Toward Improved Modeling of Retirement Income Policies: Interim Report (1995)

In our final report, we will address overall modeling strategies to support the development of retirement-income-related policy models that can adequately inform the public debate. (Such models may be newly developed or may build, at least in part, on existing models.) In addition, we will address data collection strategies, both for data that are needed for behavioral research and for data that are needed as the basis for policy models. (The two are not always the same. For example, microsimulation models often use a broad-gauge population survey as the primary database and draw on other specialized data sets for estimates of behavioral parameters to apply to the primary database records.)

We will also address the key issue of incorporating research findings into policy models, identifying, in our best judgment, the state of the art that we believe policy models should reflect (e.g., variables that policy models should include in order to estimate savings or labor supply responses to policy changes). Finally, we will recommend directions for future analytical research that can improve the base of knowledge that is relevant to important policy questions about retirement income security.

In this interim report, we make a few recommendations pertaining to retirement income policy models and data that we believe are needed to support adequate policy analysis in this area. We also make two recommendations on a topic that we will consider further in the final report—namely, how to coordinate the funding of research and modeling in the area of retirement income security so as to maximize the cost-effectiveness of scarce resource dollars.

Recommendation 1 Retirement-income-related policy models that are operated by government agencies or that are developed with government funds should be made publicly available to the policy and research communities in a timely manner.

Recommendation 2 Retirement-income-related policy models should be adequately documented so that analysts other than the model developers can readily use them.

Recommendation 3 Government agencies should take advantage of the dramatic changes in computing technology and the dramatic reductions in computing costs to develop (or sup-

port the development of) retirement-income-related policy models that are fully accessible to the research community.

Historically, government policy models have rarely been readily available, well documented, or designed for accessibility with user-oriented computing technology. More typically, models have been developed and operated by relatively small staffs within agencies or their contractors with little interaction with the broader research community (see Citro and Hanushek, 1991:Chap. 11). We argue that it is in the best interests of the policy-making process for this situation to change and for agencies to follow these three recommendations.

The first recommendation calls for agencies to establish a policy of ready access to other researchers and analysts for retirement-income-related models (and databases) that are developed with public funds. The policy should apply to in-house models that are developed by an agency itself and also to models that are developed from contracts or grants to other organizations. Of course, private organizations are free to develop and market proprietary models; however, models that receive substantial public support for their development should be available to the policy and research communities at a nominal cost.

One can question the desirability of making government models widely accessible. There are associated costs for developing adequately documented models that users other than the model developers can understand and run. Also, for contentious policy debates, there is the risk that policy makers will be overwhelmed by competing estimates developed by various groups that are based on not-well-understood assumptions or modifications to models and that understanding of the issues will be impeded rather than facilitated.

On balance, however, we believe that the advantages of providing government models to outside analysts outweigh the disadvantages. There is a compelling need for the policy process to be informed by policy models that are as up to date and accurate as possible and for the models' strengths and weaknesses to be thoroughly understood by analysts and decision makers. Yet agencies have limited resources of their own to validate and improve their policy models. By making their models more widely accessible, agencies can only enlarge the community of researchers who are interested in and working with a model. In turn, a larger user community will likely result in more attention to model validation and enhancement and thereby facilitate the development of improved models that provide more useful and accurate information to crucial policy debates. In this regard, we believe that researcher access to models of the Social Security system (including the SSA model) is vital given the importance for all Americans of the policy decisions that will need to be made about the future of the system. Similarly, it is important that models of the employer pension system and of health care financing be accessible.

As a practical matter, a model may be available but not usable if it is not well documented (training may also be needed for effective use of a model by other analysts). The preparation of documentation is often neglected for a variety of reasons, including that the press of the policy debate leaves little time for the documentation task (on this point, see Citro and Hanushek, 1991:Chap. 10, and Hollenbeck, 1991). Ideally, documentation is prepared in tandem with model development and before the model is heavily used to supply estimates to the ongoing policy discussion.

The computing environment for which models are designed also has a great impact on their accessibility. Models that are developed (or redeveloped) for use with microcomputers and workstations, with good graphical user interfaces, are much more accessible to users than models developed for other computing environments. Assuming they are well designed, such models are much easier for users to work with, modify, understand, and evaluate. The costs

and time to modify and apply these models are also much less than for models that use older technology (see, e.g., Cotton and Sadowsky, 1991). In some cases, it may not be cost-effective to develop a model for a microcomputing environment if the model can provide needed information only with some other technology (e.g., a supercomputer). But, in most cases, we believe that for agencies as well as other users, the advantages of implementing models with microcomputing technology will be compelling. We noted earlier that newer models are generally being developed to take advantage of the new computing technology. (In addition to the models listed above, the health care finance reform model of the Agency for Health Care Policy and Research, AHSIM, has recently been redesigned for personal computers.)

Recommendation 4 Relevant agencies should continue to support existing retirement-income-related panel surveys of individuals. Such surveys, which permit analyzing behavioral responses to policy changes over time, are essential for retirement-income-related research that can inform the development of adequate policy models. They must be continued if they are to provide sufficient longitudinal information for analysis purposes.

Recommendation 5 Relevant agencies should develop and implement data collection systems that provide improved information on the nature and extent of employer retirement-income-related benefits and on employer and worker behavior relevant to retirement. The new data should include panel studies of employers and samples of their employees. Researchers and policy analysts who will use these data should be involved from the outset in the design of the data systems.

Recommendation 6 Relevant statistical and administrative agencies should create up-to-date matched files of survey responses and administrative records that, with suitable protections to maintain the confidentiality of individuals or employers, are available for retirement-income-related research and policy analysis use.

Recommendations 4, 5, and 6 identify important priorities for the development of data that are needed for good models of retirement-income-related policy issues. These priorities are the continuation of panel surveys that provide longitudinal information about households; the development of panel surveys that provide longitudinal information about employers and their employees and of improved data generally about the full range of retirement-income-related employer benefits; and the creation of matched survey and administrative records data files.

Panel surveys, which obtain repeated observations of the same individuals over time, are essential to understand behavior that is relevant to future retirement income security (e.g., savings, labor supply, the retirement decision) and what factors influence changes in behavior. In turn, such knowledge is essential to modeling the effects of proposed policy changes.

Panel surveys, by definition, involve repeated interviews and, as such, are vulnerable to funding cuts. The tendency is to question the value of added rounds of data collection. Yet to provide full value for retirement-income-related analysis, panel surveys must be continued over long spans of time so that there is a sufficient period in which to observe such phenomena as patterns of asset accumulation and decumulation as a function of individual characteristics, changes in the economic environment, and policy changes.

In addition, it is important to begin new panels with younger cohorts periodically. Other-

wise, there is no way to observe if there are changes in tastes that cause younger groups to behave differently from older groups as they age. Also, without new cohorts, there is no way to observe the effects of newly initiated policies on behavior (e.g., an increase in the Social Security retirement age can have little effect on the savings or labor supply behavior of people already close to retirement).

We described earlier the important new information that is being collected in HRS and AHEAD. There are questions about the appropriate mix of funds for HRS with regard to the number of interviews for a particular cohort as traded off against the sample size for a particular cohort and the introduction of new cohorts. There are also cross-agency questions about the appropriate balance between funding for HRS and AHEAD and other useful panel surveys (e.g., NLSY). What we want to stress is the importance of funding for panel surveys generally. In particular, we note that HRS cannot fulfill its promise if it is not continued in some form for a longer period.

A major gap in the information that is needed to support analysis of retirement-income-related behavior concerns data on employers. We do not have much information on the full range of retirement-income-related employer benefits, including, for example, retiree health care plans. Also, we do not have much panel data on employers and their workers. As a consequence, we know very little about why some employers offer pensions and other types of benefits and others do not, why employers choose to offer particular types of plans, why they decide to terminate plans, why they decide to try to buy out older workers, and the like. Hence, it is very difficult to estimate the behavior of employers with regard to the retirement-income-related benefits they provide in response to either economic factors or government policy changes. In turn, this lack of understanding limits the ability to develop policy models to project likely retirement income support obtained by workers via employer-provided pensions and other benefits. To model workers' behavior, one must first have estimates of the kinds of benefits that employers will be likely to offer under current and alternative policy regimes.

With regard specifically to pensions, analyses to date of employer plans have been limited to time-series statistics (e.g., trends in type of pension plan coverage) or to detailed information about employer characteristics for very small samples of employers. Only limited evidence has been found that employers behave in ways expected by economic theory—for example, there is little evidence that they fine-tune pension plans to maximize worker productivity or minimize hiring and training costs. (See Gustman and Mitchell, 1992; recent work by Bajtelsmit, 1995, suggests that worker productivity, using an aggregate measure of output per employee, leads employers to offer more generous pensions, rather than the other way around.)

Collecting data from large samples of employers is an expensive proposition that entails problems of response and protection of respondent confidentiality. We believe it is important to initiate new data collection in this area but acknowledge the cost and other constraints. What could make sense is to take a staged approach to determining data needs and cost-effective ways of obtaining needed data. Such an approach could begin with small focus group studies of employer senior financial and personnel executives to determine what they think are important variables that affect their worker hiring, retention, and benefit policies. (Because retirement-income-related benefit programs represent long-term financial commitments by employers, it is important to involve senior financial people in these studies.) From such studies and from discussions with researchers in the field, it could be possible to design

cost-effective panel surveys for analysis purposes. (Because the universe of employers is undergoing continual turnover, with “births ” and “deaths,” the survey design would need to allow new cohorts to be introduced over time.) Ideally, such surveys would follow samples of employers and also follow samples of their employees to determine worker responses to employer initiatives and to obtain measures of individual worker characteristics, including measures of productivity, that could help explain employer benefit plan decisions. To make sure such surveys are as useful and cost-effective as possible, it will be important to involve researchers from both academia and the private sector in their design, similar to the model followed for HRS and AHEAD (see section below on organization of modeling and research).

We described earlier the plans to match Social Security and other administrative records with HRS and AHEAD. We support the development of such matches and of matches of SSA and other administrative records with other surveys, both of individuals and employers. Such matched files are needed, not only for policy models but for analytical models that address important research questions concerning the behavior of individual decision units. The lack of matched files is a major impediment to many kinds of retirement-income-related analysis and policy modeling. The matching task is operationally feasible; what is needed is the will to develop a reasonable policy for safeguarding confidentiality while providing research access.

Recommendation 7 Relevant agencies should explore ways to integrate retirement-income-related research, data collection, and policy modeling, so as to obtain the most cost-effective use of available resources. Such integration should represent a high-priority goal in order to minimize duplication of effort, ensure that important issues are not overlooked, and ensure that priorities are set in light of the full range of policy concerns.

Recommendation 8 Relevant agencies should explore mechanisms for bringing academic researchers and government analysts together on a regular, routine basis to facilitate the development of high-quality, relevant models and associated data for addressing retirement-income-related policy concerns.

We believe that an important requirement for cost-effective and appropriate retirement-income-related policy analysis and modeling is for the federal government to find a way to address retirement income policy questions from a broad and integrated perspective. We described earlier the fragmented structure for such analysis, in which each agency tends to consider its own part of the picture (e.g., SSA looking at Social Security in isolation, PWBA looking at employer pensions in isolation, the Treasury Department looking at revenue effects of retirement-income-related tax laws in isolation). This fragmentation also extends in many instances to data collection—for example, data sources about employer benefits may pertain to pensions (e.g., the Form 5500 database, which provides useful data primarily for qualified pension plans) or to health care (e.g., the National Employer Health Insurance Survey) but not usually to both. The fragmentation of policy analysis and much associated data collection makes it very difficult to develop the needed broad perspective or to allocate increasingly scarce resources for policy analysis in a cost-effective manner.

Of course, there are instances in which it makes perfect sense for particular agencies to develop and apply particular special-purpose policy models and databases—for example, the PBGC needs a model to project corporations' financial status and the implications for the

pension insurance fund, and the U.S. Office of Personnel Management needs a model to project the actuarial balance of the federal civil service retirement funds. But for policy analysis that is concerned with the retirement income security of future generations and what it will cost workers to attain an adequate level of retirement income, there should be provision for an integrated approach. In fact, we believe that it makes little sense for agencies to undertake such tasks as evaluating in detail the capabilities of current policy models or considering ways to develop improved models or data until a mechanism has been put in place with which to approach those tasks from a broad perspective.

Moreover, we believe that such a mechanism should be formal in nature. There are many instances of effective informal interagency cooperation in the area of retirement income policy (e.g., the analysis of the April 1993 CPS pension supplement in Pension and Welfare Benefits Administration [1994] was undertaken by an interagency team, and there is an informal interagency committee to advise NIA on the HRS and AHEAD surveys). Moreover, a formal mechanism will always rely on informal interagency cooperation and teamwork. But we believe that a formal mechanism is needed to ensure that the resources required for an integrated approach to retirement-income-related policy modeling and analysis are used most efficiently and are directed to the highest priorities.

It is important that the integrating mechanism address not only policy modeling per se, including the associated data requirements, but also retirement-income-related behavioral research that is needed for improved policy models. Integration of research support is necessary to make best use of scarce funding dollars and to ensure that effort is directed to the highest priority topics determined after consideration of the full range of policy concerns.

There are many possible mechanisms for coordinating the development and application of retirement-income-related policy models and research in a cost-effective and appropriate manner. One possibility is to assign lead responsibility for developing appropriately broad models and supporting the needed research to a single agency, with additional funding to be contributed by other relevant agencies and oversight provided by an interagency committee. The lead agency could be a program agency, such as SSA or PWBA, or a research agency, such as the National Science Foundation or NIA, or a new institute. It could develop models and conduct research in-house or by means of grants or contracts with private organizations. Another possibility is to have a mechanism within the Executive Office of the President (e.g., within the Office of Science and Technology Policy or the Office of Management and Budget) to review agency budgets for retirement-income-related modeling and research and to assign priorities.¹

The merits of particular mechanisms for integrating policy analysis in the area of retirement income need to be carefully considered. There is also a concern that integration may go too far and thereby stifle initiative. However, without some type of integrating mechanism, we believe there is likely to be inefficient and ineffective use of scarce resources: in some instances, resources will be wasted in duplicative effort; in other instances, resources will be inappropriately used for modeling or analysis that overlooks key interrelationships among policy areas.

It is worthwhile noting in this regard that other countries have seen the need to take a broad and integrated approach to issues of income security for retired people and the associated tax burdens on current workers. For example, after an intensive interagency review of modeling requirements in the area of retirement income, the Canadian government has em-

barked on a multiyear effort to develop a comprehensive Canada Pension Plan Policy Model (LOGISIL, 1994). The work is a joint project of the Office of the Superintendent of Financial Institutions (which includes the chief actuary) and Human Resources Development Canada, with input from other relevant agencies, such as the Finance Department and Statistics Canada. The Australian government has also explored ways to integrate the development of policy models for retirement income and other areas although there has not been a final determination of how best to proceed.

Finally, an important element for cost-effective coordination of retirement-income-related policy modeling, data collection, and research is regular interaction between government agency analysts and the outside research community, including academic researchers and those who work for the private sector (e.g., for benefit consulting companies). Researchers need to be made aware of priority policy concerns and practical considerations that affect the kinds of modeling and analysis that government agencies can do. They also need to learn how government policy models work and what are cost-effective ways to incorporate research results into policy models. In turn, agencies need to be aware of research and modeling advances in the research community that can bring important new insights and approaches to the policy process.

The development of HRS and AHEAD involved close working relationships between government agency staff and academic economists, sociologists, demographers, and survey researchers. The result appears to be two very promising data sets for addressing retirement-income-related policy issues. The data are already attracting considerable research attention and providing policy-relevant findings. This type of close collaboration between government analysts and outside researchers should be encouraged, through technical advisory committees; seminars and similar forums for exchanging information about data, research, and models; and other means. Such close cooperation is needed to ensure that research is directed to key questions of underlying behavior, that appropriate data are collected to support research and modeling, and that appropriate policy models are developed to inform what is a critical debate about the future retirement income security of the nation.