Reviewed by
Joakim
Hussénius and Peter Ericson
Swedish Ministry of Finance, Economic Affairs Department, S-103 33
Stockholm, Sweden.
The present book originated as an Occasional Paper describing the microsimulation model POLIMOD, a tax and benefit model developed by the Cambridge Microsimulation Unit. Principally, the book is a user's guide to POLIMOD, but is also of great interest to readers developing their own models and practitioners curious about what is inside the "black box" of a typical microsimulation.
Static microsimulation is a technique in which agents are exposed to certain experiments. The agents in this context are usually observed individuals or households. With extensive information about the characteristics of these agents (age, sex, income, wealth etc) and the current legislation regarding tax and benefits, it is possible to simulate the budgetary and distributional effects in different policy regimes. For example, the model should be able to show how an increase in child benefits affects the income distribution, as well as the budget. The strength of micromodels, as opposed to macromodels, is the possibility of getting detailed information about the distributional effects of policy changes. All reforms can be evaluated in terms of "winners" and "losers", and changes in measurements like the Gini Coefficient can easily be calculated.
In static microsimulation only the immediate effects are analysed. Behavioural responses to changes in the legislation, and other second order effects are disregarded. However, there is much work and experiment in progress to incorporate dynamic effects in many of the existing microsimulation models. The main topics on the research agenda are how to explain the ways in which labour supply and fertility respond to different tax and transfer schedules. The current demographic situation - low fertility and an aging population - worries many political decision makers.
Despite (or maybe as a result of) their static nature, microsimulation models are used in many different countries; by the government, fiscal authorities and institutes which are responsible for evaluating different policy reforms. But how is it that these models get so much attention? Without any behavioural response, they must surely produce the wrong results! Furthermore, the purpose of many reforms is precisely to change people's behaviour. The trend towards a reduction in marginal taxes is often motivated by a desire to enhance the incentives to work. In a static model this effect is totally ignored and lower tax rates will only result in a reduction of the tax revenue. The answer is, that despite their shortcomings, these models are considered to generate information of great value. Not all reforms are assumed to generate important changes in behaviour and the fact that both macroscopic and distributional effects can be measured is very much appreciated by policy makers.
Prior to the use of simulation models, evaluations of budgetary and distributional effects were very simple and had low accuracy. Often the calculations were based on uncertain estimates of averages and totals. The first order effects produced by the static microsimulation models are on the other hand very precise. Another advantage is that the results are politically indisputable. It is far more likely that experts and politicians will have different opinions about the direction and size of the second order effects. If the modeller would like to incorporate a labour supply response to changes in income taxes, it is possible to find research reporting wage elasticities from small negative values to values just over one. Even though the mean value of all estimates is a small positive value, it is easy to get support for different opinions somewhere in the literature. Therefore, the static models will continue to be an important tool in the future, even though we might see an increased number of dynamic microsimulation models.
At first sight, constructing a microsimulation model seems to be fairly simple. Only two things are required; microdata and a program describing all tax and benefit rules. In practice, however, construction is a very laborious process. First, all the data must be gathered. It is rare that all the information needed can be found in one place, so data have to be gathered from different sources. Before this, it is often necessary to negotiate with the Data Inspection Board to get a permission. Even if several sources can be used, it may be necessary to augment the information with questionnaires. Problems with non-response and representativeness must also be handled. The data should then be validated against external sources. Next, the legislation (at a rather detailed level of description) must be transfered to source code in a suitable programming language. One of the advantages of the model should be the possibility of investigating odd rules and the interaction between different rules. These programs are often very large, and in some cases it takes several years to detect programming errors.
Since developing a microsimulation is a long process, the overall transparency may be low, despite apparent theoretical simplicity. Because of the richness in details, the modeller is forced to consider the pros and cons of particular technical solutions and solve problems as they arise. The solutions chosen are not unique and might have been totally different with another team.
One might think this should not be a problem. All one needs to do is take a look at the program documentation which describes the model. The problem is that good documentation almost never exists. It might be possible to find a list of variables in the database, some comments in the programming code or a few reports describing simulation results. It turns out that, despite the principle that a static microsimulation model should rely on simplicity and transparency, we are always in danger of returning to a black box.
With this in mind, the book by Redmond, Sutherland and Wilson (hereafter RSW), is a rare exception to the rule. RSW not only describe the different problems encountered when building a static microsimulation model, they also present very accurate documentation of the model they have created - POLIMOD. The title of the book might suggest a general discussion about microsimulation, but in fact RSW emphasise the description of their own model.
POLIMOD utilises the microdata from the 1991 Family Expenditure Survey (FES), which is described in an appendix to the book. The appendix also describes how the databases have been updated and reweighted to represent 1996 and 1997. The book consists of four parts:
1. Tax-Benefit Models: Methods and Analysis: An overall exposition of the capacities of the model and how it can be run.
2. Model Outputs and Modelling Assumptions: This section presents a nice example of what a microsimulation model can accomplish. The tax and social security system in 1996/97 is replaced by the system in force during 1978/79. The population is grouped into equivalent household net income deciles, and the different systems are compared in terms of "winners", "losers" and revenue effects. One result is that policy reforms since 1978/79 have mainly been of benefit to high income households through lower income taxes. By contrast, households in the lower deciles have suffered from large increases in indirect taxes.
RSW also discuss how sensitive the results are to different assumptions. For example, the effect of adjusting the households' disposable income in different ways. According to economic theory and common practice, the disposable income is adjusted by a factor depending on the family composition.
3. Model Construction: Data and Methods: This section describes how VAT, National Insurance Contributions and Local Taxes are calculated under different assumptions. It also presents a method for incorporating entitlement to retirement pension and other non-means-tested social security benefits.
4. Model Validation: Validation is an important step when constructing a simulation model. This is true both in terms of external credibility and the researchers own ability to trust the results. In this section both aggregate figures and distributions from the model are validated against external sources.
In general, the book is very well written and provides good documentation to users of POLIMOD. It also has a more general value for people involved in the nitty gritty of building static microsimulation models. For readers interested in a more general survey of static microsimulation, however, the book is too detailed and not comprehensive enough.
Return to Contents
of this issue
© Copyright Journal of Artificial Societies and Social Simulation, 2000