A survey of dynamic microsimulation models: uses, model structure and methodology

More than 10 years ago O’Donoghue (2001a) surveyed the dynamic microsimulation models that had been developed up to that point. However many of the barriers mentioned have been gradually overcome in the last decade. This paper surveys the development and practices in dynamic microsimulation over the past decade, and discusses the methodological challenges today. It provides an overview of the methodological choices made in more than 60 known dynamic microsimulation models and examines the advantages and disadvantages of different practices. In addition, this paper reviews the main progress made in the field and explores how future microsimulation models could evolve potentially.


INTRODUCTION
A dynamic microsimulation model is a model that simulates the behaviour of micro-units over time. Orcutt et al. (1961) described the first dynamic microsimulation model following the inspiration of Orcutt's (1957) article. Most dynamic microsimulation models that have developed in following decades trace a direct or indirect link back to this model. Urban planners use microsimulation techniques to estimate traffic flows, while in the field of economics and social science, microsimulation is often used to analyse the social economic policies. In this paper we shall review how dynamic microsimulation in social science has developed, with main focuses on the economic models developed in the past decade.
Micro level data, such as data obtained from a household survey, is often chosen as the basis for social economic research. In order to evaluate certain impacts of public policies, e.g. the redistributive impact over the course of a lifetime, it is necessary to utilise a long panel dataset. In general, such datasets are not available, either because the analysis relates to the future, as in the case of pension forecasts, or because collected datasets do not cover sufficiently long time periods; therefore, analysts use dynamic microsimulation models to assist in their analysis, a concept which was initially suggested by Orcutt in 1957. Essentially, microsimulation is a tool to generate synthetic micro-unit based data, which can then be used to answer many "what-if" questions that, otherwise, cannot be answered.
Microsimulation models, as in the field of policy modelling, are usually categorised as "static" or "dynamic." Static models, e.g. EUROMOD (Sutherland, 2007), are often used to evaluate the immediate distributional impact upon individuals/households of possible policy changes without reference to the time dimension and extensive behavioural adjustment. Some newer static models, e.g. IZAΨMOD (Peichl et al., 2010), improved the traditional model by incorporating certain behaviour responses assuming the market adjusts to the new steady state overnight.

models in UK, APPSIM in
Australia (Harding, 2007a) and DESTINIE2 (Blanchet et al., 2009) in France etc. Meanwhile, a few generic software programmes have emerged, such as ModGen (Wolfson and Rowe, 1998), UMDBS (Sauerbier, 2002), GENESIS (Edwards, 2004) and LIAM , eliminating the need to create a model from scratch. It has allowed an internationalisation of the models with developments in Belgium (Dekkers and Belloni, 2009), Italy , Canada (Spielauer, 2009), UK (Emmerson et al., 2004) etc. Nevertheless, the decade has seen the demise of a number of models such as DYNACAN in Canada, CORSIM in U.S., NEDYMAS (Dekkers et al., 1993) in the Netherlands, the Belgian model (Joyeux et al., 1996) and MIDAS in New Zealand etc. The micro-econometric and micro-economic understandings of the processes that make up a dynamic microsimulation model have also greatly improved over this period. It is therefore worth in considering the progress made by the discipline over the past decade.
In this paper we shall describe the models developed, their uses and discuss some of the methodological choices faced. We then review the progress made by the discipline since the earliest models and suggest some directions for future development.

OVERVIEW OF MODELS AND THEIR USES
Dynamic microsimulation models can have many uses and this section provides an overview of the principle uses. Table 1 summarises many of the existing dynamic microsimulation models in terms of their main purposes, which covers projection, evaluating/designing public policies, inter-temporal behaviour studies, etc. Given that the most accessible micro datasets for social scientists are household or individual level information, most models do not incorporate information on business establishments, with a few exceptions for models like (Sweden) MOSES (Eliasson, 1977), NEDYMAS model, where business behaviours are incorporated through market equilibriums in the models. There are only a small number of firm-level microsimulation models, e.g. DIECOFIS (Parisi, 2003), and they are mostly static.
Dynamic microsimulation models can also be used to evaluate the future performance of various long-term programmes such as pensions, educational financing, and health and long-term care, by analysing simulated future cross-sectional data. The governmental models such as DYNCAN (Morrison, 2000), POLISIM (McKay, 2003), PENSIM2 (Emmerson, 2004), the Sfb3 models (Galler and Wagner, 1986), MOSART (Andreassen et al., 1996), PENMOD (Shiraishi, 2008) and SESIM (Ericson, and Hussenius, 1999;Klevmarken and Lindgren, 2008) have been extensively used for this purpose. The existence of baseline projections allows the design of a new public policy by simulating the effect of potential reforms. Models such as LIAM ), PRISM (Kennell and Sheils, 1990), the Belgian dynamic model (Joyeaux et al., 1996), the SfB3 population model (Galler and Wagner, 1986), LIFEMOD (Falkingham and Johnson, 1993), SESIM (Flood, 2007;Klevmarken, 2010) and Belgium MIDAS Dekkers and Belloni, 2009), have all been used to look at pension reform. A number of models such as DYNAMOD, the SfB3 cohort model (Hain and Hellberger, 1986), LIFEMOD (Harding, 1993), and SAGE (Zaidi and Scott, 2001) have been used to examine changes to education finance, whereby education costs are to be paid for over an individual's lifetime. Fölster (2001) used a microsimulation model to examine reforms to social insurance utilising personal savings accounts.
With the longitudinal information created from dynamic microsimulation models, researchers can study the inter-temporal processes and behaviours at both the aggregate and individual levels. For example, CORSIM (Keister, 2000), DYNAMOD (Baekgaard, 1998), and the New Zealand MIDAS model (Stroombergen et al., 1995) have all been used to look at wealth accumulation. Creedy and van de Ven (2001), Nelissen (1998) and others have used dynamic microsimulation models to explore lifetime earning redistributions. Models such as DESTINIE1/2, LIAM, LifePaths, and IFSIM have been used to examine intergenerational transfers (Rowe and Wolfson, 2000;Bonnet and Mahieu, 2000;Blanchet et al., 2009;Baroni et al., 2009;O'Donoghue et al., 2009), whilst FAMSIM (Lutz, 1997) has been used to study the demographic behaviour of women, and MICROHUS (Klevmarken and Olovsson, 1996) examined the impact of a taxbenefit system on labour market mobility. Models that simulate these processes can be used to LI, O'DONOGHUE A survey of dynamic microsimulation models: uses, model structure and methodology design policies to combat these problems, for example DYNASIM was used to study the effect of teenage childbearing, while CORSIM has been used to look at dental health within the US population (Brown et al., 1995). The FEM and POHEM models were designed to evaluate the evolution of the population's health status and its budget implications for the US and Canada (Zucchelli et al., 2010;, whist the LifePaths modelling framework has been used in Canada to examine time use issues (Wolfson and Rowe, 1998).
By combining spatial information with dynamic microsimulation models, the model can also be used to predict the geographical trend of certain social economic activities. This type of model is usually referred to as a dynamic spatial microsimulation model, e.g. MOSES (Wu et al., 2008) from U.K., there are a number of models that attempt to analyse policy changes at the national level. For instance, the SVERIGE model simulates a number of demographic processes for policy analysis in Sweden (Vencatasawmy et al., 1999;Holm et al., 2003), whilst the SMILE model (Ballas et al., 2005a;O'Donoghue et al., 2011) analyses the impact of policy change and economic development on rural areas in Ireland. In addition to modelling economic policy, SimBritain (Ballas et al., 2005b) looks at the evolution of health at the national level while models such as SustainCity (Morand et al., 2010) focus on the housing market with a dynamic setting.
However, occasionally the opposite has occurred, where dynamic microsimulation models have been used as input into macro models as in the case of MOSART (Andreassen and Solli, 2000), DYNASIM (Wertheimer et al., 1986;Favreault and Smith, 2004)

APPSIM Australia
Designed to provide answers regarding the future distributional impact of policy change and other issues associated with policy responses to population ageing (Harding, 2007a;Kelly and Percival, 2009) BRALAMMO Brazil Models the Brazilian labour market for pension welfare analysis (Zylberstajn et al., 2011) CAPP_DYN Italy Analyses the long term redistributive effects of social policies (Mazzaferro and Morciano, 2008) CBOLT US Analyses potential reforms to federal entitlement programmes and quantifies the US nation's long-term fiscal challenges (Oharra et al., 2004) CORSIM US Models changes occurring within kinship networks, wealth accumulation, patterns of intergenerational mobility, the progressivity and the life course of the current social security system, as well as potential reforms, household wealth accumulation, health status, interstate migration, time and income allocation, and international collaborations (Caldwell et al., 1996;Caldwell et al., 1997) Czech Republic Model Czech Republic Designed to analyse public pension system and potential reforms in Czech republic (Fialka et al., 2011) DEMOGEN Canada Models distributional and financial impact of proposals to include homemakers in the Canadian pension plan (Wolfson 1989) DESTINIE I/II France Models public pensions and intergenerational transfers (Blanchet et al., 2009;Bonnet and Mahieu, 2000;Bonnet et al.,1999) DYNACAN Canada Models the Canada Pension Plan and its impact on the Canadian population (Morrison, 2000;Osberg and Lethbridge, 1996)

DYNAMITE Italy
Models microeconomic issues and the impact of macroeconomic/institutional changes on the distribution of income (Ando and Nicoletti-Altimari, 1999;Ando et al., 2000) DYNAMOD

I & II Australia
Models life course policies such as superannuation, age, pensions and education, long-term issues within the labour market, health, aged care and housing policy, future characteristics of the population and the projected impact of policy changes (Antcliff 1993;Antcliff et al., 1996;King et al., 1999a;King et al., 1999b)

DYNASIM I & II US
Forecasts the population up to 2030 by employing different assumptions regarding demographic and economic scenarios, and analyses the cost of teenage childbearing to the public sector under alternative policy scenarios, also includes a link to a macro model (Citro and Hanushek, 1991a;Citro and Hanushek, 1991b) DYNASIM III US Designed to analyse the long-term distributional consequences of retirement and ageing issues (Favreault and Smith, 2004)

DYPENSI (SIPEMM) Slovenia
A Slovenia Dynamic Microsimulation Model with the focus on pension system simulation (Majcen, 2011) FAMSIM Austria Models the demographic behaviour of young women (Lutz, 1997) LI, O'DONOGHUE A survey of dynamic microsimulation models: uses, model structure and methodology

FEM US
A demographic and economic simulation model designed to predict the future costs and health status of the elderly and to explore what current trends or future shifts might imply for policy, developed by RAND (Goldman et al., 2006;Goldman et al., 2009) GAMEO France Analyses and assesses the consequences of various higher education policies (Courtioux et al., 2008) HARDING Australia Analysis of lifetime tax-transfer analysis, for analysis of policy concerning the Higher Education Contribution Scheme and redistributive impact of government health outlays over the lifetime of an individual (Harding, 1993) HealthPaths Canada Simulate the trajectory of health and life expectancies, and analyses the relative importance of determinants of healthadjusted life expectancy (Wolfson and Rowe, 2013) IFS Model UK Studies pensioner poverty under a variety of alternative tax and benefit policies (Brewer et al., 2007) IFSIM Sweden Studies intergenerational transfers and the interdependence between demography and the economy (Baroni et al., 2009)

INAHSIM
Japan Simulates demographic and social evolution, able to simulate kinship relationships in detail (Inagaki, 2010) INFORM UK Developed for forecasting of benefit caseloads and combinations of receipt, designed to incorporate significant benefit reforms planned over the coming years, based entirely on administrative data (Gault, 2009)

Italian Cohort
Italy Analyses lifetime income distribution issues (Baldini, 2001) Japanese Cohort Japan Looks at the impact on household savings of changes in demographic structure (Ando and Moro, 1995;Ando, 1996) LABORsim Italy Simulates the evolution of the labour force over future decades in Italy (Leombruni, 2006) LIAM  (Hancock, 2000) Melbourne Cohort Australia Analyses income inequality in a lifetime context (Van de Ven, 1998)

MICROHUS Sweden
Models dynamic effects of changes to the tax-benefit system on the income distribution and economicdemographic effects of immigration (Klevemarken, 1991;Klevmarken and Olovsson, 1996) MICSIM Germany Analyses German pension and tax reform (Merz et al., 2002)

MiMESIS
Sweden Evaluates Swedish Pension Reform (Mikula et al., 2003) MIDAS Multi Analyses pension system and social security adequacy (Dekkers and Belloni, 2009) MIDAS New Zealand Models wealth accumulation and distribution (Stroombergen et al., 1995) LI, O'DONOGHUE A survey of dynamic microsimulation models: uses, model structure and methodology

MIND Italy
Simulates the economic impact resulting from alternative values of the income growth rate and real interest rate (Vagliasindi et al., 2004) MINT US Forecasts the distribution of income for the 1931-1960 birth cohorts in retirement, MINT5 extends to the 1926-2018 birth cohorts ( (Panis and Lillard, 1999;Smith et al., 2007;Toder et al., 2002) MOSART 1/2/3 Norway Models the future cost of pensions, undertakes micro level projections of population, education, labour supply and public pensions, incorporates overlapping-generations, models within a dynamic microsimulation framework (Andersson et al., 1992;Fredriksen, 1998)

MOSES Sweden
Investigates the micro basis for inflation and the interactions between inflation, economic growth, and profitability (Eliasson, 1977) NEDYMAS Netherlands Models intergenerational equity and pension reform, the redistributive impact of social security schemes in a lifetime framework (Nelissen, 1996;Nelissen, 1998) PENMOD Japan Public pension system analysis (Shiraishi, 2008) PENSIM UK Models the treatment of pensioners by the social security system across the income distribution (Hancock et al. 1992;Curry, 1996) PENSIM US Analyses lifetime coverage and adequacy issues related to employer-sponsored pension plans in the US (Holmer et al., 2001) PENSIM2 UK Estimates the future distribution of pensioner incomes to analyse the distributional effects of proposed changes to pension policy (Emmerson et al., 2004) Pensions Model Belgium Analyses and forecasts the medium term impact of a change to pension regulations (Joyeaux et al., 1996) POHEM Canada A longitudinal microsimulation model of health and disease, it is used to compare competing health intervention alternatives within a framework that captures the effects of disease interactions  POLISIM US Demographic-economic and social security projection for US social security administration (Holmer, 2009;McKay, 2003) PRISM US Evaluates public and private pensions (Citro and Hanushek, 1991a;Citro and Hanushek, 1991b) SADNAP Netherlands Evaluates the financial and economic implications of the problem of ageing (Van Sonsbeek, 2009) SAGE UK Dynamic demographic/tax model for the UK (Zaidi and Rake, 2001;Zaidi and Scott, 2001) SESIM Sweden Analyses the consequences of population ageing and models budget and distributional impact of inter-temporal policy issues such as student grants, labour supply, savings decisions and pensions (Ericson and Hussenius, 1998;Ericson and Hussenius, 1999;Klevmarken and Lindgren, 2008;Klevmarken, 2010;Pylkkänen, 2001) Sfb3 Germany Analyses pension reforms, the effect of shortening worker hours, distributional effects of education transfers (Galler and Wagner, 1986;Hain and Hellberger, 1986) SimBritain UK Simulates urban and regional populations within the UK (Ballas et al., 2005a;Ballas et al., 2005b);

SMILE Ireland
Population projections with spatial details for Ireland (Ballas et al., 2005a

SVERIGE Sweden
Models human eco-dynamics, e.g. the impact of human cultural and economic systems on the environment (Vencatasawmy et al., 1999) Swedish Cohort Sweden Models the replacement of social insurance by personal savings accounts and the distribution of lifetime marginal effective tax rates (Fölster, 2001) Tdymm Italy Analyses the Italian labour market and pension system, with a focus on pension adequacy and related distributional effects (Tedeschi, 2011) XEcon Canada A model intended for theoretical exploration rather than practical empirical application (Wolfson, 1995) Although Table 1 tries to cover as many known models as possible, it is nearly impossible to give a complete list as new models are being developed every year. In addition, the list focuses more on the dynamic microsimulation models that are mainly used for social economic analyses.
Certain regional dynamic spatial models and transportation models are not included.
One can also track the development of models through a number of lineages. The original Orcutt Socio-economic System (Orcutt et al., 1961)  All these powerful dynamic microsimulation models come with the cost of high complexity.
Compared with static microsimulation, dynamic microsimulation is much more costly to develop and also has more issues with the methodologies used. This paper intends to discuss some of the methodological issues related to the construction of a dynamic microsimulation model, surveying current practice in the field around the world.

An overview of the technical choices made by dynamic microsimulation models
This section continues the discussion of methodological issues faced in constructing dynamic LI, O'DONOGHUE A survey of dynamic microsimulation models: uses, model structure and methodology microsimulation models but focuses on the technical implementation and choices made in a model. Table 2 describes the data source and sample size used by different dynamic microsimulation models while Table 3 provides an overview of the technical choices of the modelling structure. These two tables will be discussed in details in this section.

Base dataset selection
Base dataset selection is important in a microsimulation model as the quality of the input data determines the quality of the output. However, selection of a base dataset is not an easy task as hardly any micro dataset contains all the information required by a dynamic population microsimulation model and the difficulties of picking a base dataset have been discussed by Zaidi and Scott (2001), Cassells et al. (2006), Klevmarken and Lindgren (2008) and many other papers.
There are a number of different types of base data that a dynamic model can utilise. As shown in Table 2, a dynamic microsimulation model typically starts with one or several of the following types of dataset according to their sources: Administrative data often contains extensive information in taxable earnings and basic (taxrelated) social economic variables and the data is often collected for the most part of the population, with a much bigger sample size compared with survey data. Because the data is often collected for taxations or law enforcement purpose, the data could be very accurate for some variables, e.g. employee earnings, but very misleading in certain self-reported information, e.g.
household wealth. Since the administrative data collection is a government initiative, some information that is useful for social economic researchers may not be included. For this reason, models using administrative data often seek to supplement information from external sources, such in the case of Sweden (SESIM model) where certain variables were imputed based on a survey.
Legal and privacy reasons may also prevent administrative data from being accessible. Models such as CORSIM, DYNACAN and DYNAMOD use census data. While census datasets typically have higher number of observations than household surveys, they often contain less information and have to be supplemented with imputed information from other sources.  (Caldwell, 1996), whilst LIAM simulates a historical profile by exploiting retrospective variables, previous censuses and other data sources (Li and O'Donoghue, 2012).
Each of data matching/imputation methods has pros and cons and is often tailor-made to the specific datasets and projects. Statistical matching can be used when there are sufficient matching variables in a comparable dataset and this method has the desirable feature of having a "realworld" value, although the quality of matching may vary substantially depending on the quality LI, O'DONOGHUE A survey of dynamic microsimulation models: uses, model structure and methodology and quantity of matching variables. Synthetic simulation has the advantage of flexibility but longitudinal consistency may be an issue due to the limited information available. Additionally, it is also not uncommon to estimate behavioural relations from a dataset other than the base dataset and applies the equation to simulation. In this case, however, one may want to consider the comparability and consistency issues carefully due to the survey design and variable definitions.
Another issue in the base dataset selection is sample size; the larger the sample size, the more subgroups can be considered. Sample sizes are more important for inter-temporal analysis because the number of dimensions is increased, as similar individuals in a cross-sectional sample may in fact be very different due to the different paths taken to reach the same state. Regardless the source of the dataset, panel data is usually preferred as it records changes over time. Sample size also has an impact on run time of the model; the larger the samples size, the longer the run speed, resulting in a trade-off. Faster computer power does however reduce the impact of this trade-off.

Cohort model or population model
One issue that is closely related to the base dataset selection is the type of data structure that a model uses. Harding (1993) and others have categorised inter-temporal dynamic models into two types: cohort models that model a single cohort over a relative long time period (usually lifetime), and population models that model a population cross-section over a defined period of time. In addition, some models focus only on adults (i.e. ignore children) and thus, although these models may contain a cross-section of the population, they do not represent the entire age spectrum. In many cases, only persons born within a particular year are included in the model. LIFEMOD (Falkingham and Lessof, 1992), for instance, only includes 4000 people born in 1985.
From a model design perspective, the distinction between cohort and population model is less significant than the use that the model is put to. The distinction made in the literature from a historical viewpoint has more to do with computing power and data constraints rather than any major methodological differences. Cohort models were typically used because the computing costs required to simulate whole lifetimes for cross-sections with sufficient sample sizes to be able to examine specific cohorts were too high. It is also possible to model both types using the same computing platform. The potentially larger size of the cohort modelled in dynamic cohort models allows life time income patterns for smaller population groups such as recipients of disability benefits or lone parents to be studied.
Some cross-section models such as MOSART combine both types of modelling technique as they may use a very large dataset. With increasing computation and modelling capacities, newer models tend to use population models as one may get more information and draw inference to the population directly.

Ageing method in dynamic microsimulation
Aging within a microsimulation context may be defined as the process of updating a database to represent current conditions or of projecting a database for one or more years to represent expected future conditions. There are two types of ageing processes; static ageing and dynamic ageing. Static ageing, as described in Immervoll et al. (2005), is a method that attempts to align the available micro-data with other known future information (such as changes in population aggregates, age distributions or unemployment rates), without modelling the processes that drive these changes (e.g., migration, fertility, or economic downturn static ageing also has a number of disadvantages. Klevmarken (1997) highlighted that whereas static ageing may avoid some problems of drift in the projected cross-section associated with dynamic ageing because of misspecification in dynamic equations, it cannot account for mobility between states. In addition, he pointed out that from a statistical point of view, it is inefficient not to use all available historical information to project into the future. A consequence of not modelling the mobility of individuals between points in time is that it reduces the type of analyses that can be undertaken by a microsimulation model, for example, it is not possible to conduct analyses that require life event histories such as the simulation of pensions. Furthermore, future weights need to be forecast in order to age a dataset. Although macro models or other forecasting devices can be used they may not forecast weights at the level of detail required.
Besides, the weight calculation may be further complicated when multiple variables are being This may be possible in a simple model with a small number of processes, but in a fully dynamic model of work and life histories, many more processes need to be jointly estimated, a formidable requirement given the available data. Therefore, it might be necessary to make some assumptions to make estimation feasible, e.g. education choice happens before labour participation choice etc.
In addition, one may also need to assume independent error terms and some other arbitrary assumptions in order to simplify the estimation. Although these assumptions are common in practice, they may lead to theoretical pitfalls and biased results when excessively used without proper testing. In addition, projections over time at the micro-level are particularly susceptible to misspecification error as modelling at this level involves more detail than in macro models, also current knowledge regarding micro-behaviour is not good enough to specify a fully dynamic  Chénard 2000Chénard , 2001.

Discrete or continuous time modelling
Another choice in the development of dynamic microsimulation models is the treatment of time.
Discrete time models simulate which individuals experience particular events in given time intervals while continuous time models treat time as a continuous variable and determine the exact time that an event occurs (Willekens, 2006).
Discrete time microsimulation models changes once per time period. Take demography for example, demographic modules in dynamic models are often constructed using annual transition probability matrices. Individuals are passed through a collection of transition matrices in each time period of the simulation (usually a year) to determine their simulated life paths, e.g. death.
This method often assumes a sequential order of life events, however in reality they may be interdependent. As in the example given above and consequently the order in which the transition matrices are applied is very important. In the example given above, if marriage is determined first, then the potential fertility rate changes and similarly, a pre-marital pregnancy will increase the probability of getting married. Galler (1997) discussed a number of options in this situation including the procedure of random ordering as used by the DARMSTADT (Heike et al., 1987) and Hungarian models (Csicsman et al., 1987).
There are a number of other problems with this type of approach. Firstly transitions are assumed to take place at a single point in each time period and the duration of the event must last at least one time period (typically a year, but may be of shorter duration). For example if the time period is a year, this approach rules out transitions in and out of unemployment over the course of a year, which is unrealistic, as many people will have unemployment transitions for periods of less than one year as in the case of seasonal workers. Therefore, the discrete time transitions simulate net transitions (see Galler, 1997)  DYNAMOD and SOCSIM (Hammel et al., 1990) and this method was extensively discussed by Willekens (2006). Once a referencing event has occurred such as marriage, an individual is passed through each survival function that they are eligible for given their current states. For example, once an individual is married, they become eligible for divorce, the event given their current state with the nearest event time is selected and then repeated until death.
While the continuous time model has some theoretical advantages as it pinpoints the time of events, it also suffers from considerable practical limitations. The estimation of competing risks and survival functions place very high requirements on the data that are rarely matched by the actual data available (Zaidi and Rake, 2001). Given that most base datasets were collected yearly and many taxation procedures are reviewed annually, it is easier to incorporate a discrete time framework. Although a continuous time model could simulate the sequence of event occurrences, it still faces the estimation problems of interdependent processes and correlated errors. In addition, the potential interdependence of transitions for members (e.g. family) further raises the complexity of implementation. Alignment for continuous models is more difficult as cross-sectional adjustments would erode the advantages of duration models, and the potential computation cost of alignment is much higher in continuous time models. with macro-aggregates, as the process would require constant dynamic reweighting and in the case of heavy alignments, the benefits of running the model in parallel might be lost. As a result, most dynamic models in use utilise a closed model method.

Link between micro and macro models
As many policy indicators (e.g. employment rate) are at macro level, microsimulation models increasingly find the need to interact with macro economy through either an alignment process or computational general equilibrium (CGE) feedback. Alignment, as discussed in earlier, offers a simple but limited way to enforce the aggregate statistics within a simulation; however, it is usually limited to very specific variables and does not change based on the feedback from simulated micro data. Besides alignment, it is also possible to use CGE models to link macro, meso and micro models (see Ahmed and O'Donoghue, 2007;Davies 2004). CGE models offer a potential opportunity to allow macro models interact with micro models via prices in different markets, which is particular useful for analysing large scale macroeconomic shock. For instance, IFSIM (Baroni et al., 2009) links a microsimulation model with a simple CGE model assuming a single sector economy.
There are a few papers discussing the potential methods of linking a microsimulation model and a CGE model. Cockburn (2001) used an integrated approach to link a survey dataset within a CGE framework, where the main concept was to replace the traditional unit of analysis in CGE, representative household, with a real household. Another approach is to separate macro and micro components while allowing the result of the micro or macro models being fed into the other models. Depending on the direction of the output feeding and the number of iterations, this approach was further subcategorised into "Top-Down", "Bottom-Up", "Top-Down Bottom-Up" and "Iterated Top-Down Bottom-Up" approaches (Galler, 1990;Baekgaard, 1995;Savard, 2003 Robilliard and Robinson (2003) indicated that current approaches in linking micro-macro may still need to be refined before addressing distributional issues. In addition, linking with CGE requires decent quality of household income and expenditure data, which is not widely available. Furthermore, the integration between CGE and dynamic microsimulation could potentially exaggerate the uncertainty introduced in the results due to the complexities in interactions of different social economic variables and consequently result in a greatly increased computation time.
Given the complexity in incorporating complete a CGE model in microsimulation, it might be more feasible to incorporate a partial equilibrium in to the model. A static microsimulation model, IZAΨMOD (Peichl et al., 2010), allows feedback from computed labour market equilibrium to refine the labour supply behavioural response. In a spatial microsimulation model, one may consider to model the feedback from the housing market as illustrated in Hooimeijer (1996). This type of single market equilibrium implementation can avoid the complexity introduced by the social accounting matrix and inaccurate expenditure data.

Links and integration with agent based models
Although this study mostly focuses on the development of dynamic microsimulation models, it is also worth to note that microsimulation is closely related to two other individual level modelling approaches, cellular automata and agent based models (Williamson, 2007). In particular, agent based models are also used in social science to analyse macro level phenomena gathered from micro units. An agent based model (ABM) typically consists of a set of autonomous decisionmaking entities (agents), a set of agent relationships and methods of interaction, and the agents' environment (Macal and North, 2010). It is often used to show how macro level properties such as spatial patterns and levels of cooperation emerge from adaptive behaviours of individuals.
Traditionally, agent based models are highly abstract and theoretical without many direct empirical applications (Boero and Squazzoni, 2005;Janssen and Ostrom, 2006). In recent years, however, there is a growing interest in ABM literature of injecting empirical data in an attempt to simulate some real-world phenomenon (Parker et al., 2003;Hassan et al., 2008). From a practical point of view, when agent based models add more social economic attributes to the agents and when microsimulation models add more behaviour rules and interactions, they are moving toward to a common ground (Williamson, 2007). Some papers, e.g. Eliasson (1991), Baroni  With the growing number of social networking data, it becomes possible to integrate the empirically-tested adaptive behaviours from ABM into microsimulation models to produce a more realistic model. The potential introduction of network effects could benefit a set of microsimulation models, e.g. health simulation models, in which the social factors may play a role. In addition, peer effects may also help to model the evolution of marriage/fertility patterns, the formations and dissolutions of neighbourhoods in a spatial microsimulation model etc.
It should also be noted that this potential integration may also bring some disadvantages. The implementations of micro interactions would greatly increase the computational cost and complexity, thus makes the model more difficult to understand and validate. Besides, the current base datasets of the microsimulation models are often standard surveys or census data that do not cover extensive network attributes. At the current stage, the implementation of extensive interactions like ABM in microsimulation models is still at its infancy, the existing attempts are limited to the introduction of simple behaviour rules, e.g. copying consumption habits as in Lawson (2011). It is the easiest way to model potential changes with least theoretical considerations.

Modelling transitions and behaviours
Reduced form models and transition matrices are often used to simulate mortality, fertility, family formation, labour market transitions etc. As these models usually do not depend on policy parameters, they are not suitable for reform analysis, and are often restricted to simulating status quo only. The method is often used in static tax benefit microsimulation models as well as demographic components of dynamic microsimulation models.
In a structural behavioural model, individual behaviour changes are as a result of changing policies, therefore the policy parameters must have a direct or indirect impact on the model. An example could be a labour supply model that responds to changes in the tax-benefit system.
Behavioural models in microsimulation should be relevant to overall model objectives, able to introduce behavioural adjustments to policy changes. Examples of behavioural responses that fit these requirements include labour supply, retirement decisions, the effect of income and price changes on consumption, fertility and marital decisions, the take-up of social benefits etc. A behavioural model may contain multiple components. In the case of labour supply models, behaviour simulation models typically consist of three parts: an arithmetic tax benefit model to estimate budget constraints, a quantifiable behaviour model using variables that can be simulated, and a mechanism to predict the labour supply under a new policy environment (Creedy et al., 2002).
Compared with earlier microsimulation models, more models today have incorporated behavioural responses into their designs although these responses are often limited to labour market simulations. Models such as MICROHUS (Klevmarken and Olovsson, 1996), PRISM (Kennell and Sheils, 1990), NEDYMAS (Dekkers et al., 1993), LIAM  all incorporate labour supply behavioural responses to the tax-benefit system, SESIM (Klevmarken and Lindgren, 2008), DYNAMITE (Ando et al., 2000), and SADNAP (Van

Alignment with projections
As statistical models are typically estimated using historical datasets with specific characteristics and period effects, projections of the future may therefore contain errors or may not correspond to exogenous expectations of future events. In addition, the complexity of micro behaviour modelling may mean that simulation models may over or under predict the occurrence of a certain event, even in a well-specified model (Duncan and Weeks, 1998). Because of these issues, methods of calibration known as alignment have been developed within the microsimulation literature to correct for issues related to the adequacy of micro projections. Scott (2001) defines alignment as "a process of constraining model output to conform more closely to externally derived macro-data ('targets')". Clearly, in an ideal world, a system of equations would be estimated that could replicate reality and give effective future projections without the need for alignment. However, as Winder (2000) stated, "microsimulation models usually fail to simulate known time-series data. By aligning the model, goodness of fit to an observed time series can be guaranteed." Some modellers suggest that alignment is an effective pragmatic solution for highly complex models (O'Donoghue, 2010), as it offers a limited connection between micro and macro data.
Alignment also has its downsides, as highlighted by Baekgaard (2002). Concerns raised regarding alignment include the issue of consistency within the estimates and the level of disaggregation at which this should occur. The implementation of alignment may twist the relations of key variables in an undesired way (Li and O'Donoghue, 2013). The existence of an alignment mechanism may constrain model outputs to always hit aggregate targets even if there has been an underlying behavioural or structural change. An example would be if education levels rose, as this would be expected to reduce mortality rates and increase female labour force participation. If the alignment mechanism for each process did not incorporate the impact of educational achievement, then an increase in the education level would have no effect on these aggregates. It has been suggested that equations should be reformulated rather than constrained ex post. Klevmarken (2002)  Despite the potential pitfall of its statistical properties, aligning the output of a microsimulation LI, O'DONOGHUE A survey of dynamic microsimulation models: uses, model structure and methodology model to exogenous assumptions has become standard over the past decade. As Anderson (2001) noted, almost all existing dynamic microsimulation models are adjusted to align to external projections of aggregate or group variables when used for policy analysis. Continuous variables such as earnings are typically aligned with a fix ratio in order to meet the projected average or distribution, whilst binary variables, such as working status, are aligned with various methods, including multiplicative scaling, sidewalk, sorting based algorithms etc. (See Morrison, 2006).
Microsimulation models using historical datasets, e.g. CORSIM, align their output to historical data to create a more credible profile (SOA, 1997), while models that work prospectively, e.g.
APPSIM, also utilise the technique to align their simulations with external projections (Kelly and Percival, 2009).

Model complexity
Dynamic microsimulation is mostly built on the assumed parameters, estimated Markov chains and the conditional probability distributions estimated by various econometric methods. It usually involves many equations and parameters, estimated and fixed by laws and regulations.
Once the estimations and parameters are put in place, most microsimulation models follow a straightforward execution process without invoking computational complicated algorithms. The complexity of a model, as a result, often comes from the constructions of the economic models and is mostly guided by the potential policy questions that the model is required to answer.
Models focusing on pension issues usually simulate detailed labour market behaviour for decades ahead, as a change in the pension system can only mature when the youngest cohort in the labour market retires. In contrast, short term tax policy models usually forward simulate 3 to 5 years and are typically limited to tax related variables only. If a model is being utilised to answer different research questions, then it usually needs to simulate more variables for a longer period of time, which involves higher levels of complexity.
An ideal microsimulation model should have the capacity to simulate details of all possibly related variables; however, the costs of building large models, both in terms of model validity and management needs to be taken into consideration. Dynamic microsimulation models have the reputation of being complex and the potential to run "out-of-spin" with regard to some aspects.
This might be a particular concern when simulating policy reforms. Economics models, especially reduced form models, are often criticised for simulation purpose as the stability of the model structure is questionable under policy changes. This argument is also widely known as Lucas' critique. As a result, structural models are usually seen as a better choice. Since some part of the policies (e.g. tax) can be explicitly included in the structural model, the estimated utility LI, O'DONOGHUE A survey of dynamic microsimulation models: uses, model structure and methodology parameters are perceived to be more stable (Klevmarken and Lindgren, 2008), although utility parameters can sometimes be very sensitive to even a small change in the model specification.
Over fitting may also be a potential issue when the list of explanatory variables grows. Due to the number of models that one microsimulation model can invoke and the budget/time constraints, many models are primarily constructed using reduced form models with some structural models in key components.
Complex models, while having more power, are much more difficult to validate and may often contain bugs in their implementation. In addition, the complexity of the processes often means development takes longer than expected. Large general purpose microsimulation models are usually built by large teams with access to large and complex datasets. These models usually simulate a wide variety of economic and demographic processes and can therefore be used for many different applications. These forecasting models usually incorporate alignment systems in order to keep the model in line with external forecasts or are in fact linked to macro-models.

Models of this type include DYNASIM from the US, DYNACAN in Canada, SESIM in Sweden,
MOSART in Norway and the APPSIM in Australia etc.

Model validation
Given the increasing complexity of models, it becomes increasingly important to validate the model in order to maintain its credibility. Unfortunately, only limited effort has been placed on validation matters and there is no international consensus on validation procedures. Klevmarken and Lindgren (2008) suggest that validation should be put in the same context as estimation and testing, and should involve the identification of all sources of errors and their properties. Given the size and structure of a large microsimulation model, bootstrapping and Monte Carlo exercises are likely to be more practical than the analytical deduction. In addition, sensitivity analysis on the models should also be part of the microsimulation validations (Klevmarken and Lindgren, 2008). (2008) Morrison, 2000;Caldwell and Morrison, 2000). However, these models invariably incorporate historical information such as macro-aggregates into the model and as this information would not have been known to forecasters, this may produce better forecasts than would have otherwise been the case. One method to overcome this is to compare directly generated forecasts with what happened in reality, for example comparing forecasted labour participation rates with actual rates. Another method described by Caldwell (1996) is to use an indirect approach, known as a multiple module approach. An example cited by Caldwell is the case of validating the numbers of married persons with health insurance, when the directly simulated processes are marriage and medical insurance membership. Sources of error may result from errors in either or both direct processes, or because of mis-specified interactions.

In Morrison
Some types of dynamic models, however, may have no comparable data source for validation.
For example, some theoretical models that solely look at a single cohort living in a steady state have nothing with which they can be validated through external data source as they do not attempt to mimic real life in a comprehensive style. These types of models, due to the lack of validations, are often restricted in their interpretations of policy impact. Additionally, countries that have only recently developed their micro-data resources may not have alternative sources of data with which to validate, although this problem will become progressively less with time.
Recent developments suggest an alternative validation method using a simplified model. Since no future data is available to validate a forecasting dynamic microsimulation model, Morrison (2008) suggests comparing a model's result to a trustworthy model's result. Dekkers (2010) argues that the general trend of certain indicators estimated by a simple model could be seen as a benchmark for more complicated microsimulation model as there is no black box in a simple model. The Belgium MIDAS model used this approach to validate against a "simple stylised" model, which is essentially a representative household model with only demographic and pension indexation components. This approach, however, raises another question on the criteria of a "trustworthy" Visual Basic (e.g. SESIM) and a number of other languages. Evaluation based on the prototype microsimulation models by Percival (2007) suggested that there is a substantial speed advantage when the prototype model is coded using C++. However, the cost of development is much higher when compared with other approaches, as all the potential numeric calculations and related data management need to be programmed. In practice, it is likely that policy modellers are not adequately proficient in programming, while professional programmers may not fully understand the economic theories and scientific principles that are to be implemented.
The second approach is to develop microsimulation models based on existing statistical or mathematics packages, such as GAMEO, DYNASIM, and PENSIM2 which were built on SAS.
Modern statistical packages are capable of executing computation commands in batch and the code of a model is commonly referred to as a "script", "do-file" or "syntax file" etc. The main advantage of this approach is short development time and easy access to the statistical power of the package. However, the speed of the model might be lower and the data structure may not be optimised for simulation purpose.
Another way to develop microsimulation models is to use modelling packages developed by the third parties. Software in this category ranges from generic purpose modelling software, e.g.
AnyLogic etc. to microsimulation specific packages. Whilst agent based simulation modellers use more generic purpose packages, microsimulation modellers tend to use more specific software.  (Wolfson and Rowe, 1998) o LIAM2, developed by the Belgium Federal Planning Bureau (Bryon et al, 2011) o UMDBS, developed by Sauerbier (2002) o GENESIS, developed by UK Department for Work and Pensions (Edwards, 2004)  Modgen provides a C++ library for developers to incorporate required common actions and modules. It is often used to develop continuous microsimulation models, such as LifePath. In contrast, LIAM2 is a microsimulation scripting engine which is capable of reading its own syntax.
The engine itself is coded in Python with extensive references to libraries written in C.
Microsimulation packages such as this one offer the great benefit of rapid development. There are also a few dynamic microsimulation models that were built with generic deployment in mind.
For example, LIAM avoids hardcoded parameters and variable names during the development, which greatly reduce the repetitive work load of a new modeller. GENESIS offers a platform to create a SAS based microsimulation model by reading the model specification from an Excel Sheet. JAMSIM combines several opens source packages, e.g. R, Ascape, to create a user friendly interface to model and execute microsimulation models. However, these models still require end users to understand the internal mechanism and make changes at source code level when building larger models.
Most models today are based on a statistical package or a generic purpose programming language.
However, it is not uncommon to see mixed combinations of environments in order to utilise the advantages of different software, especially for pre-or post-simulation analysis, e.g. DYNACAN, LIAM.

Progress of dynamic microsimulation modelling since 1970s
In reviewing progress made by the field, it is useful to consider an early model development, the DYNASIM model developed by Orcutt et al. (1976)   These challenges can be broadly categorised into five different areas: behaviour response modelling (a-c), microdata quality (d), development cost (e), limited computation capacity (f) and model validation (g-h). Comparing with some recent discussions in issues of microsimulation (Harding, 2007b), it is clear that most issues mentioned are still relevant and high on the list several decades later. By comparing the DYNASIM model structure with today's dynamic microsimulation models and the challenges faced by the modellers in 1980s and today, what we are seeing are gradual advancements in the methodologies rather than breakthrough in model designs and applications.
Improved computer hardware has allowed both improved speed and increased databases as have model software developments such as Scott (2001) validate the simulation output (Morrison, 2008). Another major advancement in the past decades is the emerge of generic models, including Modgen (Wolfson and Rowe, 1998), UMDBS (Sauerbier, 2002), GENESIS (Edwards, 2004), LIAM ) and LIAM2 (Bryon et al, 2011). These generic models can greatly reduce the workload of new modellers by providing commonly used microsimulation routines.

Obstacles in the advancement of microsimulation
While the field of microsimulation has progressed greatly in many aspects since the original paper of Orcutt (Williamson et al, 2009), the rate of progress in dynamic microsimulation, nonetheless, is arguably slow given that we still share the same model design and face similar problems as early DYNASIM modellers did nearly 40 years ago. There are a number of reasons could be ascribed to this lack of progress, including: can be gained and motives private R&D, given the relatively small demand for these tools by clients with the capacity to pay for them, it seems to be a business model that will stymie intellectual development. With low funded demand, the returns will be low, limiting private investment, while the private good nature of the intellectual property will limit transmission. In addition, the protection of the source code makes replicating others' findings very difficult, which could harm the credibility of the research. Moving from black box" modelling to a glass box" modelling could ease many potential users' concerns and raise the method's scientific status (Wolfson, 2009). The availability of less closed model frameworks such as GENESIS, LIAM and LIAM2 can facilitate the development of new models. However it will require the publication of full models to fully realise the benefits of scientific interaction.
Another reason for the lack of progress might be the perceived "failure" of the earlier models.
However this failure to some extent can be attributed to failing to meet unachievably high expectations. Orcutt et al. (1961) focused on the capacity to undertake prediction at a micro level to facilitate planning. Human behaviour is of such complexity and is endogenous to economic analysis that dynamic microsimulation models cannot hope to make highly accurate predictions.
Even well specified econometric model may over or under predict the outcomes (Duncan and Weeks, 1998). As George Box (1987) said "All models are wrong, some are useful". In being useful we can hope, by using good theory, data and statistical and computational methods, dynamic microsimulation models can provide a consistent and reasonable framework with which to undertake policies analyses incorporating inter-temporal events and the distribution of the population.
Funding may also be a major issue facing many microsimulation modellers. Building and using large microsimulation models requires teams of researchers representing different disciplines and experiences. In addition, the scale of the model also suggests the need for long-term funding, which does not always fit well into an ordinary university department with its normal rotation of  Buddelmeyer et al., 2009), modelling disease spread , time use simulation (Anderson et al., 2009), and even to assist personal financial planning (Avery and Morrison, 2011). The use of dynamic microsimulation models can be even further expanded as more micro-level data becomes available. With the better availability of the longitudinal data and administrative data, it is possible to better understand the consequence of ageing. In addition, the raise of the network data could help to model the disease spread and knowledge diffusion in a more realistic way.
While large dynamic models have their advantages for providing more comprehensive simulation outputs, the complexity also increases the difficulties in validation, model usages and also funding issues. It might be beneficial to also develop some specialised simple dynamic models. Smaller models could be better validated and make it easier to publish the model details within the length limit of a journal article. This does not mean that there is no need for large models. Instead, the easy-to-validate smaller models could be absorbed into a more complicated microsimulation model where more complex interactions are allowed.  Li and O'Donoghue (2011). The method may assist us to better understand the many inter temporal processes, e.g. fertility decision, education choices etc. The raising interests from the academic side would benefit the field development and ensures the sustainability of the knowledge.

Model assumptions and potential expansions
While the models that focus on the individual social economic patterns and behaviours have improved in the past decade, there is limited effort modelling the institutions except tax benefit system within microsimulation. Future models could potentially benefit from incorporating some of these important social fabrics, e.g. companies, unions in a labour market simulation exercise, or hospitals, clinics, emergency response in a health model. This also relates to the discussions of better behaviour modelling and CGE models linkage referred in earlier section. These improvements could potentially provide us a more accurate picture of the consequences of large scale policy change.
A number of the typical assumptions in microsimulation can also be further explored. For instance, the unit of analysis is traditionally individual or nuclear family. It is useful to consider more complex household arrangements to be comparable with the welfare economics literature.
Also, the sharing within household assumption may not hold if there's complex family structure.
By modelling the kinships within the family, like INAHSIM (Inagaki, 2010), it allows us to better understand the welfare network and poverty alleviations in the society. In addition, most microsimulation models these days ignore the budget and political constraint from the government side, either by assuming fiscally neutral or unlimited resources. The macroeconomic constraint and the political feasibilities are not discussed in most papers. Future microsimulation modellers may want to incorporate these factors to assess the feasibilities of a proposed reform.

Methodologies and technical choices
In terms of methodological development, a primary need is to codify the various methodologies that are currently being used in dynamic microsimulation models. There are many methods being used; most without any published description or evaluation. As noted above, this can impede the progress of the field. Formally documenting the methods used and publishing in a peer-reviewed journal could improve the knowledge diffusion and increase the public good returns by academics, providing incentive to innovate. Additionally, publications could preserve the LI, O'DONOGHUE A survey of dynamic microsimulation models: uses, model structure and methodology knowledge that could have been lost due to the end of project funding. It is hoped that in time that the newly founded International Journal of Microsimulation can provide an opportunity for citable peer reviewed publications.
The widely used alignment technique is still under documented. While some models have published their alignment implementation details, it is still unclear how alignment should be used when combined with more complex econometric models, e.g. categorical behaviour models.
Generally, there is limited understanding on the simulation properties of many algorithms used, including alignment, error term manipulation, complex reweighting, random numbers etc. In most models, each equation is estimated separately without the considering the potential correlated error terms and the simulation structure where the equations will be applied to. This may lead to undesired bias due to inconsistent assumptions when simulating some reforms. In addition, to improve the model credibility, it is worthwhile to pay attention to the testing and validation process of a simulation model. Additionally, papers using microsimulation model typically provide the result of only one-run although a few papers, e.g. Pudney and Sutherland (1994) found that the microsimulation results could have a wide confidence interval. Given the raising computing capacity available for researchers these days, modellers could potentially provide more information about the simulation, e.g. the confidence intervals of the result using Monte Carlo techniques.
Despite the discussions and the general consensus to improve validation process in microsimulation, there is still little guideline how dynamic microsimulation models should be validated (Harding, 2010). While DYNASIM documented many issues involved in the model validation, there are still many areas that need to be explored, such as behaviour responses validation, longitudinal consistency validation, module interactions etc. Besides the validation from the technical side, it is also worth considering to validate the simulation with historical data, from which we can learn what has been done right, how the simulation performs under different assumptions etc.

CONCLUSIONS
This paper has discussed some of main issues involved in constructing a dynamic Over the past decades, microsimulation models have been applied to many different policy areas and a comparison of models as given in Table 1 illustrates the scope of application of dynamic microsimulation models. Most dynamic microsimulation models listed can be categorised as discrete models using dynamic ageing approach. For newer models, alignment has become a standard component allowing interactions with macro aggregates and more recently, simulation packages that are dedicated solely to microsimulation have become a viable option in model development. These packages, together with increased co-operation through meetings and code sharing (e.g. LIAM, Modgen model series), could significantly increase the development process.
The increasing use of microsimulation models has raised many technical challenges to meet the needs of more complex and accurate policy analyses. For instance, there is a growing interest in integrating CGE into microsimulation models, although the actual implementations of CGEmicrosimulation are at this stage restricted due to data and technical limitations. Behavioural responses in microsimulation could also be further improved and one should consider more lifecycle models when simulating inter-temporal choices. Microsimulation models could potentially implement some elements from ABM to allow dynamic behaviour interactions and adaptations.
In addition, considering different unit of analysis, budget and political constraints, may also broaden the field of microsimulation applications. Furthermore, certain practices within the simulations, such as alignment for complex models and error term simulation, should be more thoroughly studied.
Besides the technical challenges, there are also some general issues in the field. The lack of documentations often forces new modellers to reinvent the wheel; closed sourced models which slow down the knowledge transmission. The unrealistically high expectation in long run simulation may challenge the creditability of the model and make applying for funding more difficult in long run. Future modellers may help to address these issues by publishing model details in academic journals and be more open on the algorithm implementations. Newer modelling platforms like LIAM2 attempt to be more open and transparent in the software source code, which would potentially benefit the field development and knowledge transmission. In addition, the field can also explore topics other than taxations and standard government policies.
Topics like the impact of climate change, the social consequence of ageing, for instance, could also potentially gain benefits from microsimulation techniques. For some examples of the multi-dimensional reweight algorithm, see Deville et al. (1993), Tanton et al. (2011) etc.