Conceptually, microsimulations belong between highly individual approaches, like agent-based simulations and aggregate-only approaches, similar to the system dynamics approach. They combine aspects of both of these ideas as a somewhat between approach. (Historically, microsimulations were developed after system dynamics but before agent-based approaches Macy and Willer, 2002). Instead of focusing on a generic flow rate (impacted through various other aspects), microsimulations account for differences in flows based on attributes. For example, a microsimulation model about population might acknowledge differences in birth and death rates based on age, allowing a more accurate estimation than using a single shared attribute to the model. Figure 6.3 illustrates the differences. On a system simulation we assume a shared birth rate and a shared death rate, but in reality, we know those are different across age groups. In Finland, people tend to give birth between 25 and 35 years and die in their 70s. (However, the numbers in Figure 6.3 are illustrative only and do not correspond to population statistics.) Therefore, the inflow of new children and outflow of the deceased depends on the population. If the Finnish population would consist only of the elderly in their 60s and 70s, there would be no influx of new children.
This adds some individual components into the model, but the simulation is not based on agents and their rules. When developing these models, we focus not on individuals but on flows and stacks - similar to system dynamics. The level of analysis is highly aggregated. Furthermore, microsimulation models build on data. They require information about the stack to build models that correspond to reality. However, these benefits are bounded by limitations. Microsimulation executes rules based on probabilities, but it does not interact or react to other agents nor does it have memory about past decisions it has done. Because of these benefits and limitations, the focus with microsimulations is often predictions and policy advice, unlike agent-based simulations that are used as thought experiments.
The baseline for a microsimulation is a representative sample of the target population. Representativeness is important to allow generalisation from the simulated sample to the target population. Often these are for population estimation, but similar to agent-based simulations, this is not necessarily true. The units could be all kinds of entities. Table 6.1 illustrates how this data could look. If we step the simulation by one year following the model presented on Figure 6.3b, Table 6.1b presents one opportunity of how the population looks at time . Furthermore, for women we would compute their probabilities for giving birth during the year (for simplicity and to avoid double counting, we would do this only for women). The population has three women: Unit 2, Unit 4 and Unit 5. Unit 5 is 50 years old, so her probability of giving a birth is 0%. Units 2 and 4 are between 25 and 35 years old, so their probability of giving birth is 35%. Based on drawing random numbers, we assume that this time it was Unit 2 who gave birth to unit . This increases her children number and creates a new unit in the population table. Often our analysis would only focus on the distribution of the population based on their characteristics like age - not understanding individual units in detail like in this description.
Beyond predicting the future, microsimulations are also used to analyse the impact of policy changes or interventions. They are suitable for this task as the simulation runs on a representative sample of the population. Thus, the simulation can help to project outcomes of policy changes for the whole population by examining how it would change the aggregate level of a variable. To illustrate these aspects, let us assume that the government would seek to identify new tax schemas for the population in Table 6.1. There can be various ideas for such policies. Code Examples 6.3 and 6.4 illustrate two potential solutions for the problem, seeking to make policies more friendly for larger families. (Here the differences in countries play clearly: Finland has a person-centric tax system, but in Germany the taxes are calculated per family unit. Therefore, this would directly influence the units of our analysis!) By running the simulation, we can estimate its impacts on individuals and governments and answer questions such as: Does the new tax schema increase money available for families with children? or What is the impact of them on the taxes gathered? By complicating the model, we could even add economic thresholds to the likelihood of giving birth.
The use of microsimulations is often applied. Microsimulations support forecasting futures and thus are commonly used to support decision-making or examining policy alternatives. Because of this, academic works focus more on the process of building these simulation models and describing how parameters were chosen rather than on helping develop theories (e.g. see von Randow et al., 2011; Sutherland and Figari, 2013; Tobias and Mosler, 2017; Pearson et al., 2010). For example, Pearson et al. (2010) examine the impacts that aging societies have on healthcare systems. The paper does not address different empirical outcomes from the simulation model, like in agent-based simulation models. Rather, their paper emphasises data sources used for the simulation and techniques they have used to improve the data sources. When reporting results, they are used to validate the model (running the simulation on a year where many of the output measurements would be known) or show opportunities for scenarios.