERS AND EDUCATIONAL OUTCOMES 775
SCHOOL
lower for low-income and disadvantagedminority groups.9
Table 1 also reports estimates of patterns ofcomputer use among teenagers. Not surpris-ingly, teenagers use their home com-puters—94.6% of teenagers who have accessto a home computer use it. Computers alsoappear to be useful for completing schoolassignments. Conditioning on computer own-ership, only 81.6% of teenagers not enrolled inschool use computers at home compared to95.2% of enrolled teenagers. Among schoolenrollees who use home computers, 93.4%report using them to complete school assign-ments. Another interesting finding is that
71.1% of enrolled computer users use theircomputer for word processing, whereas only38.8% of nonenrolled computer users use theircomputer for word processing.
Teenagers also use home computers formany other purposes. The most commonuses of home computers among teenagersare for the Internet (86.9%), games (72.6%),and e-mail (78.2%). Use of home computersfor graphics and design (45.0%) and spread-sheets or databases (22.1%) in addition toword processing are also fairly common. Noneof these uses among high school students,however, is as prevalent as using home com-puters to complete school assignments. Con-cerns that home computers are only usedfor noneducational purposes such as playinggames, listening to music, and e-mailingfriends appear to be exaggerated.
At a minimum, estimates from the CPSindicate that home computers are useful forcompleting school assignments. Whether thesestudents wrote better reports or could havecompleted similar quality school assignmentsat a library, community center, or school,however, is unknown. Furthermore, the prev-alence of noneducational uses of home com-puters suggests that home computers mayalso provide a distraction that lessens or neg-ates their educational impact. We now turn toexamining the relationship between home com-puter ownership and high school graduation.
Table 2 reports estimates of high schoolgraduation rates by previous computer owner-ship. The CPS sample includes children aged
TABLE 1
Home Computer Use among Children Aged
16–18, CPS, 2003
AllChildren
Enrolledin School
NotEnrolled
Percent of childrenwith access toa home computer
79.6 81.1 56.4
Sample size 4,388 4,119 269
Percent of childrenwith access to a homecomputer who usethat computer
94.6 95.2 81.6
Sample size 3,543 3,392 151
Percent of homecomputer users who
Use computer forschool assignments
93.4 93.4
Use computer forthe Internet
86.9 87.4 74.5
Use computer forgames
72.6 72.9 64.9
Use computer forelectronic mail
78.2 78.8 62.9
Use computer forword processing
70.0 71.1 38.8
Use computer forgraphics and design
45.0 45.5 33.4
Use computer forspreadsheets ordatabases
22.1 22.3 16.1
Sample size 3,357 3,234 123
Notes: The sample consists of children aged 16–18 whohave not graduated from high school and live with at leastone parent. All estimates are calculated using sampleweights provided by the CPS.
TABLE 2
High School Graduation Rates, Matched CPS
(2000–2004) and NLSY97
No HomeComputer
HomeComputer Difference
High school graduationrate by second surveyyear CPS (%)
56.7 73.3 16.6
Sample size 308 1,419
High school graduationrate by age 19NLSY97 (%)
70.7 94.2 23.5
Sample size 659 3,280
Notes: The CPS sample consists of teenagers aged16–18 who have completed 11th or 12th grade, but havenot received a high school diploma in the first survey year.All estimates are calculated using sample weights providedby the CPS and NLSY97.
9. See Hoffman and Novak (1988). U.S. Departmentof Commerce (2002), Fairlie (2004), Goldfarb and Prince(2008), and Ono and Zavodny (2007) for more details ondifferences in computer and Internet use.
ECONOMIC INQUIRY776
16–18 yr who live with at least one parent andreport completing the 11th or 12th grade, buthave not graduated from high school witha diploma in the first survey year. Computerownership is determined in the first surveyyear, and high school graduation is deter-mined in the second survey year.10 Thus, thegraduation rate that we use is defined as thepercent of all teenagers at risk of graduatingby the second survey date who actually grad-uate by the second survey date. In theNLSY97, home computer access is determinedbetween the ages of 15 and 17 and high schoolgraduation is measured by age 19. Using thesedefinitions of high school graduation, we donot capture individuals eventually returningto complete high school or a General Educa-tional Development test (GED) after age 19 inthe NLSY97 or after the second survey year inthe CPS.11
For both measures, high school graduationrates are much higher among teenagers withaccess to a home computer than teenagerswithout access to a home computer. Estimatesfrom the CPS indicate that 73.3% of teenagerswho have home computers graduate fromhigh school by the following year comparedto only 56.7% of teenagers who do not havehome computers. Estimates from the NLSY97provide evidence of a similarly large differencein graduation rates. Nearly 95% of childrenwho had a home computer between the agesof 15 and 17 graduated from high school byage 19 compared to only 70.7% of childrenwho did not have a home computer.
Estimates from the CPS and NLSY97clearly indicate that teenagers with home com-puters are more likely to graduate from highschool than children without home com-puters. The difference in graduation rates islarge and not much smaller than differencesgenerated by extreme changes in parentaleducation or family income. Although these
not control for other factors, such as parentaleducation and family income, which are likelyto be strongly correlated with computerownership.
V. ESTIMATING THE EFFECTS OF HOMECOMPUTERS ON HIGH SCHOOL GRADUATION
To control for parental education, familyincome, and other characteristics, we estimateprobit regressions for the probability of grad-uating from high school using the two datasets. We discuss the results from the CPS first,which are reported in Table 3. All specifica-tions include the sex, race, immigrant status’and age of the child; number of children inthe household; family income; home owner-ship; region of the country; central city status;state-level unemployment rate; average expen-ditures per pupil; and dummy variables for theage requirements of compulsory schoolinglaws, in addition to home computer ownership(see Table A1 for means).12 For both themother and father, we control for presencein the household, education level, labor forcestatus, and occupation. All the independentvariables are measured in the first surveyyear before measurement of high school grad-uation. Mother’s and father’s education levelsgenerally have a positive effect (although notstatistically significant) on the graduation prob-ability, and home ownership has a positiveeffect on graduation. Latino children, boys,and children with many siblings are less likelyto graduate from high school, all else equal.
Home computers are associated with grad-uating from high school by the following year.The reported marginal effects on the homecomputer variable are large, positive, and sta-tistically significant. It implies that havinga home computer is associated with an 8.1 per-centage point higher probability of graduatingfrom high school.13 The effect of this variableon the probability of high school graduation isroughly comparable in magnitude to that
10. Regression estimates are not sensitive to excludingthe relatively small number of children reporting complet-ing 12th grade, but not graduating in the first survey year,or the children graduating with a GED.
11. Dropping out of school, however, is associatedwith a much lower probability of returning to and com-pleting high school. For example, estimates from theNLSY indicate that 50% of dropouts from 1979–1986returned to school by 1986 (Chuang 1997), and estimatesfrom the CPS indicate that only 42% of 22–24 yr olds whodid not complete high school received a GED (U.S.Department of Education 2001).
12. State-level unemployment rates are from Bureauof Labor Statistics (2002), and the age requirements forcompulsory schooling laws and average expendituresper pupil are from U.S. Department of Education (2002).
13. We also estimate a specification that includesInternet access at home in addition to home computer.The marginal effects estimate on home Internet accessis small, negative, and statistically insignificant.
FAIRLIE, BELTRAN & DAS: HOME COMPUTERS AND EDUCATIONAL OUTCOMES 777
estimates indicate large differences, they do
TABLE
3P
rob
it,B
iva
ria
teP
rob
it,a
nd
2S
LS
Reg
ress
ion
sfo
rH
igh
Sch
oo
lGra
du
ati
on
an
dH
om
eC
om
pu
ter
(Ma
rgin
alE
ffec
ts)
Ma
tch
edC
PS
,20
00
–2
00
4
Explanatory
Variables
Specification
(1)
(2)
(3)
(4)
DependentVariable
High
SchoolGraduation
HighSchoolGraduation
ComputerBivariate
HighSchoolGraduationBivariate
HighSchoolGraduation
Model
Type
Probit
Probit
Probit
Probit
2SLS
Fem
ale
0.0
64
9(0
.02
67
)0
.06
48
(0.0
26
7)
0.0
19
9(0
.01
72
)0
.06
46
(0.0
27
1)
0.0
61
8(0
.02
68
)
Bla
ck�
0.0
31
9(0
.04
60
)�
0.0
31
8(0
.04
59
)�
0.0
65
2(0
.03
59
)�
0.0
30
5(0
.04
77
)�
0.0
24
7(0
.06
63
)
La
tin
o�
0.0
99
7(0
.05
13
)�
0.0
99
7(0
.05
13
)�
0.1
27
9(0
.04
31
)�
0.0
97
4(0
.05
66
)�
0.0
85
7(0
.09
10
)
Imm
igra
nt
0.0
18
6(0
.05
64
)0
.01
86
(0.0
56
3)
�0
.00
51
(0.0
31
7)
0.0
18
9(0
.05
71
)0
.03
00
(0.0
62
5)
Fa
mil
yin
com
e:m
issi
ng
�0
.09
35
(0.0
54
4)
�0
.09
38
(0.0
54
3)
0.0
29
1(0
.02
65
)�
0.0
95
0(0
.05
49
)�
0.0
84
5(0
.06
91
)
Fa
mil
yin
com
e:$
15
,00
0to
$3
0,0
00
�0
.03
22
(0.0
55
8)
�0
.03
23
(0.0
55
8)
0.0
42
7(0
.02
34
)�
0.0
33
4(0
.05
66
)�
0.0
25
3(0
.06
73
)
Fa
mil
yin
com
e:$
30
,00
0to
$5
0,0
00
0.0
27
0(0
.04
79
)0
.02
67
(0.0
47
9)
0.0
71
5(0
.02
05
)0
.02
48
(0.0
55
1)
0.0
26
1(0
.09
10
)
Fa
mil
yin
com
e:$
50
,00
0to
$7
5,0
00
�0
.04
90
(0.0
55
2)
�0
.04
94
(0.0
55
0)
0.0
91
0(0
.02
06
)�
0.0
51
7(0
.06
38
)�
0.0
45
3(0
.10
24
)
Fa
mil
yin
com
e:g
rea
ter
tha
n$
75
,00
0�
0.0
09
3(0
.05
19
)�
0.0
09
7(0
.05
18
)0
.09
28
(0.0
26
1)
�0
.01
16
(0.0
58
1)
�0
.01
36
(0.0
91
3)
Ho
me
ow
ner
ship
0.0
89
9(0
.04
05
)0
.09
00
(0.0
40
5)
0.0
78
2(0
.02
81
)0
.08
82
(0.0
43
6)
0.0
85
6(0
.06
88
)
Mo
ther
—h
igh
sch
oo
lg
rad
ua
te0
.01
73
(0.0
48
6)
0.0
17
3(0
.04
86
)0
.06
84
(0.0
23
2)
0.0
15
1(0
.05
61
)0
.02
20
(0.0
92
9)
Mo
ther
—so
me
coll
ege
0.0
74
1(0
.04
87
)0
.07
41
(0.0
48
7)
0.0
95
7(0
.02
28
)0
.07
14
(0.0
60
0)
0.0
74
3(0
.11
70
)
Mo
ther
—co
lleg
eg
rad
ua
te0
.03
47
(0.0
57
8)
0.0
34
9(0
.05
78
)0
.08
34
(0.0
26
0)
0.0
32
3(0
.06
59
)0
.04
29
(0.1
10
1)
Fa
ther
—h
igh
sch
oo
lg
rad
ua
te0
.07
47
(0.0
51
2)
0.0
74
6(0
.05
12
)�
0.0
62
6(0
.03
75
)0
.07
54
(0.0
54
7)
0.0
77
4(0
.06
48
)
Fa
ther
—so
me
coll
ege
0.0
51
2(0
.05
55
)0
.05
11
(0.0
55
5)
0.0
20
9(0
.03
29
)0
.05
07
(0.0
56
9)
0.0
47
9(0
.05
94
)
Fa
ther
—co
lleg
eg
rad
ua
te0
.05
50
(0.0
61
0)
0.0
55
0(0
.06
10
)0
.05
70
(0.0
35
2)
0.0
54
5(0
.06
27
)0
.05
08
(0.0
63
4)
Ho
me
com
pu
ter
0.0
81
1(0
.04
14
)0
.08
19
(0.0
41
9)
0.0
96
1(0
.17
80
)0
.10
67
(0.5
11
0)
New
est
com
pu
ter
pu
rch
ase
din
firs
tsu
rvey
yea
r�
0.0
03
4(0
.03
68
)
Fa
ther
use
sIn
tern
eta
tw
ork
0.0
61
0(0
.02
32
)
Mo
ther
use
sIn
tern
eta
tw
ork
0.0
45
4(0
.02
12
)
An
oth
erte
ena
ger
pre
sen
tin
ho
use
ho
ld0
.05
00
(0.0
23
8)
Mo
ther
’so
ccu
pa
tio
nco
ntr
ols
Yes
Yes
Yes
Yes
Yes
Fa
ther
’so
ccu
pa
tio
nco
ntr
ols
Yes
Yes
Yes
Yes
Yes
q�
0.0
24
8(0
.28
55
)
R2/p
seu
do
-R2
0.1
17
40
.11
74
0.1
41
9
Mea
no
fd
epen
den
tv
ari
ab
le0
.70
50
0.7
05
00
.82
11
0.7
05
00
.70
50
Sa
mp
lesi
ze1
,71
11
,71
11
,71
11
,71
1
Notes:
Th
esa
mp
leco
nsi
sts
of
teen
ager
saged
16–18
wh
oh
ave
com
ple
ted
11th
or
12th
gra
de
bu
th
ave
no
tre
ceiv
eda
hig
hsc
ho
old
iplo
ma
inth
efi
rst
surv
eyyea
r.M
arg
ina
lef
fect
sa
nd
thei
rst
an
da
rder
rors
(in
pa
ren
thes
es)
are
rep
ort
ed.
All
spec
ifica
tio
ns
incl
ud
ea
con
sta
nt;
nu
mb
ero
fch
ild
ren
inh
ou
seh
old
;d
um
my
va
riab
les
for
ag
e,re
gio
n,
cen
tra
lci
tyst
atu
s,su
rvey
yea
r,ro
tati
on
gro
up
,mo
ther
’sa
nd
fath
er’s
pre
sen
cein
the
ho
use
ho
ld,a
nd
lab
or
forc
est
atu
s;th
est
ate
-lev
elu
nem
plo
ym
ent
rate
;ex
pen
dit
ure
sp
erp
up
il;a
nd
ag
ere
qu
irem
ents
of
com
pu
lso
rysc
ho
oli
ng
law
s.A
lles
tim
ate
sare
calc
ula
ted
usi
ng
sam
ple
wei
gh
tsp
rovid
edb
yth
eC
PS
.
ECONOMIC INQUIRY778
It is also less than one-half the raw differencein high school graduation rates reported inTable 2, indicating that computer ownershipis strongly correlated with the controls.
A. Additional Probit Estimates
One concern with these results is that somestudents may have limited exposure to recentlypurchased computers, thus reducing the esti-mated effect on high school graduation.Although the CPS does not provide informa-tion on the timing of when all computer pur-chases were made, it provides information onwhen the newest computer was obtained bythe family. To insure longer exposure tohaving a computer and to further elimin-ate concerns regarding reverse causation orjoint determination, we include an additionaldummy variable measuring whether the new-est computer was purchased in the first surveyyear (Specification 2). A problem with thismeasure is that a computer purchased in thefirst survey year may represent a replacementfor an older model. The marginal effects esti-mate on home computer, which now measuresthe relationship for computers purchased atthe latest in the year before the first survey year(or 21–34 mo before measurement of highschool graduation), is very similar to the orig-inal estimate. The interaction estimate is smalland statistically insignificant. Therefore, thelarge estimated relationship between homecomputers and high school graduation isnot sensitive to the inclusion of recently pur-chased computers.
Although not reported, we also estimatea specification that includes the number ofcomputers per person in the household. A lim-itation of the data, however, is that the measureof the number of computers in the CPS is cen-sored at 3. Thus, we include a per capita mea-sure for households with one or two computersand a dummy variable for three or more com-puters. We find a large, positive, and nearly sta-tistically significant marginal effects estimateon the per capita computer measure. We alsofind a positive and statistically significant esti-mate on the dummy variable for three or morecomputers. Although we do not have completeinformation on the number of computers, theresults indicate that the level of access to homecomputers is also associated with the probabil-ity of graduating from high school.
As a falsification test, we also examinewhether cable television is associated witha higher probability of graduating from highschool. The 2003 CPS includes informationon whether the household has cable television.Similarly to computer ownership, cable televi-sion may be correlated with unobserved familywealth or permanent income. Because we donot expect access to cable television to increasethe probability of high school graduationamong teenagers, the finding of a similarlysized estimate as the one for home computersmay indicate that the estimated home com-puter effect is simply capturing the correlationwith an unobserved family characteristic. Wefind a small and statistically insignificant mar-ginal effects estimate on the cable televisiondummy variable when it is included alone orin addition to the home computer dummy var-iable. The home computer estimate remainslarge, positive, and statistically significant.
B. Bivariate Probit Estimates from the CPS
Although the probit models include numer-ous controls for individual, parental, and fam-ily characteristics, estimates of the effects ofhome computers on high school graduationmay be biased. For example, if children withhigher levels of academic ability or childrenwith more ‘‘educationally motivated’’ parentsare more likely to have access to home com-puters, then the probit estimates may overstatethe effects of home computers on high schoolgraduation. On the other hand, if parents ofchildren with less academic ability or timeto spend with their children are more likelyto purchase computers, then the probit esti-mates may understate the effects. In eithercase, the effects of unobserved factors, suchas academic ability and parental motivation,may invalidate a causal interpretation of theprevious results.
A potential solution to this problem is toestimate a bivariate probit model in whichequations for the probability of high schoolgraduation and the probability of havinga home computer are simultaneously esti-mated. We exclude dummy variables forwhether the child’s mother and father usethe Internet at work and whether anotherteenager is present in the household fromthe equation determining high school gradua-tion. The exclusion of these variables is useful
FAIRLIE, BELTRAN & DAS: HOME COMPUTERS AND EDUCATIONAL OUTCOMES 779
implied by being a girl or owning a home.
for identification, improving the stability of
estimates.14 These three variables shouldaffect the probability of purchasing a com-puter, but should not have a large effect onhigh school graduation (after controlling forfamily income, parental education, parentaloccupations, and number of children). Inter-net use at work may be associated with higherearnings, but this effect should be controlledfor by the inclusion of family income. Simi-larly, the presence of an additional teenagermay increase demand for home computersbecause of high rates of use for this age group,but it is unlikely to have a large effect on highschool graduation after controlling for thenumber of children in the household.
Before discussing the bivariate probitresults, we provide some evidence on the valid-ity of these exclusion restrictions by examiningcorrelations with having a home computer andhigh school graduation (reported in Table 4).Computer ownership rates are higher whenthe mother uses the Internet at work, thefather uses the Internet at work, and there isanother teenager present in the household,indicating that all three instruments arestrongly correlated with having a home com-puter. In addition, controlling for other vari-ables in probit models, we find that theestimates on all instruments are individuallyand jointly statistically significant with lowp values. On the other hand, all of the instru-ments are uncorrelated with high school grad-uation rates after controlling for othervariables including home computer ownershipin probit regressions. In all cases, the marginaleffect estimates are small and statisticallyinsignificant. Although this is not a formal testof the validity of the exclusion restrictions, itsuggests that the excluded variables are corre-lated with home computers, but do not havea strong independent correlation with highschool graduation.
Estimates from the bivariate probit modelfor the probability of high school graduationand having a home computer are reported inSpecification 3 of Table 3. We first brieflydiscuss the results for the home computerequation reported in the first column of Spec-ification 3. The probability of owning a homecomputer generally increases with parental
wealth or permanent income and havean effect on the budget constraint or mayhave an effect on preferences for computersthrough pure tastes, exposure, perceived use-fulness, or conspicuous consumption. Familyincome and home ownership are also impor-tant determinants of owning a computer.The estimated positive relationships are likelyto be primarily due to their effects on the bud-get constraint through income and wealth;however, they may also be due to effects onpreferences. African American and Latinochildren have lower probabilities of havinga home computer than do white children, allelse equal.
All three excluded variables have large,positive, and statistically significant marginaleffects estimates in the home computer equa-tion. Father’s Internet use at work, mother’sInternet use at work, and having an additionalteenager increase the probability of havinga home computer by 6.1, 4.5, and 5.0 percent-age points, respectively.
The second column in Specification 3reports the bivariate probit results for the highschool graduation equation. The marginaleffects estimate on home computers remainslarge and positive but is no longer statisticallysignificant.15 The point estimate implies thatthe presence of a home computer increasesthe probability of school enrollment amongchildren by 9.6 percentage points. The magni-tude of the estimate is comparable to theprobit estimate. In fact, we cannot reject thenull hypothesis that the unobserved factorsaffecting home computer ownership and highschool graduation are uncorrelated (i.e., q5 0).The test statistic is very small providing evi-dence that the original probit estimates areconsistent and that estimation of the bivariateprobit may not be needed.
The finding of a positive bivariate probitestimate is consistent with estimates from ear-lier data for school enrollment. Using cross-sectional data from the 2001 CPS, Fairlie(2005) estimates a bivariate probit model forhome computer ownership and school enroll-ment and finds positive estimates. Only therelationship between home computers andschool enrollment among teenagers is
14. Identification is possible using the nonlinearity ofthe bivariate probit, but the estimates are not very stable toalternative specifications.
15. The number of parameters being estimated in thebivariate probit is more than double the original probitusing the same sample size.
ECONOMIC INQUIRY780
education. Education may be a proxy for
examined, however, because of the use ofcross-sectional data. Computer and Internetuse by the child’s mother and father are usedas excluded variables.
We also estimate the model with two-stageleast squares (2SLS) to investigate whether thechoice of functional form is driving the results(reported in Specification 4).16 In the 2SLSregression, the coefficient estimate on homecomputer is roughly similar in magnitude tothe bivariate probit marginal effects estimate(.1067 compared to .0961). The standard error,however, is large and the coefficient estimate isnot statistically significant. Although the statis-tical imprecision is troubling and we cannotrule out zero effects with the 2SLS estimates,we are at least reassured that the estimatesare similar to the bivariate probit estimates.The bivariate probit estimates do not appearto be driven simply by the functional form ofthe model. The results of a Hausman test alsoprovide no evidence that home computers areendogenous in the 2SLS model and that ordi-nary least squares (OLS) estimates are biased.The OLS estimates are very similar to the probitmarginal effects and are statistically significant.
Returning to the bivariate probit model, wealso check the sensitivity of the bivariateprobit estimates to various combinations ofexclusion restrictions. Although we do notfind evidence that the original probit estimatesare inconsistent, the analysis is useful for com-pleteness and addresses concerns that one of
the excluded variables is problematic. Specif-ically, we estimate bivariate probit models inwhich we remove mother’s Internet use atwork (which had the weakest relationship withhome computers), and use only father’s Inter-net use at work or the presence of anotherteenager as the exclusion restriction (seeTable 5). In all cases, the marginal effects esti-mate on home computer is large, positive, androughly similar in magnitude to the originalestimates. None of the estimates, however,is statistically significant. Overall, the homecomputer marginal effects estimate is not sen-sitive to the choice of exclusion restrictions inthe bivariate probit models.
As a final check of the sensitivity of thebivariate probit estimates, we add anotherexclusion restriction to the model. If networkeffects exist in the adoption of computers thenthe rate of computer ownership in the localarea should affect the probability of owninga computer (Goolsbee and Klenow 2002).At the same time, local levels of computerownership should not have a large effect onhigh school graduation rates after controllingfor education, family income, and home own-ership. Therefore, we use computer ownershiprates in the metropolitan area as an additionalexclusion restriction in the bivariate probit.Estimates are reported in Specification 4 ofTable 5. The addition of this exclusion restric-tion has little effect on the home computermarginal effects estimate.
The findings from the bivariate probit and2SLS models do not contradict our originalfindings of a positive association between hav-ing a home computer and graduating from
TABLE 4
Selected Statistics for Excluded Variables, Matched CPSs, 2000–2004
Home Computer High School Graduation
Probit Probit
RawDifference
MarginalEffect
WaldStatistic p value
RawDifference
Marginaleffect
WaldStatistic p value
Father uses the Internet at work 0.2248 0.0624 5.22 0.0223 0.0594 �0.0475 1.55 0.2126
Mother uses the Internet at work 0.1828 0.0501 4.55 0.0330 0.0793 0.0081 0.05 0.8197
Another teenager presentin household
0.0494 0.0528 4.51 0.0337 �0.0429 0.0019 0.00 0.9579
Joint significance test of allexclusion restrictions
12.99 0.0047 1.75 0.6260
Notes: See notes to Table 3. Probit regressions include the instrument (alone), and the independent variables listed inTable 3.
16. The first-stage regressions are not reported, butinclude the same controls and additional variables asthe home computer equation reported in Specification 3.
FAIRLIE, BELTRAN & DAS: HOME COMPUTERS AND EDUCATIONAL OUTCOMES 781
high school from probit regressions. Althoughthe estimated magnitude of the relationship isroughly similar in the probit, bivariate probit,and 2SLS models, there is no evidence ofcorrelated unobservables, and the bivariateprobit estimates are not sensitive to differentestimation techniques and exclusion restric-tions, we are still left with some uncertaintybecause of the lack of precision in the bivariateprobit and 2SLS estimates. We now turn to ananalysis of the relationship using data fromthe NLSY97.
C. Estimates from the NLSY97
Estimates from probit regressions for theprobability of graduating from high schoolusing the NLSY97 are reported in Table 6.The dependent variable equals 1 if the individ-ual graduates from high school by age 19.Computer ownership is measured betweenages 15 and 17 and most other variables aremeasured in the first survey year, 1997.17 Allspecifications include similar individual,parental, and family characteristics as in theCPS specifications. In addition to these con-trols, we include dummy variables for moredetailed living arrangements, whether thechild’s mother was a teen mother, whetherany grandparent is a college graduate, house-hold net worth, and a continuous measure ofhousehold income in Specification 1. Highschool graduation generally increases withparents’ and grandparents’ education, house-hold net worth, and household income.
The NLSY97 provides additional evidenceof a strong positive relationship between com-puter ownership and high school graduationafter controlling for individual, parental,and family characteristics. The marginaleffects estimate on home computer is large,positive, and statistically significant.18 Havinga home computer as a teenager is associatedwith a .0685 higher probability of graduatingfrom high school.19 The estimate implies a largerdifference in graduation probabilities thaneither having a college graduate mother orhaving a college graduate father (relative to highschool dropouts).
The NLSY97 also includes information onreligion and private school attendance. Weinclude these measures as additional controlsin Specification 2. Their inclusion has littleeffect on the home computer marginal effectsestimate. To further account for potentialunobserved factors correlated with havinga home computer, we add two typically unob-servable measures of the home environment inSpecification 3—whether a language otherthan English is spoken at home and whetherthere is a quiet place to study at home.Although the estimate is insignificant at con-ventional levels, speaking another language athome is associated with a lower probability ofgraduation. The marginal effects estimate onwhether there is a quiet place to study is very
TABLE 5
Additional Bivariate Probit Regressions (Marginal Effects), Matched CPS, 2000–2004
Explanatory Variables
Specification
(1) (2) (3) (4)
Home computer 0.0633 (0.1918) 0.0675 (0.1985) 0.0922 (0.1773) 0.0852 (0.1703)
Excluded variables
Father uses Internet at work 0.0674 (0.0268) 0.0680 (0.0237) 0.0602 (0.0261)
Mother uses Internet at work 0.0464 (0.0230)
Another teenager presentin household
0.0494 (0.0240) 0.0512 (0.0247) 0.0500 (0.0236)
MSA-level home computer rate 0.2019 (0.1195)
q 0.0296 (0.3102) 0.0226 (0.3256) �0.0182 (0.2828) �0.0067 (0.2740)
Sample size 1,711 1,711 1,711 1,711
Note: See notes to Table 3.
17. Children living alone in 1997 are excluded fromthe sample.
18. We also estimate separate regressions that includeinteractions between home computers and race, income orgender and, in almost all cases, do not find large, statisti-cally significant interaction effects.
19. We find a larger positive marginal effects estimatewhen the dependent variable is high school graduation inthe last survey year, 2002.
ECONOMIC INQUIRY782
TABLE
6
Pro
bit
Reg
ress
ion
sfo
rH
igh
Sch
oo
lG
rad
ua
tio
n(M
arg
ina
lE
ffec
ts),
NL
SY
97
Explanatory
Variables
Specification
(1)
(2)
(3)
(4)
(5)
Fem
ale
0.0
225
(0.0
065)
0.0
222
(0.0
065)
0.0
224
(0.0
064)
0.0
214
(0.0
063)
0.0
210
(0.0
063)
Bla
ck0.0
251
(0.0
070)
0.0
278
(0.0
071)
0.0
266
(0.0
070)
0.0
246
(0.0
071)
0.0
245
(0.0
071)
Lati
no
�0.0
037
(0.0
098)
�0.0
128
(0.0
113)
�0.0
014
(0.0
117)
�0.0
007
(0.0
116)
�0.0
004
(0.0
116)
Asi
an
0.0
350
(0.0
116)
0.0
336
(0.0
121)
0.0
342
(0.0
110)
0.0
336
(0.0
109)
0.0
333
(0.0
112)
Imm
igra
nt
�0.0
165
(0.0
156)
�0.0
194
(0.0
162)
�0.0
157
(0.0
154)
�0.0
148
(0.0
152)
�0.0
141
(0.0
151)
Liv
esw
ith
mo
man
dst
epd
ad
�0.0
338
(0.0
152)
�0.0
307
(0.0
149)
�0.0
303
(0.0
147)
�0.0
273
(0.0
142)
�0.0
261
(0.0
141)
Liv
esw
ith
dad
an
dst
epm
om
�0.0
425
(0.0
331)
�0.0
470
(0.0
346)
�0.0
466
(0.0
341)
�0.0
439
(0.0
333)
�0.0
425
(0.0
329)
Liv
esw
ith
mo
mo
nly
�0.0
346
(0.0
110)
�0.0
324
(0.0
108)
�0.0
331
(0.0
108)
�0.0
324
(0.0
107)
�0.0
313
(0.0
106)
Liv
esw
ith
dad
on
ly�
0.1
173
(0.0
396)
�0.1
140
(0.0
400)
�0.1
176
(0.0
406)
�0.1
165
(0.0
404)
�0.1
165
(0.0
403)
Liv
esw
ith
gu
ard
ian
�0.0
637
(0.0
237)
�0.0
632
(0.0
238)
�0.0
631
(0.0
237)
�0.0
632
(0.0
237)
�0.0
625
(0.0
236)
Mo
mw
as
teen
ager
at
firs
tb
irth
�0.0
128
(0.0
088)
�0.0
114
(0.0
087)
�0.0
111
(0.0
086)
�0.0
112
(0.0
085)
�0.0
112
(0.0
085)
Mo
ther
hig
hsc
ho
ol
gra
du
ate
0.0
032
(0.0
077)
0.0
022
(0.0
077)
0.0
019
(0.0
076)
0.0
014
(0.0
076)
0.0
008
(0.0
076)
Mo
ther
som
eco
lleg
e0.0
217
(0.0
082)
0.0
206
(0.0
083)
0.0
201
(0.0
082)
0.0
188
(0.0
082)
0.0
185
(0.0
083)
Mo
ther
coll
ege
gra
du
ate
0.0
474
(0.0
083)
0.0
468
(0.0
083)
0.0
459
(0.0
081)
0.0
452
(0.0
081)
0.0
451
(0.0
081)
Fath
erh
igh
sch
oo
lgra
du
ate
0.0
274
(0.0
074)
0.0
273
(0.0
074)
0.0
257
(0.0
073)
0.0
247
(0.0
073)
0.0
245
(0.0
073)
Fath
erso
me
coll
ege
0.0
248
(0.0
083)
0.0
257
(0.0
082)
0.0
247
(0.0
081)
0.0
235
(0.0
081)
0.0
233
(0.0
082)
Fath
erco
lleg
egra
du
ate
0.0
409
(0.0
084)
0.0
407
(0.0
083)
0.0
399
(0.0
082)
0.0
378
(0.0
084)
0.0
376
(0.0
084)
Gra
nd
pare
nt
coll
ege
gra
du
ate
0.0
234
(0.0
091)
0.0
243
(0.0
089)
0.0
234
(0.0
089)
0.0
228
(0.0
088)
0.0
227
(0.0
088)
Ho
use
ho
ldn
etw
ort
h(1
0,0
00s)
0.0
013
(0.0
006)
0.0
014
(0.0
006)
0.0
013
(0.0
006)
0.0
013
(0.0
005)
0.0
014
(0.0
006)
Ho
use
ho
ldn
etw
ort
hsq
uare
d(1
0,0
00s)
0.0
000
(0.0
000)
0.0
000
(0.0
000)
0.0
000
(0.0
000)
0.0
000
(0.0
000)
0.0
000
(0.0
000)
Ho
use
ho
ldin
com
e(1
0,0
00s)
0.0
045
(0.0
028)
0.0
042
(0.0
028)
0.0
039
(0.0
028)
0.0
038
(0.0
028)
0.0
037
(0.0
028)
Ho
use
ho
ldin
com
esq
uare
d(1
0,0
00s)
�0.0
001
(0.0
002)
�0.0
001
(0.0
002)
�0.0
001
(0.0
002)
�0.0
001
(0.0
001)
�0.0
001
(0.0
002)
Pri
vate
sch
oo
l�
0.0
230
(0.0
152)
�0.0
227
(0.0
150)
�0.0
219
(0.0
148)
�0.0
216
(0.0
148)
Oth
erla
ngu
age
spo
ken
at
ho
me
�0.0
225
(0.0
153)
�0.0
226
(0.0
154)
�0.0
228
(0.0
154)
Qu
iet
pla
ceto
stu
dy
inh
ou
seh
old
0.0
036
(0.0
096)
0.0
040
(0.0
096)
0.0
028
(0.0
095)
Yo
uth
att
end
sex
tra
class
es0.0
189
(0.0
066)
0.0
184
(0.0
067)
Dic
tio
nary
pre
sen
tin
ho
use
ho
ld0.0
168
(0.0
147)
Ho
me
com
pu
ter
by
age
17
0.0
685
(0.0
133)
0.0
691
(0.0
134)
0.0
679
(0.0
133)
0.0
648
(0.0
130)
0.0
632
(0.0
129)
Rel
igio
nd
um
mie
sN
oY
esY
esY
esY
es
Pse
ud
o-R
20.2
300
0.2
344
0.2
366
0.2
389
0.2
389
Mea
no
fd
epen
den
tvari
ab
le0.9
028
0.9
028
0.9
027
0.9
025
0.9
027
Sam
ple
size
3,7
15
3,6
73
3,6
70
3,6
50
3,6
48
Notes:
(1)
Th
esa
mp
lein
all
spec
ifica
tio
ns
con
sist
so
fte
enager
sli
vin
gw
ith
thei
rp
are
nts
in1997.
(2)
All
spec
ifica
tio
ns
incl
ud
ea
con
stan
t,n
um
ber
of
chil
dre
nin
the
ho
use
ho
uld
,d
um
my
va
ria
ble
sfo
rth
eq
ua
rter
of
bir
th,
reg
ion
,ce
ntr
alci
tyst
atu
s,a
nd
mis
sin
gca
teg
ori
esfo
rso
me
va
riab
les,
an
dsc
ho
olsi
ze,st
ud
ent-
tea
cher
rati
oa
nd
loca
lu
nem
plo
ym
ent
rate
.
FAIRLIE, BELTRAN & DAS: HOME COMPUTERS AND EDUCATIONAL OUTCOMES 783
small and statistically insignificant. The addi-tion of these home environment controlshas no effect on the estimated relationshipbetween home computers and high schoolgraduation.
As a final sensitivity check, we estimatea specification that includes a dummy variableindicating whether the child takes extra classesor lessons, such as music, dance, or foreignlanguage lessons. This variable is likely torepresent a good proxy for educational moti-vation. Indeed, we find a positive and statisti-cally significant marginal effects estimate onthe variable. Even after controlling for thisvariable, however, we continue to find a strongpositive relationship between access to a homecomputer and high school graduation.
Estimates from the NLSY97 indicate thathome computers are associated with morethan a .06 higher probability of graduatingfrom high school, which is similar in magni-tude to the estimates from the CPS. These esti-mates are extremely robust to controlling forthe exceptionally rich set of individual, paren-tal, family, and home environment character-istics available in the NLSY97.
D. Dictionaries as a Falsification Test
The NLSY97 provides another falsificationtest for interpreting the estimated relationshipbetween home computers and high schoolgraduation. The NLSY97 includes informa-tion on whether a dictionary is present inthe household. It is likely that the presenceof a dictionary is correlated with the educa-tional motivation of the family, but it isunlikely that dictionaries have a large effecton educational outcomes. A dictionary maybe useful for completing some school assign-ments, but it is unlikely to have a discernableeffect on the likelihood that a child graduatesfrom high school. Specification 5 of Table 6reports estimates from a model that includesthe home dictionary variable. The marginaleffects estimate on the presence of a dictionaryat home is statistically insignificant and ismuch smaller than the home computer esti-mate. The home computer marginal effectsestimate is now .0632, which is only slightlysmaller than the previous specification.Finally, we find a small and statistically insig-nificant estimate on the presence of a dictio-nary at home when we include it without
provide additional evidence that is consistentwith the hypothesis that the presence of homecomputers increases the probability of gradu-ating from high school.
E. Grades and Home Computers
Estimates from the CPS and NLSY97 indi-cate a strong positive relationship betweenhome computers and high school graduation;however, we know very little about the under-lying causes of this relationship. The similarityof the bivariate probit results and the rich setof controls included in the NLSY97 regres-sions suggest that the relationship is not solelydriven by an unobserved factor. An examina-tion of the relationship between home com-puters and additional educational outcomesmay shed some light on the underlying causesof the relationship and provide further evi-dence on the educational impacts of homecomputers.
The NLSY97 includes information onoverall grades obtained in high school, whichcan be used to estimate the student’s gradepoint average (GPA). The theoretical modelpresented above indicates that home com-puters may increase GPAs by making it easierto complete school assignments, keeping chil-dren out of trouble, or increasing interest inschoolwork. On the other hand, home com-puters may decrease GPAs by providing a dis-traction through video games or emphasizingpresentation over content.
Table 7 reports estimates for linear regres-sions for GPAs.20 The mean GPA in the sam-ple is a 2.8 or roughly a B average. We includethe same sets of control variables as thosereported in Table 6. Home computers areassociated with higher GPAs. The coefficienton home computer is large, positive, and sta-tistically significant. It corresponds to anincrease of .216 points, which is roughlytwo-thirds the value of a plus or minus grade.The implied effect is comparable in magnitudeto having a college-educated mother.
In Specifications 2–4, we include the addi-tional measures of religion, private school,
20. The measure of GPA in the NLSY97 is categoricalcapturing major cutoffs. We also estimated an orderedprobit model with fewer independent variables and findsimilar results as the linear regression. We find that homecomputers have a positive and statistically significant rela-tionship with GPAs.
ECONOMIC INQUIRY784
the home computer variable. These results
TABLE
7
OL
SR
egre
ssio
ns
for
Hig
hS
cho
ol
GP
A,
NL
SY
97
Explanatory
Variables
Specification
(1)
(2)
(3)
(4)
(5)
Fem
ale
0.3
51
4(0
.02
31
)0
.35
02
(0.0
232
)0
.35
34
(0.0
232
)0
.34
15
(0.0
233
)0
.34
00
(0.0
233
)
Bla
ck�
0.1
40
6(0
.03
29
)�
0.1
425
(0.0
345
)�
0.1
451
(0.0
345
)�
0.1
590
(0.0
345
)�
0.1
602
(0.0
345
)
La
tin
o�
0.0
48
8(0
.03
57
)�
0.0
566
(0.0
376
)�
0.0
580
(0.0
439
)�
0.0
581
(0.0
438
)�
0.0
586
(0.0
439
)
Asi
an
0.2
30
7(0
.09
19
)0
.21
89
(0.0
919
)0
.21
76
(0.0
928
)0
.23
56
(0.0
935
)0
.23
38
(0.0
935
)
Imm
igra
nt
0.0
285
(0.0
510)
0.0
312
(0.0
513)
0.0
245
(0.0
517)
0.0
412
(0.0
516)
0.0
432
(0.0
516)
Liv
esw
ith
mo
ma
nd
step
da
d�
0.1
50
4(0
.03
95
)�
0.1
475
(0.0
399
)�
0.1
451
(0.0
399
)�
0.1
392
(0.0
397
)�
0.1
396
(0.0
397
)
Liv
esw
ith
da
da
nd
step
mo
m�
0.2
10
4(0
.08
00
)�
0.2
254
(0.0
805
)�
0.2
279
(0.0
804
)�
0.2
078
(0.0
805
)�
0.2
072
(0.0
806
)
Liv
esw
ith
mo
mo
nly
�0
.12
29
(0.0
318
)�
0.1
216
(0.0
320
)�
0.1
214
(0.0
321
)�
0.1
191
(0.0
320
)�
0.1
189
(0.0
321
)
Liv
esw
ith
da
do
nly
�0
.21
68
(0.0
643
)�
0.2
178
(0.0
649
)�
0.2
191
(0.0
649
)�
0.2
216
(0.0
648
)�
0.2
234
(0.0
648
)
Liv
esw
ith
gu
ard
ian
�0
.07
76
(0.0
556
)�
0.0
808
(0.0
562
)�
0.0
814
(0.0
562
)�
0.0
784
(0.0
562
)�
0.0
806
(0.0
562
)
Mo
mw
as
teen
ager
at
firs
tb
irth
�0
.11
41
(0.0
332
)�
0.1
048
(0.0
334
)�
0.1
043
(0.0
334
)�
0.1
051
(0.0
334
)�
0.1
047
(0.0
334
)
Mo
ther
hig
hsc
ho
ol
gra
du
ate
0.0
72
3(0
.03
39
)0
.07
42
(0.0
341
)0
.07
37
(0.0
341
)0
.07
74
(0.0
341
)0
.07
55
(0.0
341
)
Mo
ther
som
eco
lleg
e0
.11
65
(0.0
394
)0
.11
37
(0.0
396
)0
.11
47
(0.0
396
)0
.10
86
(0.0
395
)0
.10
69
(0.0
395
)
Mo
ther
coll
ege
gra
du
ate
0.2
27
4(0
.04
59
)0
.22
08
(0.0
460
)0
.22
35
(0.0
459
)0
.21
71
(0.0
459
)0
.21
55
(0.0
459
)
Fa
ther
hig
hsc
ho
ol
gra
du
ate
0.1
25
0(0
.03
52
)0
.12
43
(0.0
354
)0
.12
35
(0.0
355
)0
.12
52
(0.0
355
)0
.12
50
(0.0
355
)
Fath
erso
me
coll
ege
0.1
859
(0.0
441)
0.1
868
(0.0
442)
0.1
828
(0.0
443)
0.1
731
(0.0
443)
0.1
726
(0.0
443)
Fath
erco
lleg
egra
du
ate
0.2
837
(0.0
460)
0.2
838
(0.0
462)
0.2
790
(0.0
462)
0.2
646
(0.0
462)
0.2
641
(0.0
463)
Gra
nd
pare
nt
coll
ege
gra
du
ate
0.0
918
(0.0
353)
0.0
892
(0.0
356)
0.0
899
(0.0
356)
0.0
848
(0.0
356)
0.0
843
(0.0
355)
Ho
use
ho
ldn
etw
ort
h(1
0,0
00
s)0
.00
52
(0.0
016
)0
.00
50
(0.0
016
)0
.00
49
(0.0
016
)0
.00
47
(0.0
016
)0
.00
46
(0.0
016
)
Ho
use
ho
ldn
etw
ort
hsq
ua
red
(10
,00
0s)
0.0
00
0(0
.00
00
)0
.00
00
(0.0
000
)0
.00
00
(0.0
000
)0
.00
00
(0.0
000
)0
.00
00
(0.0
000
)
Ho
use
ho
ldin
com
e(1
0,0
00
s)0
.01
14
(0.0
091
)0
.01
26
(0.0
092
)0
.01
30
(0.0
092
)0
.01
21
(0.0
092
)0
.01
20
(0.0
092
)
Ho
use
ho
ldin
com
esq
ua
red
(10
,000
s)�
0.0
00
5(0
.00
04
)�
0.0
005
(0.0
004
)�
0.0
005
(0.0
004
)�
0.0
005
(0.0
004
)�
0.0
005
(0.0
004
)
Pri
va
tesc
ho
ol
�0
.02
79
(0.0
419
)�
0.0
286
(0.0
419
)�
0.0
273
(0.0
419
)�
0.0
297
(0.0
420
)
Oth
erla
ngu
ag
esp
ok
ena
th
om
e0
.01
00
(0.0
450
)0
.00
59
(0.0
450
)0
.00
49
(0.0
450
)
Qu
iet
pla
ceto
stu
dy
inh
ou
seh
old
0.1
253
(0.0
390
)0
.12
47
(0.0
389
)0
.11
97
(0.0
393
)
Yo
uth
att
end
sex
tra
cla
sses
0.1
516
(0.0
264
)0
.15
17
(0.0
264
)
Dic
tio
na
ryp
rese
nt
inh
ou
seh
old
0.0
540
(0.0
545
)
Ho
me
com
pu
ter
by
ag
e1
70
.21
63
(0.0
329
)0
.21
53
(0.0
330
)0
.20
94
(0.0
331
)0
.20
60
(0.0
330
)0
.20
31
(0.0
331
)
Rel
igio
nd
um
mie
sN
oY
esY
esY
esY
es
R2
0.1
993
0.1
985
0.2
004
0.2
080
0.2
08
0
Mea
no
fd
epen
den
tv
ari
ab
le2
.81
98
2.8
252
2.8
268
2.8
272
2.8
27
8
Sa
mp
lesi
ze4
,06
74
,00
84
,00
13
,97
83
,97
5
Note:
See
no
tes
toT
ab
le6
.
FAIRLIE, BELTRAN & DAS: HOME COMPUTERS AND EDUCATIONAL OUTCOMES 785
home environment, and whether the youthattends extra classes. Although some of thesevariables have large effects on GPAs, the coef-ficient estimate on home computer is not sen-sitive to their inclusion. Specification 5 reportsthe results of our falsification test using thepresence of a dictionary at home. The coeffi-cient is relatively small and statistically insig-nificant and essentially has no effect on thehome computer estimate.
These estimates provide further evidencethat is consistent with the hypothesis thathome computers have a positive effect on edu-cational outcomes. They are also consistentwith earlier estimates from the 1988 NationalEducational Longitudinal Survey (NELS).Attewell and Battle (1999) find that test scoresand grades are positively related to home com-puter use even after controlling for differencesin individual, parental, and family character-istics. Similar to the NLSY97, the NELSallows them to control for several typicallyunobservable characteristics of the educa-tional environment in the household.21 Theseresults also suggest that home computers mayaffect school performance instead of onlyaffecting the likelihood that a child is enrolledand finishes high school.
F. School Suspension
Personal computers may provide utilityfrom games, e-mail, chat rooms, download-ing music, and other noneducation uses. Al-though these types of activities may providea distraction for children as noted above inthe theoretical model, they might reduce delin-quency and criminal activities among children,thus increasing the likelihood of graduatingfrom high school. The NLSY97 includesdetailed information on delinquency andcriminal activities. We first present resultsfor the relationship between home computersand school suspension. Probit estimates forthe probability of being suspended from schoolin the survey year are reported in Table 8.Access to a home computer is measured in
the year before the school suspension measure.In our sample, 11.3% of children in any givenyear experience a suspension from school.
Having a home computer is associated witha lower probability of school suspension. Themarginal effects estimate is large, negative,and statistically significant. Children whohave access to a home computer are 2.8 per-centage points less likely to be suspended fromschool than are children who do not have ahome computer. The estimated effect is notsensitive to the inclusion of the additional con-trols. Even after including detailed home envi-ronment controls and whether the child takesextra classes, the marginal effects estimate onhome computer remains large, negative, andstatistically significant and similar to the esti-mate in the base specification. The marginaleffects estimate is also not sensitive to the inclu-sion of the presence of a home dictionary. Thepresence of a dictionary at home is not associ-ated with being suspended from school with orwithout controlling for home computers.
The time series variation in this variableallows us to estimate two additional modelsthat may help identify causal effects. First,we estimate a fixed effects regression that con-trols for all unobserved individual, parental,and family characteristics that do not changeover time and time-varying characteristics suchas family structure and income. Estimates arereported in Specification 1 of Table 9. Thehome computer effect is now identified fromchanges over time in access to home computersand school suspension. The marginal effectsestimate on home computer is smaller in mag-nitude and now statistically insignificant atconventional levels, but remains somewhatlarge. The point estimate implies an effect of�.0090, which is 8% of the mean school suspen-sion probability of .1147. The lack of statisticalsignificance of this estimate, however, may bedue to the relatively short time span and lackof time series variation in having a home com-puter. We have at most 4 yr of data for eachchild while they are in school with 40% of chil-dren having 3 yr or less of data. Less than 20%of children experience a change in home com-puters from 1 yr to the next. Although our sam-ple does not represent an ideal application fora fixed effects model, it is somewhat reassuringthat the point estimates from these models donot contradict our previous results.
As a final check of the validity of our resultsfor school suspension, we follow Schmitt and
21. They include measures of the frequency of child-parent discussions of school-related matters, parents’familiarity with the parents of their child’s friends, atten-dance in ‘‘cultural’’ classes outside of school, whether thechild visits science or history museums with the parent,and an index of the educational atmosphere of the home(e.g., presence of books, encyclopedias, newspapers, andplace to study).
ECONOMIC INQUIRY786
TABLE
8
Pro
bit
Reg
ress
ion
sfo
rS
cho
ol
Su
spen
sio
n(M
arg
ina
lE
ffec
ts),
NL
SY
97
Explanatory
Variables
Specification
(1)
(2)
(3)
(4)
(5)
Fem
ale
�0
.07
53
(0.0
054
)�
0.0
753
(0.0
054
)�
0.0
768
(0.0
054
)�
0.0
758
(0.0
054
)�
0.0
758
(0.0
054
)
Bla
ck0
.02
20
(0.0
078
)0
.02
15
(0.0
080
)0
.02
14
(0.0
081
)0
.02
09
(0.0
081
)0
.02
06
(0.0
081
)
La
tin
o�
0.0
26
4(0
.00
74
)�
0.0
279
(0.0
075
)�
0.0
282
(0.0
089
)�
0.0
289
(0.0
089
)�
0.0
293
(0.0
088
)
Asi
an
�0
.02
31
(0.0
199
)�
0.0
195
(0.0
206
)�
0.0
198
(0.0
208
)�
0.0
185
(0.0
212
)�
0.0
187
(0.0
211
)
Imm
igra
nt
�0
.02
14
(0.0
101
)�
0.0
190
(0.0
103
)�
0.0
183
(0.0
105
)�
0.0
174
(0.0
106
)�
0.0
176
(0.0
106
)
Liv
esw
ith
mo
ma
nd
step
da
d0
.05
12
(0.0
111
)0
.05
16
(0.0
113
)0
.05
05
(0.0
113
)0
.05
17
(0.0
114
)0
.05
13
(0.0
114
)
Liv
esw
ith
da
da
nd
step
mo
m0
.06
69
(0.0
237
)0
.06
81
(0.0
241
)0
.06
92
(0.0
243
)0
.07
14
(0.0
245
)0
.07
07
(0.0
245
)
Liv
esw
ith
mo
mo
nly
0.0
41
2(0
.00
83
)0
.03
97
(0.0
083
)0
.03
93
(0.0
083
)0
.04
00
(0.0
083
)0
.04
02
(0.0
083
)
Liv
esw
ith
da
do
nly
0.1
05
7(0
.02
41
)0
.10
46
(0.0
243
)0
.10
55
(0.0
245
)0
.10
45
(0.0
246
)0
.10
42
(0.0
246
)
Liv
esw
ith
gu
ard
ian
0.0
84
5(0
.01
80
)0
.08
12
(0.0
179
)0
.07
88
(0.0
179
)0
.07
78
(0.0
180
)0
.07
72
(0.0
179
)
Mo
mw
as
teen
ager
at
firs
tb
irth
0.0
23
9(0
.00
76
)0
.02
51
(0.0
077
)0
.02
51
(0.0
077
)0
.02
51
(0.0
077
)0
.02
48
(0.0
077
)
Mo
ther
hig
hsc
ho
ol
gra
du
ate
�0
.00
90
(0.0
070
)�
0.0
089
(0.0
070
)�
0.0
088
(0.0
070
)�
0.0
088
(0.0
071
)�
0.0
089
(0.0
071
)
Mo
ther
som
eco
lleg
e�
0.0
20
0(0
.00
79
)�
0.0
193
(0.0
078
)�
0.0
193
(0.0
079
)�
0.0
190
(0.0
080
)�
0.0
192
(0.0
080
)
Mo
ther
coll
ege
gra
du
ate
�0
.03
60
(0.0
089
)�
0.0
348
(0.0
090
)�
0.0
344
(0.0
091
)�
0.0
340
(0.0
091
)�
0.0
341
(0.0
091
)
Fa
ther
hig
hsc
ho
ol
gra
du
ate
�0
.02
53
(0.0
070
)�
0.0
255
(0.0
070
)�
0.0
261
(0.0
070
)�
0.0
259
(0.0
071
)�
0.0
258
(0.0
071
)
Fath
erso
me
coll
ege
�0
.01
89
(0.0
090
)�
0.0
207
(0.0
089
)�
0.0
209
(0.0
090
)�
0.0
198
(0.0
091
)�
0.0
197
(0.0
091
)
Fath
erco
lleg
egra
du
ate
�0
.03
85
(0.0
088
)�
0.0
400
(0.0
087
)�
0.0
403
(0.0
088
)�
0.0
397
(0.0
089
)�
0.0
396
(0.0
089
)
Gra
nd
pare
nt
coll
ege
gra
du
ate
�0
.01
49
(0.0
080
)�
0.0
184
(0.0
077
)�
0.0
188
(0.0
077
)�
0.0
193
(0.0
077
)�
0.0
193
(0.0
077
)
Ho
use
ho
ldn
etw
ort
h(1
0,0
00
s)�
0.0
00
5(0
.00
04
)�
0.0
005
(0.0
004
)�
0.0
005
(0.0
004
)�
0.0
005
(0.0
004
)�
0.0
005
(0.0
004
)
Ho
use
ho
ldn
etw
ort
hsq
ua
red
(10
,00
0s)
0.0
00
0(0
.00
00
)0
.00
00
(0.0
000
)0
.00
00
(0.0
000
)0
.00
00
(0.0
000
)0
.00
00
(0.0
000
)
Ho
use
ho
ldin
com
e(1
0,0
00
s)�
0.0
00
1(0
.00
22
)�
0.0
006
(0.0
022
)�
0.0
007
(0.0
022
)�
0.0
006
(0.0
022
)�
0.0
005
(0.0
022
)
Ho
use
ho
ldin
com
esq
ua
red
(10
,000
s)0
.00
00
(0.0
001
)0
.00
00
(0.0
001
)0
.00
00
(0.0
001
)0
.00
00
(0.0
001
)0
.00
00
(0.0
001
)
Pri
va
tesc
ho
ol
0.0
258
(0.0
105
)0
.02
60
(0.0
106
)0
.02
64
(0.0
107
)0
.02
64
(0.0
107
)
Oth
erla
ngu
ag
esp
ok
ena
th
om
e�
0.0
003
(0.0
103
)0
.00
19
(0.0
104
)0
.00
20
(0.0
104
)
Qu
iet
pla
ceto
stu
dy
inh
ou
seh
old
�0
.01
21
(0.0
091
)�
0.0
124
(0.0
091
)�
0.0
116
(0.0
092
)
Yo
uth
att
end
sex
tra
cla
sses
�0
.00
60
(0.0
061
)�
0.0
059
(0.0
061
)
Dic
tio
na
ryp
rese
nt
inh
ou
seh
old
�0
.00
71
(0.0
116
)
Ho
me
com
pu
ter
�0
.02
79
(0.0
058
)�
0.0
272
(0.0
058
)�
0.0
272
(0.0
058
)�
0.0
274
(0.0
058
)�
0.0
272
(0.0
059
)
Rel
igio
nd
um
mie
sN
oY
esY
esY
esY
es
Pse
ud
o-R
20
.10
69
0.1
091
0.1
095
0.1
091
0.1
09
5
Mea
no
fd
epen
den
tv
ari
ab
le0
.11
32
0.1
126
0.1
131
0.1
129
0.1
13
0
Sa
mp
lesi
ze1
7,3
26
17
,081
16
,926
16
,806
16
,79
4
Notes:
See
no
tes
toT
ab
le6.A
ge
du
mm
yvari
ab
les
are
als
oin
clu
ded
inall
spec
ifica
tio
ns.
Ro
bu
stst
an
dard
erro
rsth
at
all
ow
for
corr
elate
dre
sid
uals
over
tim
ea
rein
pa
ren
thes
es.
FAIRLIE, BELTRAN & DAS: HOME COMPUTERS AND EDUCATIONAL OUTCOMES 787
Wadsworth (2006) and estimate a regressionthat includes future computer ownership inaddition to previous computer ownership.22
Future computer ownership may serve asa proxy for unobserved characteristics thatare correlated with having a home computerand educational outcomes, but cannot havea causal effect on current school suspension.Thus, the finding of a negative coefficient esti-mate on future computer ownership of similarmagnitude to the coefficient estimate on pre-vious computer ownership suggests that thecorrelation in unobserved factors may be theunderlying cause of the estimated negative rela-tionship. Specifications 2–5 of Table 9 reportprobit estimates for the probability of schoolsuspension. The marginal effects estimate onhome computer remains large, negative, andstatistically significant, whereas the estimateon future home computer is much smaller andstatistically insignificant in three of fourspecifications.23 Previous computer ownership,not future computer ownership, appears tohave a strong negative correlation with theprobability of school suspension, which is con-sistent with the hypothesis that home computershave a positive effect on educational outcomes.These findings for the relationships betweenhome computers, future home computers,and school suspension are also consistent withSchmitt and Wadsworth’s (2006) findings forthe effects of home computers on British schoolexaminations. They find statistically insignifi-cant estimates on future computer ownership,whereas the estimate on past computer owner-ship generally remains positive and statisticallysignificant in their regression models.
G. Criminal Activities
If home computers reduce criminal activi-ties then they may have an indirect effect oneducational outcomes. We investigate thishypothesis by estimating separate probitregressions for the probability of committingany criminal activity, being arrested, and gangactivity. Estimates are reported in Table 10for the main specification, a specification thatincludes the presence of a dictionary at home,a fixed effects model, and a specification that
TABLE
9A
dd
itio
na
lR
egre
ssio
ns
for
Sch
oo
lS
usp
ensi
on
(Ma
rgin
al
Eff
ects
),N
LS
Y9
7
Explanatory
Variables
Specification
(1)
(2)
(3)
(4)
(5)
Ho
me
com
pu
ter
�0
.00
90
(0.0
075
)�
0.0
39
8(0
.00
85
)�
0.0
38
4(0
.00
86
)�
0.0
384
(0.0
086
)�
0.0
384
(0.0
086
)
Fu
ture
ho
me
com
pu
ter
�0
.01
61
(0.0
077
)�
0.0
15
1(0
.00
77
)�
0.0
149
(0.0
078
)�
0.0
155
(0.0
077
)
Ma
inco
ntr
ols
Tim
eV
ary
ing
Yes
Yes
Yes
Yes
Rel
igio
n/p
rivate
sch
oo
lN
oN
oY
esY
esY
es
Ho
me
env
iro
nm
ent
No
No
No
Yes
Yes
Ex
tra
cla
sses
No
No
No
No
Yes
Fix
edef
fect
sY
esN
oN
oN
oN
o
R2/p
seu
do
-R2
0.0
099
0.1
064
0.1
083
0.1
08
80
.10
81
Mea
no
fd
epen
den
tv
ari
ab
le0
.11
47
0.1
263
0.1
255
0.1
25
50
.12
52
Sa
mp
lesi
ze1
7,7
51
13
,432
13
,238
13
,22
11
3,1
27
Note:
See
no
tes
toT
ab
le6
.A
ge
du
mm
yv
ari
ab
les
are
als
oin
clu
ded
ina
llsp
ecifi
cati
on
s.R
ob
ust
sta
nd
ard
erro
rsth
at
all
ow
for
corr
ela
ted
resi
du
als
ov
erti
me
are
inp
are
nth
eses
.
22. Future computer ownership is measured in the 2yrs after school suspension is measured.
23. The marginal effects estimate on future computerownership is small and statistically insignificant whenincluded alone.
ECONOMIC INQUIRY788
includes future home computers. We first dis-cuss the results for children committing anycriminal activity, which includes damagingproperty, stealing, other property crimes,assaults, and selling drugs. The reported mar-ginal effects estimates for home computers aregenerally negative but are not statistically sig-nificant at conventional levels. Most of thepoint estimates imply large effects, roughlyequal to about 5% of the mean. The marginaleffects estimate on the presence of a dictionaryis negative, but has a large standard error, andthe estimate on future home computers is pos-itive, but statistically insignificant.
Table 10 also reports estimates for regres-sions for the probability of arrests. The mar-ginal effects estimates are large, negative, andstatistically significant in most of the specifica-tions. The fixed effects estimate is not signifi-cant at conventional levels and is smaller thanthe other estimates but implies a large effect.The range of reported point estimates indicates
that home computers are associated with a de-crease in the probability of being arrested by.0080 to .0179. The average arrest probabilityin the sample is .06. The presence of a dictionaryat home and future computer ownershipappear to have no relationship with arrests.
The marginal effects estimate on homecomputers in the regressions for the probabil-ity of being in a gang are large and negative inall specifications. None of the estimates, how-ever, is statistically significant at conventionallevels. The marginal effects estimate on futurehome computers is very small, but the mar-ginal effects estimate on the presence of a dic-tionary is negative and large, although notstatistically significant.
Overall, the estimates provide some evi-dence of a negative relationship between homecomputers and criminal activities. The mostconsistent and statistically significant resultsare for arrests. For the other criminal activitymeasures, many of the estimates are large and
TABLE 10
Regressions for Criminal Activity (Marginal Effects), NLSY97
Explanatory Variables
Specification
(1) (2) (3) (4)
Any criminal activity
Home computer �0.0120 (0.0088) �0.0113 (0.0088) 0.0001 (0.0090) �0.0074 (0.0132)
Dictionary present in household �0.0154 (0.0197)
Future home computer 0.0078 (0.0137)
Fixed effects No No Yes No
R2/pseudo-R2 0.0405 0.0394 0.0116 0.0355
Mean of dependent variable 0.2449 0.2448 0.2342 0.2641
Sample size 18,192 18,178 21,909 13,355
Arrests
Home computer �0.0179 (0.0041) �0.0176 (0.0042) �0.0080 (0.0055) �0.0146 (0.0055)
Dictionary present in household �0.0036 (0.0079)
Future home computer 0.0023 (0.0055)
Fixed effects No No Yes No
R2/pseudo-R2 0.0758 0.0748 0.0000 0.0840
Mean of dependent variable 0.0597 0.0595 0.0604 0.0597
Sample size 18,178 18,164 21,895 13,300
Gang activity
Home computer �0.0020 (0.0013) �0.0019 (0.0012) �0.0022 (0.0031) �0.0028 (0.0021)
Dictionary present in household �0.0028 (0.0024)
Future home computer 0.0002 (0.0016)
Fixed effects No No Yes No
R2/pseudo-R2 0.1246 0.1239 0.0000 0.1227
Mean of dependent variable 0.0211 0.0209 0.0200 0.0237
Sample size 18,240 18,226 21,966 13,380
Note: See notes to Table 6. Age dummy variables are also included in all specifications. Robust standard errors thatallow for correlated residuals over time are in parentheses.
FAIRLIE, BELTRAN & DAS: HOME COMPUTERS AND EDUCATIONAL OUTCOMES 789
negative and consistent across specifications,but are not statistically significant.
VI. CONCLUSIONS
The personal computer is ubiquitous in theclassroom; however, one-quarter of all chil-dren in the United States do not have accessto a home computer. Although many childrendo not have a computer at home, surprisinglylittle previous research has examined the edu-cational consequences of this disparity inaccess to technology. Using the two majorrecent U.S. panel data sets with informationon computer ownership, the matched CPSand the NLSY97, we employ several empiricalstrategies to examine the causal effects ofhome computers on high school graduationand other educational outcomes. First, mar-ginal effects estimates from probit regressionsfor the probability of high school graduationindicate that home computers are associatedwith a 6–8 percentage point higher probabilityof graduating from high school even after con-trolling for numerous individual, parental,family, and home environment characteristics(including several proxies for educationalmotivation using the NLSY97). Althoughwe find no statistical evidence indicating thatthe probit estimates are biased, we also esti-mate bivariate probit and 2SLS models forthe joint probability of computer ownershipand high school graduation to further ruleout the effects of unobserved factors. Usingparental use of the Internet at work and thepresence of another teenager in the householdas instruments, we find marginal effect esti-mates that are similar to the original probitestimates, although statistically insignificant.Third, estimates from falsification tests usingcable television and the presence of dictionar-ies at home provide some evidence that ourresults are not being driven by unobservables.
Estimates from the NLSY97 also indicatea strong positive relationship between homecomputers and grades and a strong negativerelationship between home computers andschool suspension. Fixed effects estimates,which control for individual, parental, andfamily unobservable characteristics that donot change over time, are smaller in magnitudeand insignificant but continue to imply non-trivial effects. We also find that future com-puter ownership does not have a strongnegative correlation with school suspension,
whereas previous computer ownership contin-ues to have a strong negative correlation.Finally, we find some evidence suggesting thathome computers may decrease crime. This evi-dence may be used to suggest a possible mech-anism by which home computers can increasehigh school graduation rates: by reducingnonproductive activities, such as truancyand crime, among children.
The general consistency of the sign and mag-nitude of estimates across data sets, inclusion ofdifferent sets of controls, timing of computerpurchases, exclusion restrictions, and estima-tion strategies suggest that home computersmay have positive effects on educational out-comes. The main weakness of the analysis isthat some of the techniques, such as the bivar-iate probits, 2SLS, and fixed effects models,produced imprecisely measured estimates. Onthe other hand, the probit models, falsificationtests, and future home computer results pro-vide more precise estimates that are consistentwith the hypothesis that home computersimprove educational outcomes. More evidence,possibly from large random experiments, how-ever, is needed on whether access to home com-puters improves educational outcomes andidentifying the underlying mechanisms.
The findings presented here have importantpolicy implications. They suggest that dispar-ities in access to technology may translate intofuture disparities in educational, labor mar-ket, and other economic outcomes, thus mak-ing the low rates of access to home computersamong disadvantaged minorities and poorchildren especially alarming. Policies thataddress the financial, informational, and tech-nical constraints limiting the optimal level ofinvestment in personal computers among dis-advantaged families may be needed. One solu-tion is to expand the relatively new programsthat provide students with laptop computersto allow students to take computers homeon a regular basis. Tax breaks or special loansfor educational computer purchases, trainingprograms, and computer donations representa few additional examples. The findings alsoraise concerns about funding cuts for technol-ogy-related programs affecting disadvantagedgroups, such as community technology centers(Servon 2002; Servon and Nelson 2001).Finally, home computers in the educationalprocess may become more important overtime as schools are increasingly digitizing con-tent and there is growing momentum for the
ECONOMIC INQUIRY790
controversial issue of replacing textbooks withCD-ROMs or Internet-based materials.
APPENDIX
REFERENCES
Attewell, P., and J. Battle. ‘‘Home Computers and SchoolPerformance.’’ The Information Society, 15, 1999,1–10.
Bureau of Labor Statistics. ‘‘State and Regional Unem-ployment, 2001 Annual Averages.’’ 2002. http://www.bls.gov/lau/staadata.text. Accessed April 7, 2009.
Center for Human Resource Research. NLSY97 Users’Guide. Columbus, OH: The Ohio State University,2003.
Chuang, H.-L. ‘‘High School Youths’ Dropout and Re-Enrollment Behavior.’’ Economics of EducationReview, 16, 1997, 171–86.
Crandall, R. W. ‘‘Bridging the Digital Divide: UniversalService, Equal Access, and the Digital Divide.’’ Paperpresented at Bridging the Digital Divide: CaliforniaPublic Affairs Forum, Stanford University, 2000.
Cuban, L. Oversold and Underused: Computers in theClassroom. Cambridge, MA: Harvard UniversityPress, 2001.
Fairlie, R. W. ‘‘Race and the Digital Divide.’’ The Berke-ley Electronic Journals. Contributions to EconomicAnalysis & Policy, 3, 2004, 1–38, Article 15.
———. ‘‘The Effects of Home Computers on SchoolEnrollment.’’ Economics of Education Review, 24,2005, 533–47.
Freeman, R. B. ‘‘The Labour Market in the New Informa-tion Economy.’’ National Bureau of EconomicResearch Working Paper No. 9254, 2002.
Fuchs, T., and L. Woessmann. ‘‘Computers and StudentLearning: Bivariate and Multivariate Evidence onthe Availability and Use of Computers at Home andat School.’’ CESIFO Working Paper No. 1321, 2004.
Giacquinta, J., J. Bauer, and J. Levin. Beyond Technol-ogy’s Promise: An Examination of Children’s Educa-tional Computing at Home. New York: CambridgeUniversity Press, 1993.
Goldfarb, A., and J. Prince. ‘‘Internet Adoption andUsage Patterns Are Different: Implications for theDigital Divide.’’ Information Economics and Policy,20(1), 2008, 2–15.
Goolsbee, A., and P. J. Klenow. ‘‘Evidence on Learningand Network Externalities in the Diffusion ofHome Computers.’’ Journal of Law and Economics,45(Pt. 1), 2002, 317–44.
Hoffman, D., and T. Novak. ‘‘Information Access: Bridg-ing the Racial Divide on the Internet.’’ Science, 280,1998, 390–91.
Keefe, D., P. Farag, and A. Zucker. Annotated Bibliogra-phy of Ubiquitous Computing Evaluations, UbiquitousComputing Evaluation Consortium. Arlington, VA:SRI International, 2003.
Kirkpatrick, H., and L. Cuban. ‘‘Computers Make KidsSmarter–Right?’’ Technos Quarterly for Educationand Technology, 7, 1998, 2.
Lenhart, A., M. Simon, and M. Graziano. The Internetand Education: Findings from the Pew Internet &American Life Project. Washington, DC: Pew Inter-net & American Life Project, 2001.
Lowther, D. L., S. M. Ross, and G. R. Morrison. Evalu-ation of a Laptop Program: Successes andRecommen-dations. Paper presented at the National EducationalComputing Conference,‘‘Building on the Future,’’Chicago, IL, 2001.
McFadden, D. 1974. ‘‘Conditional Logit Analysis of Qual-itative Choice Behavior,’’ in Frontiers in Econometrics,
TABLE A1
Sample Means of Selected Explanatory Variables in theCPS and NLSY97
Variable CPS NLSY97
Female 0.4964 0.5075
Black 0.1399 0.1455
Latino 0.1329 0.1214
Asian 0.0247
Immigrant 0.0764 0.0504
Household net worth (10,000s) 12.8000
Household income (10,000s) 4.2162
Family income: missing 0.1332 0.2292
Family income: $15,000 to $30,000 0.1261
Family income: $30,000 to $50,000 0.1856
Family income: $50,000 to $75,000 0.1777
Family income: greater than $75,000 0.2427
Home ownership 0.7860
Mother—high school graduate 0.3209 0.3275
Mother—some college 0.2751 0.2264
Mother—college graduate 0.2300 0.2235
Father—high school graduate 0.2205 0.3195
Father—some college 0.1912 0.1521
Father—college graduate 0.2511 0.2256
Grandparent college graduate 0.1820
Lives with mom and step dad 0.1083
Lives with dad and step mom 0.0264
Lives with mom only 0.2303
Lives with dad only 0.0354
Lives with guardian 0.0433
Mom was teenager at first birth 0.1892
Private school 0.1095
Other language spoken at home 0.1171
Quiet place to study in household 0.9101
Youth attends extra classes 0.2918
Dictionary present in household 0.9631
Home computer 0.8043 0.8807
Father uses the internet at work 0.2669
Mother uses the internet at work 0.2557
Another teenager present in household 0.45141
Notes: For the CPS, the sample consists of teenagersaged 16–18 who have completed 11th or 12th grade, buthave not received a high school diploma in the first surveyyear. All variables are measured in the first survey year.For the NLSY97, the sample consists of teenagers livingwith their parents in 1997. Home computer access is mea-sured between ages 15 and 17, fixed characteristics (e.g.,gender and race) are measured in 1997, and possiblytime-varying characteristics (e.g., household income) aremeasured at age 15. Sample weights provided by theCPS and NLSY97 were used to calculate these estimates.
FAIRLIE, BELTRAN & DAS: HOME COMPUTERS AND EDUCATIONAL OUTCOMES 791
edited by P. Zarembka. New York: Academic Press,105–42.
Mitchell Institute. One-to-one Laptops in a High SchoolEnvironment, Piscataquis Community High SchoolStudy Final Report. Portland, ME: Great MaineSchools Project, 2004.
Noll, Roger G., D. Older-Aguilar, G. L. Rosston, andR. R. Ross. ‘‘The Digital Divide: Definitions, Mea-surement, and Policy Issues.’’ Paper presented atBridging the Digital Divide: California PublicAffairs Forum, Stanford University, 2000.
Ono, H., and M. Zavodny. ‘‘Digital Inequality: A FiveCountry Comparison Using Microdata.’’ Social Sci-ence Research, 36, 2007, 1135–55.
Peck, C., L. Cuban, and H. Kirkpatrick. ‘‘Techno-Promoter Dreams, Student Realities.’’ Phi DeltaKappan. 83, 2002, 472–80.
Puma, M. J., D. D. Chaplin, and A. D. Pape. E-Rate andthe Digital Divide: A Preliminary Analysis fromthe Integrated Studies of Educational Technology.Washington, DC: Urban Institute, 2000.
Rockman, et al. AMore Complex Picture: Laptop Use andImpact in the Context of Changing Home and SchoolAccess. A Project for Anytime Anywhere Learningby Microsoft Corporation. San Francisco: Rockmanet al., 2000.
Schmitt, J., and J. Wadsworth. ‘‘Is There an Impact ofHousehold Computer Ownership on Children’sEducational Attainment in Britain?’’ Economics ofEducation Review, 25, 2006, 659–73.
Selwyn, N. ‘‘The Effect of Using a Home Computer onStudents’ Educational Use of IT.’’ Computers andEducation, 31, 1998, 211–27.
Servon, L. Bridging the Digital Divide: Community, Tech-nology and Policy. Oxford, UK: Blackwell, 2002.
Servon, L. J., and Marla K. Nelson. ‘‘Community technol-ogy centers: Narrowing the digital divide in low-income, urban communities.’’ Journal of UrbanAffairs, 23, 2001, 279–90.
Silvernail, D. L., and D. M. M. Lane. The Impact ofMaine’s One-to-One Laptop Program on MiddleSchool Teachers and Students, Phase One SummaryEvidence, Research Report #1. Gorham, ME: MaineEducation Policy Research Institute, University ofSouthern Maine Office, 2004.
SRI International. ‘‘The Integrated Studies of EducationalTechnology: Professional Development and Teachers’
Use of Technology.’’ SRI International Report,Menlo Park, CA, 2002.
Stevenson, K. R. ‘‘Evaluation Report—Year 3, MiddleSchool Laptop Program, Beaufort County School Dis-trict.’’ Columbia: University of South Carolina,1999.
Stoll, C. Silicon Snake Oil: Second Thoughts on the Infor-mation Highway. New York: Doubleday, 1995.
Underwood, J., Billingham, M., and Underwood, G.‘‘Predicting Computer Literacy: How Do the Tech-nological Experiences of Schoolchildren PredictTheir Computer Based Problem-Solving Skills?’’Journal of Information Technology for Teacher Edu-cation, 3, 1994, 115–25.
Universal Services Administration Company. ‘‘AnnualReport.’’ 2005. http://www.universalservice.org/_res/documents/about/pdf/usac-annual-report-2005.pdf.Accessed April 7, 2009.
Urban-Lurain, M., and Y. Zhao. Freedom to Learn Eval-uation Report: 2003 Project Implementation Execu-tive Summary. Prepared for Michigan VirtualUniversity by The Center for Teaching and Technol-ogy, Michigan State University, 2004.
U.S. Department of Commerce. A Nation Online: HowAmericans are Expanding Their Use of the Internet.Washington, DC: U.S.G.P.O, 2002.
U.S. Department of Education. Dropout Rates in theUnited States: 1999. Washington, DC: NationalCenter for Education Statistics, 2001.
U.S. Department of Education. Digest of Education Sta-tistics, 2001. NCES 2002–130. Washington, DC:National Center for Education Statistics, 2002.
U.S.DepartmentofEducation.OfficeofEducationTechnol-ogy. ‘‘Toward a New Golden Age in American Educa-tion: How the Internet, the Law, and Today’s StudentsAre Revolutionizing Expectations.’’ Washington,DC: Office of Education Technology, 2004. http://nationaledtechplan.org/default.asp. Accessed April 7,2009.
U.S. Department of Education. ‘‘Internet Access in U.S.Public Schools and Classrooms: 1994–2003.’’ NCES2005-015. Washington, DC: National Center forEducation Statistics, 2005. http://nces.ed.gov/pubs2005/2005015.pdf.
Zavodny, M. ‘‘Does Watching Television Rot YourMind? Estimates of the Effect on Test Scores.’’ Eco-nomics of Education Review, 25, 2006, 565–73.
ECONOMIC INQUIRY792