Administrative costs in higher education: how fast are they really growing?

Abstract

It is widely believed that administrative expenditures in US higher education are growing too rapidly, particularly in relation to expenditures that are directly related to instruction, and that this so‐called ‘administrative bloat’ is a major factor in the rising cost of higher education. We argue that this perception of rapid growth is exaggerated, and that it results from focusing on simple expenditure aggregates that obscure important variation across institutions. A more careful analysis using panel‐data methods supports a more benign conclusion that administrative expenditures, and their ratio to instructional expenditures, are stationary over time. This conclusion is supported by panel unit‐root tests. This suggests that some of the concern about the role of administrative expenditures in rising higher‐education costs may be misdirected.

Notes

1. ‘Current funds expenditures and transfers’ are defined as ‘the costs incurred for goods and services used in the conduct of the institution’s operations.’ This category includes expenditures for instruction, research, public services, academic and institutional support, student services, operation and maintenance of plant, auxiliary enterprises, hospitals, and independent operations. For complete definitions of all expenditure variables used in this paper, see the IPEDS Glossary (http://nces.ed.gov/ipeds/glossary/). Beginning in 2001, the Government Accounting Standards Board (GASB) required all public institutions to issue financial reports using the reporting model and standards of GASB Statements 34 and 35. Our sample ends with the year 2000 because subsequent data are not directly comparable.

2. FTE is the more commonly used measure of enrollment. Use of headcount data may introduce a bias to the extent that, for example, an institution with a high ratio of commuters to resident students may have a high ratio of headcount to FTE and a lower ratio of administrative costs per student. Our approach mitigates this bias while presuming that the nature of the institution, and hence its ratio of headcount to FTE, remains fairly stable over time.

3. A simple linear trend regression yields the result that the difference between the growth rates of administrative and instructional expenditures per FTE is growing by 0.035 percentage points per year. The full equation is [%ΔAdmin_per_FTE − %ΔInstruct_per_FTE] = −0.010 + 0.035t, where Admin_per_FTE and Instruct_per_FTE are administrative and instructional expenditures per FTE, respectively, and t is time (in years). Both the intercept and the trend coefficient are significant at the 1% level.

4. For example, suppose one large institution spends $1000/FTE on administration in year one, and $1100/FTE in year two. Meanwhile, nine other institutions each spend $100/FTE on administration in both years. The aggregate data suggest an increase from $2000 to $2100, or an increase of 5%. But the average institution experienced a growth rate of 1% (i.e., zero for nine of the institutions, and 10% for the large influential outlier).

5. Results are available from the authors on request. The tests are those developed by Emerson and Kao (2001, 2006), most of which are extensions of Vogelsang (1997). For administrative and institutional expenditures per FTE, the tests are significant at the 1% critical level. For the instruction–administration ratio there is not a significant structural break at the 10% critical level.

6. The dramatic differences between the OLS and fixed‐effects results can be explained in terms of an omitted‐variable specification error. If there are institution‐specific fixed effects α _i but these are omitted from the model (say, by constraining the intercept to be the same for all institutions), then the error term in the pooled OLS regression is ϵ _it + α _i − α. If the α _i values are positively correlated with the right‐hand side variable y _{i, t−1}, then the OLS estimate of β will be biased upward.

7. The panel unit root tests were performed using EViews 5.1 and the respective standard settings with regard to lag length (BIC) and bandwidth selection (Newey–West using the Bartlett kernel). For a description of these tests, see Quantitative Micro Software (2005, pp. 530–541). The Levin–Lin–Chu and Breitung tests impose a common β across institutions in computing the test statistic, while the Im–Pesaran–Shin and Fisher tests estimate a separate β _i for each institution and construct a test statistic by combining information from individual unit root tests on the separate β _i values. The differences among the results of the tests may be a result of the differing assumptions that they make. First, these tests are, strictly speaking, valid only asymptotically as both the number of cross‐sectional units (N) and the number of time periods (T) approach infinity. The tests differ in their assumptions regarding the relative rates at which N and T grow, and their small‐sample distributions in situations such as this (in which T is only 17) can differ substantially from their asymptotic distributions, leading to differences in conclusions. In addition, these tests rely critically on the assumption of independence across cross‐sectional units, which may be violated if their expenditure series are correlated. These tests may be more appropriate for ‘narrow, long’ panels (small N, large T) such as cross‐country macro data than in ‘wide short’ (large N, small T) micro panels such as this one. For more details on these tests, see Breitung (2000), Im et al. (2003), Levin et al. (2002); and for the Fisher tests, see Maddala and Wu (1999). Baltagi (2005) also has a good discussion, and Wagner (2006) provides a critique. Unit root tests for the 1984–1990 and 1991–2000 subsamples are available from the authors on request.