Microsimulation and Public Policy

Order this book

Reviewed by
Cathal O'Donoghue
Microsimulation Unit, Department of Applied Economics, University of Cambridge, CB3 9DE.

Microsimulation and Public Policy is a volume of selected papers from the International Association for Research into Income and Wealth's special conference on microsimulation, held in Canberra in 1993. Like Conte et al. (1997), this book is a good representation of the state of the field of microsimulation in the early nineties, consisting of 21 papers from all areas of the subject and an introduction by the editor.

For readers unfamiliar with the field, microsimulation is an approach which spans a number of social sciences, focusing primarily on the economic behaviour of agents and investigating the impact of public policy at the micro-level. The models typically take representative samples of micro agents and measure the effect of government policy on these samples. The field originates from a paper by Guy Orcutt, "A New Type of Socio-economic System" (1957). He was concerned that macroeconomic models of the time had little to say about the impact of government policy on things like the income distribution of agents (households and firms) in the economy. He suggested the development of simulation models using micro-agents for policy use.

The seventies were an era of large scale microsimulation development in the USA with the construction of the DYNASIM model (Orcutt and Glazer 1980). Interest also spread to Germany in the eighties. The resulting models tended to be large scale efforts requiring enormous computing, data and manpower resources. Also during this period, governments and academics started to develop smaller scale applications which could use the more limited data and computing power generally available. The Harding volume is the latest a long line of multi-author volumes including Haveman and Hollenbeck (1980), Orcutt et al. (1986), Brunner and Petersen (1990), Lewis and Michel (1990), Citro and Hanushek (1991) and Hancock and Sutherland (1992). Other recent survey articles in the field include, Merz (1991), Mot (1992), Harding (1993) and Sutherland (1995).

In her introduction, Harding surveys the main types of microsimulation model which currently exist. Microsimulation has tended to focus on the household sector although a number of researchers have also simulated the behaviour of firms at the micro-level. Concentrating on household microsimulation, Harding discusses the two main types of model, static and dynamic, together with more general issues which concern microsimulation such as data requirements.

The book contains chapters which represent each of these strands and is divided into five parts dealing with static models, behavioural response, lifetime and retirement incomes, micro-macro links (including firm behaviour) and finally relations between models and data. In this review, I recategorise the papers into four sections: static models, statistical dynamic models, behavioural response and validation/reliability issues, focusing on papers which look at the household sector. (This review does not cover the three chapters in the volume which use firms as the unit of analysis.) As the book contains a broad representation of the state of the field of microsimulation in the early nineties, I use shall also use this review as an opportunity to survey the field briefly.

Static Models

The first set of models described are static in the sense that they look at economic agents at one point in time. Models of this kind are simply accounting mechanisms and do not consider behaviour over time. They tend to be based on representative samples of a population and are primarily used to investigate the first round impact of government policy reform. Simulation of this type is used to disentangle the complexity of government policy rather than simulating behaviour. For example, in many countries, individuals will pay social insurance contributions on income, social insurance benefits will be based on social insurance contributions and income taxes will be based on income net of social insurance contributions and benefits. Social assistance benefits may then be based on incomes net of these three instruments and finally passported benefits, such as health benefits or free school meals, may be dependent on the receipt of social assistance. These government policies thus involve a complex web of interactions and changes to one policy may have unexpected impacts on the others.

Early static models were used to examine the impact of government policies (and their changes) on particular family types. For example, at budget time, newspapers commonly include calculations of this type, indicating how certain "typical" families gain or lose from policy changes. However, the population of households in a country is far more heterogeneous than a small number of hypothetical households. In addition, the interaction of government policies with the distribution of households adds a further layer of complexity. It is for this reason that static models were developed. Chapters 2 (Eason) and 3 (Gupta and Kapur) describe the uses of static microsimulation in the UK and Canada, two countries whose governments have highly developed programs of this kind. Both papers describe standard approaches. Each country takes a micro-data set (either a sample survey or samples taken from the administrative statistics of the micro units being examined: individuals, households or firms) and applies tax and social policy rules to measure the impact of government policy. By altering appropriate parameters, analysts can assess the immediate effect (before behavioural second round impacts) of a change in government policy. Simulated effects of policy reforms include changes in total cost, impact on the income distribution and numbers or micro-unit types (such as single pensioners or two earner couples) of gainers and losers. Both papers face similar trade-offs between using data with a high level of detail and lower coverage or selecting data sets with greater coverage, but less detail. For example, both countries have samples of tax returns in great detail, allowing many types of tax deduction to be simulated. However, these data are collected to assess tax and therefore only cover those individuals who have to pay tax. Thus they tend to under represent low income households. By contrast, household surveys cover the whole population, but their data are less detailed and thus only approximations of the government policy rules can be simulated. Both countries make the same compromise, constructing two kinds of tax-benefit model, which make use of the strengths of each type of data.

Predicting the impacts of policy reforms (and changes in the population structure) over time is another object of microsimulation models; forecasting tax yiel ds over the next three years for example. In order to simulate government policies in the future, one needs to know the structure of the population at different times. One method for forecasting with static models is to use ageing techniques. Static ageing involves applying adjustment factors to account for changes in the population structure, inflation, the distribution of income and changes in policy rules. Adjusting for changes in the population structure uses a similar method to that which deals with non-response bias in surveys. Static ageing accounts for changes in the population structure by assigning weights in such a way that the external control totals represent forecasts rather than describing the situation in the year in which the survey was collected. The second and third adjustments, which account for inflation and changes in the distribution of income, use differential 'uprating' factors applied to income by source. Lastly, tax and benefit rules have to be adjusted for planned or forecast changes. Chapters 2 and 3 briefly describe how static ageing techniques are applied in the UK and Canadian government models. In both cases, forecasts of differential wage inflation and changes in population structure are supplied by macro models. In chapter 16, Meagher discusses the linkage between macro forecasting models and static microsimulation models.

In addition to comparing the impact of policies within a country, microsimulation models can also be used to compare the effectiveness of policies between countries with interesting results. The approach taken by De Lathouwer (chapter 4) is to apply the Unemployment Benefit system from the Netherlands to data from Belgium. This technique is one first used by Atkinson et al. (1988) and is useful for comparing the impact of systems in isolation from underlying differences in the population structure. However, this isolation of factors also forms the basis for a criticism of the approach since national tax-benefit systems are obviously created with national characteristics in mind.

Statistical matching is a procedure used to impute missing information required for the input databases of microsimulation models. For example, income surveys do not always contain expenditure information which is needed to simulate indirect taxation such as VAT. This is a problem that Salomaki faces in chapter 5. The Finnish static model 'Tuja' is like those from chapters 2 and 3 and contains a wide range of income labour market and demographic variables. However, the input database which is based on register information does not include data on expenditure or the use of welfare services. The same information is collected in household budget surveys (HBS). Salomaki describes two methods for merging the different data sets. As unique household identifiers do not exist, the data cannot be matched perfectly. Therefore statistical methods are necessary. The two methods which Salomaki considers are average statistical matching and hot rank merging. Both methods rely on the fact that the data sets contain a number of variables in common. Average statistical matching involves categorising households in both data sets using the common variables (Salomaki uses 160 groupings) and then, in the case of expenditure and welfare services data, assigning the average per unit disposable income expenditure (on each of 56 expenditure groups and 43 welfare service categories) for each group in the HBS to the microsimulation input data set. However, this method does not allow for any within group consumption rate variability between households. A second method which can do this is hot rank merging. In this case, households within the groups are ranked according to their income. Expenditure and use of welfare services for a household is assumed to be same as that for the household of the same rank in the HBS. Salomaki found (as expected) that although both methods produced accurate estimates of total expenditures and welfare service use, the latter method, not surprisingly, produced better estimates of the variability. However, these methods were only checked for the whole set of households. A potential criticism, raised by Harding in chapter 1, is that Salomaki did not test for representativity by sub-group.

In order to simulate many taxes or benefits, one year of data is sufficient. However, certain benefits, such as social insurance pensions, require information from a number of years, sometimes as many as forty. This type of simulation therefore requires time series information which is typically unavailable for most countries. For this reason, simulations of this type are either avoided or carried out using approximate methods. Unusually, the Swedish National Social Insurance Board (NSIB) maintains detailed information on incomes and labour market characteristics over time and thus can provide the longitudinal information necessary to model these types of benefits precisely. Eklind et al. (chapter 14) report on a model which uses this data. They present a static model, similar to those described above, with a cross-sectional data set at its core. In addition, they are able to link the longitudinal information required to simulate pensions (provided by the NSIB from 1961 to 1990) using 'exact matching' or unique person identity numbers. In order to complete the working life cycles used, a statistical matching exercise is carried out, while to maintain forecast demographic trends, static ageing is used.

Statistical Dynamic Models

Because of the availability of good time series information, the Swedish modellers were able to simulate policy instruments which depend on life cycle information. Many modellers in other countries do not have such a luxury. In addition, static modelling using time series information does not allow for experiments to be carried out. For example, what would happen to welfare bills if fertility rates or working patterns changed? Dynamic models have been developed to cope with these issues. These use a process called dynamic ageing to project the households forward through time. While static ageing changes the weights of a single cross-section, dynamic ageing simulates transitions at the individual level and thus produce hypothetical sets of panel data. Simulated transitions include changes in demographic characteristics, educational qualifications, labour market patterns and income mobility. Like static microsimulation models, the rules for government policies and use of state services can be applied to each of the waves of the generated panel. Unfortunately, simulating transitions is a very data intensive procedure, requiring extensive time series data sets to specify transition equations. Even where the data does exist, as Harding points out in her introduction, these data sources are generally subject to period, cohort and age effects and may therefore be unsuitable for forecasts. Typically however, this data is not available at all and so dynamic microsimulation is confined to a small number of countries.

Two types of models have been developed: dynamic population models and dynamic cohort models. The former type takes a cross section of the population at one point in time and projects it forward over a number of years, creating new cross-sections at intervals (generally annually). Population models thus produce hypothetical panel data sets which correspond to household panel surveys such as the PSID and GSOEP. Typically, these models have been used to study pensions and the impact of changing demographic and economic patterns on pension and long term care expenditure. Cohort models, on the other hand, generate panel data sets which cover the entire lifetime, but generally for only one cohort. Simulating transitions over the entire life cycle, these models have been used to look at intra-personal redistribution over the life cycle, to produce estimates of lifetime income and to compare these with distributions based on measurements of income over shorter periods such as a year.

In addition to the distinction between cohort and population approaches, models can be classified as using statistical or behavioural methods. We shall concentrate on models of the first type in this section. Statistical models use transitions which are designed to reproduce existing (or expected) mobility within a population. They do not incorporate behavioural decisions based on reactions to government policy.

The book contains five chapters which use statistical dynamic models, chapter 11 (Falkingham and Harding), chapter 12 (Nelissen), chapter 13 (Galler), chapter 15 (Andreassen, Frederikssen and Ljones) and chapter 22 (Caldwell). Falkingham and Harding describe an application comparing the intra-personal and inter-personal redistributional characteristics of social insurance and social assistance based benefit systems using two cohort models for the UK and Australia. Nelissen applies a similar analysis using a dynamic population model to investigate the lifetime redistributive impact of the social security system in the Netherlands. Galler focuses on one particular issue which influences model forecasts, the inter-dependency of incomes within couples. Andreassen et al. describe a dynamic population model (MOSART) which has been built for Norway using a 10% sample of the population to estimate the transition equations. They also compare the impact of a number of demographic scenarios on the future position of the Norwegian public pension system. Caldwell outlines CORSIM, a population model built for the USA. He also discusses methods for validation and alignment of models to macro aggregates.

Falkingham and Harding use two cohort models for the UK and Australia (HARDING and LIFEMOD) to compare the lifetime redistributive impact of the tax-transfer systems in each country. Both models were built using a common framework, each simulating the lifetime transitions of 4000 hypothetical individuals living in a steady state world and using some common code. There are a number of differences in the treatment of the labour market however. Cohort models like these concentrate on a steady state world. In other words, all the lives of individuals in the model are assumed to follow behavioural patterns and transitions for a specific period of time. This has the advantage that tax-benefit systems are analysed in isolation from changes in behaviour and the external economic environment. The lifetime redistributive effect of systems can therefore be assessed. However, by not allowing simulated individuals to interact with changed circumstances, these models have the disadvantage that they are not actually modelling the real periods in which these individuals lead their lives. Changes in factors such as the earnings distribution, improvements in general education levels and alterations to demographic behaviour may in fact have greater impacts on lifetime income distribution than the tax-benefit system.

Caveats aside, to my knowledge, this paper is the first study of its kind, using dynamic models for international comparison. By definition, it is also the first to make comparisons of the lifetime redistributive impact of tax-benefit systems across countries. The authors unsurprisingly found that lifetime incomes in both countries were more equitable than incomes measured over shorter periods. Making comparisons of mean incomes is quite difficult across countries for a number of reasons. These include the choice of exchange rate, the impact of different time periods and the fact that the data on which the transition and income equations are based may not be fully comparable. (In chapter 20, Blackburn and Richter describe how international efforts have not yet been made to ensure the same comparability of micro data across countries which already exists for macro data. They also suggest a number of ways in which comparability could be improved.) Falkingham and Harding found that within country inequality was higher inAustralia than in the UK and that inter-personal redistribution was more important in Australia (with a social assistance based system) than in the UK (with a social insurance based system in which contributions are more closely linked to benefits). Conversely, lifetime intra-personal redistribution was higher in the UK.

As mentioned above, Andreassen, Frederiksen and Ljones describe MOSART, a new dynamic population model built for Norway. Models of this kind project full cross-sections forward through time using dynamic ageing. Like the Eklind et al. paper (chapter 14) the MOSART model is based on register data. From a modelling perspective, this is a tremendous advantage. The modellers have access to a 10% sample of the Norwegian population, containing detailed information on benefit receipts and transitions in educational, demographic and labour market status going back up to 25 years. The existence of this historical database creates similarities to the Eklind et al. paper on Sweden. However, it differs in that future cross-sections are simulated using a dynamic ageing processes. For researchers relying on much smaller survey based data sets, access to such a wealth of data is enviable. However the model does not seem to take full advantage of the information available in the data set. MOSART is a mainly demographic model relying on simple transitions estimated using the data. Labour market transitions are based primarily on demographic information, while demographic transitions contain no feedback from economic or educational states. Clearly much more could be done to disentangle socio-economic influences on a wide range of behaviours.

The authors go on to describe an application for which dynamic models are particularly suited, namely population projections. They apply some baseline assumptions to investigate the structure of the population and the operation of the public pension system into the future. A number of alternative assumptions are tested to investigate the impact on the population structure for differences in the immigration rate and changes in the fertility rate. The impact of these alternatives on the size of public pension expenditures, ensuing contribution rates and intergenerational redistribution are then examined.

Caldwell describes CORSIM, the third of the statistical dynamic models included in the volume. In terms of the level of detail modelled, CORSIM is one of the more comprehensive microsimulations with 900 equations and over 10000 parameters. Unlike the MOSART model, CORSIM does not rely on historical register information about transitions. Instead the model takes a sample from the US census in 1960 and ages it forward dynamically to 1995, the time at which the research was carried out. Projections of the population to 2030 are then carried out in similar way. Although historical transitions are not available, the historical part of the model is controlled to match historical aggregate information. An alignment procedure is used to make sure that model aggregates agree with historical totals such as fertility and mortality rates. This procedure involves two stages. Firstly micro-level transition equations are estimated using panel survey data. The population is then passed through the series of equations from 1960 to the present, with transitions being estimated using Monte-Carlo techniques. However, no matter how detailed the system of transition equations, it is not possible for simulated micro-behaviour to match actual behaviour. In consequence, historical aggregate transition information (or in fact macro forecasting) is then used to constrain aggregate transitions within the model. Alignment procedures are therefore important devices for ensuring that micro forecasts do not deviate too much from external aggregates.

Nelissen (chapter 12) describes an analysis of the lifetime redistributive impact of the social security system in the Netherlands using the NEDYMAS microsimulation model. The objective of the model is to compare the degree of verticalredistribution of social security schemes on lifetime incomes for four cohorts of ten years born between 1930 and 1960. In order to do this, a sample cross-section needs to be available from the time at which the first cohort entered the labour market. However, as such data is not available, a hypothetical cross-section is generated so as to have the same characteristics as the 1947 census. In many ways, NEDYMAS is therefore similar to the cohort type of models discussed in the Falkingham and Harding chapter. Unlike cohort models, this model does not assume a steady state world and instead attempts to model the conditions through which individuals have lived (and will live) their lives. Starting with a historical cross-section and simulating the birth of later generations, the model allows the lifetime redistributive impact of tax-benefit systems to be compared for a number of cohorts. One particular comparison Nelissen examined using the model was between the lifetime redistributive impact of the Atlantic-type flat rate benefits found in the UK and Continental-type earnings related benefits, given the existence of ceilings on earnings related benefits. Means tested benefits were found to have a more equalising effect.

Unlike other research dealing with statistical dynamic models, Galler (chapter 13) discusses a specific problem rather than simply describing a model briefly and sketching its applications. Galler's objective is to look at the impact of introducing a pension reform for widows. Currently, because pension benefits depend on contributions paid as a proportion of earnings, many women with short working careers can find themselves with very low pension entitlements. This paper looks at a modelling strategy to investigate the introduction of a minimum pension and a contributions based pension for survivors rather than the current benefit for dependents. Contributions would now be paid by spouses who remained at home, on the basis of a formula, although the state would pay the contributions for carers. Correlations between earnings profiles for members of a couple are important in the analysis of widow's pensions as the incidence of unpaid work, caring for dependents, will influence government expenditures. Dependencies between earnings profiles depend on joint decisions taken about labour supply within a family. Traditional human capital approaches assume independence between earnings within couples. By contrast, Galler finds that this assumption underestimates the within couple dispersion and he emphasises that such dependencies need to be introduced into microsimulation models. However, estimation procedures also need improvement to produce model estimates using maximum likelihood methods.

Behavioural Response

A major criticism of statistical dynamic and static models is that no attempt is made to model any influence which government policy has on behaviour. These models assume that behaviour remains the same after a policy change. However, because microsimulation models simulate taxes and benefits in great detail, they have a comparative advantage when it comes to their use in modelling behavioural responses to policy decisions. One example of behavioural response is the decision to participate in the labour market as a result of changes in the tax-benefit system. Actual labour supply or number of hours worked may also be influenced. Demand for goods is likely to change as a result of price differences arising from commodity taxes. Finally, changes in the tax-benefit system may influence the decision to apply for benefits to which an individual is entitled.

The Harding volume considers a number of these issues. Baekgaard (chapter 7) assesses the behavioural response in demand for child care in Denmark resulting from price changes. Symons and Warren (chapter 8) investigate the impact of incorporating behavioural responses to reforms in indirect taxation in Australia. Chapter 9 (Merz) investigates methodological issues relating to non-market labour supply. Klevmarken and Olovsson (chapter 10) incorporate behavioural response as part of a dynamic population model for Sweden.

The Danish government has a static model, the Law Model, which operates in a similar manner to those already described for the UK and Canada. Like those models, the Law Model simply looks at the 'day after' effects of policy reforms. In 1991, the Danish government introduced a price reduction for the cost of public day care. The existing model, based on a static approach, took no account of increased demand resulting from this price change. Baekgaard describes a technique designed to cope with this behavioural change. He regresses actual demand for day care (a dummy variable) against charges for child care and interactions with other demographic, labour market and income variables using a logistical regression approach. The Law Model includes a module simulating charges for child care (by municipality) which is used as an input into the behavioural model. This in turn produces feedback about the demand for child care. This feedback can then be used as an input in the microsimulation model to determine the revenue effect of the price reduction.

The second applied behavioural response paper which uses a static model is the analysis of indirect tax reforms in Australia provided by Symons and Warren. Adjusting for altered consumption patterns resulting from price changes (such as changes in indirect taxation) is one of the most common types of behavioural response included in microsimulation models. Modelling behavioural responses to indirect tax reforms requires information about how the demand for a commodity will change in relation to changes in its own price (the own price elasticity) and also about the impact of prices changes for other goods (cross price elasticities). Estimates of these elasticities are not available for Australia and so, in order to estimate behavioural response to indirect tax changes, the authors use a demand system estimated using UK survey data as an input to their indirect tax model STATAX. Although it would be desirable for consumer demand systems to be estimated using own country data, the authors consider that expenditure patterns are similar enough between the countries to make this assumption acceptable. One criticism of this approach is that a similar expenditure pattern may have evolved as a result of a different behavioural mechanism based on different price combinations. The second part of the STATAX program is the determination of prices faced by consumers. Changes to the prices of commodities which are inputs in the production process will also impact on the prices faced by consumers. In order to account for these price changes, a simple input-output framework is used.

Klevmarken and Olovsson describe a dynamic microsimulation model for Sweden, MICROHUS, which uses a small degree of behavioural response to policy reforms. In many respects the model is similar to the MOSART or CORSIM models with transitions between demographic, educational, labour market and income states. However, there is a difference in the specification of the labour market module. In MICROHUS, although transitions into and out of the labour market are determined by state transition probabilities (and not through behavioural equations), actual hours worked are determined by Hausman type labour supply models which take into account taxes and benefits (Burtless and Hausman 1978). Although the general procedure and the inclusion of Hausman type models in microsimulation are not new, this is one of the first instances of a behavioural labour supply model in a dynamic microsimulation. The importance of this is that, because of labour market rigidities and short term constraints to behaviour, behavioural response to policy changes may take time to have an effect. Thus dynamic models may be more appropriate because they incorporate the time dimension.

Merz expands the type of labour supply modelling framework used in Klevmarken and Olovsson to include non-market activities. Because of changed opportunity costs, he believes that tax-benefit changes may lead to changes in non-market activities within a household, in addition to market activities such as paid work in primary and secondary jobs. Merz uses a three stage modelling approach to this time allocation problem, modelling first the participation probability, then the wage and finally the labour supply in hours. Merz applies his methodology to the 1990 reform in Germany (which reduced marginal income taxes) and finds a very slight substitution of labour supply from first jobs to secondary ones. In addition, he estimates a shift from household production to pure leisure activities in the region of 2-3%. In particular, use of a microsimulation model allows the impact of the reform to be assessed for different sub-groups in the population. Lone parents were found to be particularly likely to substitute away from household production to paid market activities.

Validation and Reliability

Although behavioural and dynamic models have been developed over the last twenty years, their use has been quite limited in policy analysis by governments. Many governments simply use estimates of first round effects taken from static models. The reason is that there is concern about the level of uncertainty associated with more sophisticated models. So far there has been little published work on the validity of these models, nor has much work been done to estimate the degree of error associated with microsimulation results.

A number of papers in this volume begin to address these issues. Firstly, Caldwell (chapter 22) describes a number of validation mechanisms which can be applied to dynamic models. These include in-sample validation, which assesses the predictive power of the model in describing the data on which it was estimated. Out-of-sample validation attempts to measure the predictive power of the model in explaining data of a similar type which were not used in the estimation of the model. This is a more realistic measure of the predictive power as the stochastic variation is different. The third type of validation procedure is out-of-type validation. For example, a model may be good at reproducing transitions derived from panel data, but less good at reproducing the structure of a cross-section. The fourth type of validation is multiple-module validation. This procedure is quite interesting. As data availability will not support the joint estimation of all processes, dynamic models tend to estimate marginal processes. Caldwell provides the example of marriage and health insurance processes which are both estimated individually within the model. A multiple module approach would be to test the accuracy of results for married couples covered by health insurance. In other words, this method assesses the validity of a process which is not directly simulated in the model. This is quite a useful typology to organise future work on validation. However, it would have been interesting if some empirical estimates of the validity of CORSIM had been presented to illustrate these techniques.

Baekgaard (chapter 7) estimates ex-post behavioural response to day-care reforms in Denmark and is thus able to assess the predictive ability of his behavioural equations by comparing forecast and actual demand. Although reasonably accurate forecasts result, the specification of the model could not take into account a number of issues. For example, no notice was taken of changes in pricing structure by municipalities. Having been estimated on cross-sectional data, the model was also unable to allow for an underlying trend in the demand for child care.

De Vos and Zaidi (chapter 6) carry out a similar validation exercise, estimating the accuracy of static ageing. They firstly forecast the UK Family Expenditure Survey from 1985 to 1988 using static ageing and then compare results with those actually found in the 1988 FES. The authors' objective is to test whether static ageing could be a useful method for updating poverty estimates. They find that reweighting for demographic changes is not sufficient. Differential increases in income were found to be more important. In other words, changes in the income distribution were more significant overall than changes in the demographic structure. However, once these changes were accounted for, forecasts over the three year period were quite reasonable.

When microsimulation results are reported, they are generally presented using a single estimate and do not typically show the degree of error attached to it. Pudney and Sutherland (chapter 21) attempt to develop confidence intervals for microsimulation results which are able to account for errors associated with sampling variability, parameter estimation and stochastic effects. This paper extends previous work to incorporate the effect of statistical uncertainty introduced by behavioural response, illustrating the analysis with a discrete choice female labour supply model. Sampling is found to be the dominant source of error. Parameter estimation for the behavioural model adds little to the confidence of most estimates. However the confidence interval of particular results (such as the number of women working) increases by as much as 50%. The authors conclude that the additional uncertainty associated with behavioural response equations may be so large as to make them of no practical use when examining policy changes.

Conclusion

To sum up, the Harding volume is a very useful and comprehensive description of the current state of the field of microsimulation. As a volume with broad coverage, it is a useful guide to the type of work taking place in the field. Although this review primarily addresses the simulation side of the papers, the book also contains many new and interesting applications of microsimulation modelling, such as the comparison of lifetime redistribution across countries and cohorts. The book also usefully describes and assesses a number of tools commonly used in microsimulation, such as static ageing and statistical matching. Significant contributions have also been made to methodological development; for example the estimation of the impact of tax changes on non-market labour supply and the inclusion of behavioural response in dynamic models. The volume is also important because it begins to address issues which relate to the credibility of microsimulation. In other words, papers in the book begin to develop validation techniques and produce estimates of confidence intervals associated with model results.

On a more critical level, the book is almost too broad in scope. In this, it is similar in style to other microsimulation volumes and as such it reflects the broad range of users of this type of methodology. Users include government analysts interested in accurate tools which can respond quickly to policy questions, social scientists who want flexible investigative models with which to explore behaviour in complex systems and statisticians or econometricians who are developing new methodological techniques. It might have been useful to have separate volumes focusing on specific issues like the uses of microsimulation in government, validation techniques, modelling behaviour and issues associated with dynamic or cross-country microsimulation.

This book details the widespread international use of microsimulation techniques. However, development in the field has been slow, with little progress being made between the late eighties and the initial development of DYNASIM in the seventies. Caldwell discusses a number of reasons for this lack of progress. In the seventies, the field was very heavily funded by the US government as a result of very optimistic expectations about the prospective benefits. In consequence, lack of subsequent progress has to be measured against relatively rapid progress at that time. The absence of a quick return in terms of powerful and accurate analytical capabilities, combined with the very high development costs, meant that many other countries (with the exception of Germany) were not prepared to provide the level of resources needed to build large scale models. Many countries therefore embarked on less ambitious programs. Rapidly reducing computing costs and improved access to data in the late eighties have seen the field expand again. This period has also seen the development of most of the models described in the Harding volume. Unfortunately, many of these models were developed in isolation and had to learn lessons of model construction independently. Recently, however, there has been a welcome increase in cross-model co-operation. The transfer of technology from established programs to developing programs should further reduce development outlays and thus increase the ratio of benefits to costs. One example is provided by the development of a European static model (EUROMOD) involving a consortium of national microsimulators and other experts. Another involves the transfer of code and expertise from the CORSIM project to new models in Canada and Sweden. The existence of books like this can only be of assistance in encouraging further such co-operation. I would thus recommend Harding's book to anyone with an interest in the field of microsimulation.

References

ATKINSON A. B., F. Bourguignon and P.-A. Chiappori 1988. What Do We Learn About Tax Reform From International Comparisons? France and Britain. European Economic Review, 32: 343-352.

BURTLESS G. and J. Hausman 1978. The Effect of Taxation on Labour Supply: Evaluating the Gary Negative Income Tax Experiment. Journal of Political Economy, 86.

BRUNNER J. K. and H. G. Petersen, editors, 1990. Simulation Models in Tax and Transfer Policy. Campus, Frankfurt.

CITRO C. F. and E. A. Hanushek, editors, 1991. The Uses of Microsimulation Modelling, Volume 2: Technical Papers. National Academy Press, Washington.

CONTE R., R. Hegselmann and P. Terna, editors, 1997. Simulating Social Phenomena. Springer Verlag, Berlin.

HANCOCK R. and H. Sutherland 1992. Microsimulation Models for Public Policy Analysis: New Frontiers. STICERD Occasional Paper, London School of Economics.

HARDING A. 1993. Lifetime Income Distribution and Redistribution: Applications of a Dynamic Cohort Microsimulation Model. North Holland, Amsterdam.

HAVEMAN R. H. and K. Hollenbeck, editors, 1980. Microeconomic Simulation Models for Public Policy Analysis, Volume 2. Academic Press, New York, NY.

LEWIS G. H. and R. C. Michel, editors, 1990. Microsimulation Techniques for Tax and Transfer Analysis. Urban Institute Press, Washington.

MERZ J. 1991. Microsimulation - A Survey of Principles, Developments and Applications. International Journal of Forecasting, 7: 77-104.

MOT E. S. 1992. Survey of Microsimulation Models: Inventory and Recommendations, Social Security Research Committee. UUGA, The Hague.

ORCUTT G. H. 1957. A New Type of Socio-Economic Systems. The Review of Economics and Statistics, 58: 773-797.

ORCUTT G. H. and A. Glazer 1980. Microanalytic Modelling and Simulation. In B. Bergmann, G. Eliasson and G. H. Orcutt, editors, Micro Simulation Models: Methods and Applications. IUI Conference Reports 1980:1, Industrial Institute for Economic and Social Research, Stockholm.

ORCUTT G. H., J. Merz and H. Quinke, editors, 1986. Microanalytic Simulation Models to Support Social and Financial Policy . North Holland, New York, NY.

SUTHERLAND H. 1995. Microsimulation Models in Europe - A Survey. Discussion Paper, Microsimulation Unit, Department of Applied Economics, University of Cambridge.

Return to Contents of this issue

© Copyright Journal of Artificial Societies and Social Simulation, 1998