Estimating the social value of higher education: willingness to pay for community and technical colleges

Heckman, Lochner and Todd (2006) scrutinize this research based on the Mincer (1974) equation and estimate more general, nonparametric earnings models that allow for earnings to vary by year after completion (nonlinearity) and allow for the nonstationarity of earnings over time. Their analysis shows (1) assuming linearity leads to a downward bias to the return, (2) taking into account taxes has little impact on the return estimates, (3) taking into account tuition costs of schooling lowers the return to college by a few percentage points, and (4) psychic costs, in addition to money costs, can be a barrier to college education. Their work emphasizes that the private returns to education are substantial.

Wolfe and Haveman (2002) identify and describe intrafamily productivity, marital choice efficiency, health of children, crime reduction, charitable giving, and social cohesion as schooling outcomes that are part of nonmarket private returns and social returns.

See also Turner, Tamura, Mulholland and Baier (2007) and Yamarik (2008).

Two professionally moderated focus groups consisting of Kentuckians were conducted to ensure that respondents’ understanding and interpretation of the survey questions matched the intention of the survey authors. One group consisted of eight members of the Donovan Scholar Program, who are individuals over age 65 who were attending selected classes at the University of Kentucky. The second focus group consisted of eight returning students who were attending the Maysville Community and Technical College. Focus groups were recorded and the results were used to refine elements of the survey. The complete survey instrument is available on line at http://cber.uky.edu/pdf/CBER_UL_KCTCSReport_10-2007.pdf.

Knowledge Networks invited 370 members of its web panel to participate in the web-based sample. Two hundred and seventy-five responded yielding a response rate of 74%. The mail-based sample consisted of an initial mailing of 10,000 households. Eight hundred and four were undeliverable. A total of 2681 surveys were returned for a response rate of 29% (2681/9196). Not all 2956 web and mail observations are usable due to: a wording error on two versions of the survey (1486), protestors who did not vote for the referendum and indicated in a follow-up question “my household should not have to pay more taxes to fund the expansion” (261), and item nonresponse for variables in the logit regression (186). The number of remaining usable observations from the web (109) and mail (914) surveys is 1023.

Another indication, and one that might tell something about unobservable characteristics, is that when we control for whether an observation comes from the high response web survey or the lower response mail survey, the coefficient on the dummy variable for the web survey is not statistically different from zero. This result will be reported in Table 3 below for the logit analysis of the contingent valuation referendum responses.

Kling et al. (2012) assess the state of contingent valuation with emphasis on related research during the last 20 years. In addition to certainty statements, they report on successful avoidance of hypothetical bias by making the contingent valuation consequential. In other words, if respondents believe their responses will influence policy, then they report what they would really do. We do not include a consequentiality script. However, we do include the statement that responses will help administrators make decisions that reflect the views of the people of Kentucky in the second section of the survey instrument about budget choices, and we do ask for a vote in a referendum format.

For example, Blumenschein et al. (2008, Figure 2, p. 127) show a plot of price against percentage buyers for a field experiment in which a health management program was offered to individuals for real. Similar “demand curves” are shown for contingent valuation of the same good for both all “yes” responses and for calibrated “yes” responses. The calibration is that only “definitely sure yes” responses are classified as true “yes” responses. The demand curve for all “yes” responses is noticeably (and statistically) higher than the real demand curve. The hypothetical demand curve based on certainty-calibrated “yes” responses is virtually and statistically indistinguishable from the real demand curve. In other words, any hypothetical bias is not detectable after the calibration.

Iterative valuation techniques tend to offer more precise estimates of willingness to pay but the precision comes at a cost. The iterations alter the incentives of respondents to reveal their willingness to pay. In addition, the initial round of valuation in iterative settings may contain unintended information that consequently alters an individual’s valuation; see Whitehead (2002).

Epstein (2003) evaluates the case for using contingent valuation and notes that, in general, possible negative values should not be ignored. Although some households may place a low or zero value on higher education, there was no indication of negative values in the focus groups.

Two additional variables are used to control for version of the survey. Based on a split sample study design some respondents were presented with a referendum and tax amount to prevent either a 10 or 25% reduction in the KCTCS and were also given a “cheap talk” exhortation to avoid hypothetical bias; see Cummings and Taylor (1999). Because we focus on the 10% expansion and use the follow-up certainty questions to mitigate hypothetical bias and we want to control for any combined reduction, cheap talk effect, we include the two variables for reduction/exhortation. Because of a wording error on the survey we do not have parts of our sample that permit clean tests for the effects of cheap talk or reductions separately, but we control for their combined effects. See footnote 17 for a discussion of the implications for sample size.

The results reported above are based on the pooled sample that includes responses from the web and mail surveys. We stratified and estimated logits of the yes/no responses for the web and mail subsamples. Differences across the two are not significant at the 5% level.

In addition to asking about willingness to pay for expansion of the KCTCS, we asked respondents about perceived benefits they receive from education. We asked respondents to allocate points to the various benefit categories. Respondents were told that allocating more points to a given category indicated that they believed education provided more benefit in the given category. Allocating no points to a given category indicated that they believed education produced no benefits to the given category. If responses are grouped, individual, private benefits in the form of “wages of attendees” and “health of attendees” are at least about 24% of the total. Spillover productivity benefits in the form of “economic development,” “technology,” and “wages of non-attendees” are about 39%. If “crime” and “better public decision-making” and “health of non-attendees” are added to spillover productivity benefits, they are about 68% of the total. Despite a separate category for “local purchases,” respondents may be considering the local impact of a nearby community college rather than the local spillover benefits from enhanced human capital. They may be thinking about the cash inflow from state-provided payrolls and expenditures and the impact on local sales. See Siegfried, Sanderson and McHenry (2007) for an exemplary discussion that makes a clear distinction between distributional impacts and efficiency spillovers associated with colleges and universities.

In Appendix Table A1 we report logit results that include two variables that combine the points allocated to quality of life (Crime, Better Public Decision Making, and Health of Non-Attendees) and productivity growth (Economic Development, Technology, and Wages of Non-Attendees). We also explored variables for the effect of a KCTCS campus being located in the county of residence, population density of the county of residence, and years the respondent has lived in Kentucky. None of these variables were statistically significant at conventional levels. The coefficient on Tax Amount, the key variable for estimating mean WTP, is influenced little by their inclusion. A set of dummy variables for regions in Kentucky was included in preliminary regressions, but they were jointly statistically insignificant and were dropped with little effect on remaining variables.

If the sample is restricted to only respondents who were asked about a 10% expansion, the two control variables for cheap talk and reductions combined can be eliminated. This greatly reduces the sample size from 1023 to 526 and slightly reduces the estimate of mean WTP from $55.84 to $51.67. If the means from ACS 2007 are used where available instead of the means from our sample in evaluating the logit, the estimate of mean WTP is increased slightly from $55.84 to $57.92 [43.05, 72.79]. The nonparametric point estimates of mean WTP are substantially higher. The Turnbull estimate is $72.66 with a 90% confidence interval of [62.02, 83.31] which overlaps the confidence interval for the parametric estimate [41.75, 69.92]. The Kriström estimate of $94.95 [86.28, 103.61] does not overlap. Although we believe our sample is representative overall, we have less confidence that it is representative for the cells for each of the eight tax amounts. The parametric estimates control for differences in income, age, education and other observable characteristics and are our preferred estimates.

In this study, household and family are used interchangeably for simplicity.

In order to be consistent with our estimates of total social value, all dollar amounts have been converted to 2007 dollars using the CPI-U.

We exclude individuals without a high school diploma from our analysis to ensure that the Census data are as comparable as possible with the administrative data from KCTCS.

To be consistent with previous literature, we express our log coefficients in terms of percentages. However, the precise interpretation of a coefficient b in percentage terms is (e^b–1), where e is the exponential function. For comparison, a log coefficient of 0.4 is approximately 49% and a log coefficient of 0.2 is around 22%.

According to our administrative data from the KCTCS, more than half of the highest degrees awarded are certificates and diplomas.

For more detail on the data and estimation, see Jepsen et al. (2014).

Our estimates of the value of a 10% expansion of the KCTCS system, however, are based on the distribution of ages when degrees, diplomas, and certificates are actually earned.

The lifetime earnings estimates in Table 6 are based on the estimated values of earnings from equations (2) and (3). These estimated values are based on the coefficients for age (and age squared), highest degree, and the constant term, and all these coefficients differ between the Census and KCTCS data. Differences in the coefficients for age and the constant term explain why the estimated lifetime earnings returns to an associate’s degree are higher in KCTCS data than in the Census data even though the coefficients for associate’s degree are lower in the KCTCS data than in the Census data.

We assume that an expansion of 10% would increase output by 10% because we do not have a strong argument for an alternative. Some programs may have excess capacity and could expand without more funds. Others, particularly the fast growing health fields, are restricted due to current funding for faculty and labs. Moreover, expansion of programs could induce some current students to switch to new programs rather than attracting more students. Switching would lead us to overestimate the gain. However, to the extent that the expansion leads to better matches with students and jobs, then there will be greater productivity that will offset some of the overestimate.

Review of the focus group tapes confirmed that participants understood that the increase and payment were one-time.

For an associate’s degree, the estimated costs are $8003 in direct costs of tuition, fees and books, and 1 year of foregone earnings, the average earnings of a high school graduate the year prior to degree receipt. The average of earnings foregone is $19,950. We assume that the costs for a diploma are 75% of the costs for an associate’s degree, and the costs for a certificate are 50% of the costs for an associate’s degree.

The estimated private return for the Census data contains no controls for occupation. Because a worker’s occupation varies with education level, we also estimate the private returns with Census data that include controls for occupation, and find that the private returns fall from $53.4 million to $42.0 million. This finding suggests that part of the private return of an associate’s degree operates through changes in occupation. The KCTCS administrative data do not contain occupation information.

The NBER website http://users.nber.org/∼taxsim/marginal-tax-rates/reports estimates of tax rates based on the TAXSIM model. Estimates of average marginal tax rates on income for Federal and Kentucky taxes combined are approximately 27% for 2007.

For our best estimate of before-tax private financial value of $53.4 million and a tax rate of 27%, the additional tax revenue is $14.4 million. Because of the additional revenue, other taxes could be reduced at the same level of expenditure or additional expenditures could be made without increasing taxes and excess burden could be reduced. Hines (2008) review of the excess burden of taxes suggests the loss could be as high as 75%. Boardman, Greenberg, Vining and Weimer (2011) suggest that a rate of 23% is probably appropriate for income taxes. If we use a marginal excess burden rate of 30%, then the benefit of the additional $14.4 million revenue implies a reduction in excess burden of $4.3 million. Under the assumption that respondents did not consider this subtle benefit, the total social value increases to $97.0 million.

All estimates shown in Table 8 are positive, but negative values are possible with other combinations of assumptions. For example, if private non-market value is the same size as private financial value and the 90% lower bound on total social value is used, the education externality is negative $8.7 million. We consider this combination unlikely.

We also estimate a simple model to explore whether there is an area-wide education externality. The model is broadly similar to those found in Rauch (1993), Acemoglu and Angrist (2000), and Moretti (2004a), as well as the reviews by Moretti (2004b) and Lange and Topel (2006). However, no attempt is made to account for sorting. Focusing specifically on the associate’s degree offered by KCTCS, a 1 percentage point increase in the percentage of individuals in an area with at least an associate’s degree is associated with a 0.7% increase in earnings. Sorting has not been addressed, but this result hints that part of the private returns reported in Table 4 that reports earning equations using Census data is actually an education spillover. Results are shown in Appendix B.

Assumptions made about estimates of total social value matter too. If the lower bound of the 90% confidence interval of willingness to pay ($69.3 million) is used, then net social benefits are –$34.1 million with a benefit cost ratio of about 0.7.

McMahon (2007) uses a dynamic model of endogenous growth to estimate education externalities that are direct effects as well as externalities that are indirect effects that play out over time in growth and development. We estimate the direct effects and do not attempt to include any indirect effects.