Simulated data for census-scale entity resolution research without privacy restrictions: a large-scale dataset generated by individual-based modeling

Using Vivarium

Vivarium is a mature, open-source simulation framework¹⁴ that uses standard scientific Python tools, such as NumPy and pandas^15,16. A simulation in Vivarium consists of user-written components that encapsulate the simulation logic, a machine-readable model specification that describes what components are in the model and how they are configured, and a data file containing all data used in the simulation. The framework provides a set of services to assist users in writing their model components, an engine for executing the simulations from both an interactive Python session and from the command line, and abstractions to help manage and format model input data. Models built in Vivarium are typically individual-based, representing people in a population as agents or “simulants,” each with their own age, sex, and other characteristics relevant to the specific model. They typically use discrete “time steps” at which events may occur. In this work, each Vivarium simulant represented a person living in the United States, and each discrete time step included changes relevant to data used in record linkage, such as births, deaths, moving to another address, and changing jobs.

We use a pre-established workflow when developing a Vivarium simulation, with roles for researchers and software engineers. The researchers lead the model development process through background research, conceptualizing modeling strategies, validating strategies with domain experts, guiding the conceptual development of the modeling software, and generating analytics for simulation inputs and outputs. The software engineers lead the development of simulation code, including model components and outputs, and tools supporting model and input data analytics.

In the following sections, we will cover the different input data sources and data processing strategies used to inform our simulation of the US population. We describe the Vivarium model components for simulant characteristics including basic demographics, household structure, mortality, fertility, migration, and employment dynamics. We also describe the addition of simulant names, physical addresses, employer names, and other attributes which we implemented as a post-processing step, rather than during the simulation itself.

Simulation time

We initialized our simulation to begin on January 1, 2019, and step forward in time with 28-day time steps until the simulation clock exceeded May 1, 2041. We chose this time step duration to balance the complexity of changes in demographics, housing, employment, etc. with the computational demand of running a simulation with over 300 million simulants.

Concept model

The concept model diagrams (Figure 1 and Figure 2) provide a visualization of the logical dynamics underlying this simulation and indicate how the various components of the simulation relate. The simulation components can be divided into three overarching categories:

Figure 1. Simplified version of simulation concept model diagram, denoting the four different overarching simulation events that can occur (migration, employment, mortality, and fertility) and how each of these events are observed during the simulation run.

Figure 2. Concept model diagram showing the interaction between different components of the simulation, illustrated at the level of an individual simulant.

Each arrow in this diagram represents a dependence between two distinct components of the simulation; an arrow from component X to component Y indicates that X affects Y (for example, the employment component simulates changes in jobs, which leads to a change in income; the basic demographics of a simulant affect the probability of death so that older simulants are realistically more likely to die during a simulation time step).

Figure 1 shows what influences the occurrence of each simulation event and how these events are captured in our data collection, while Figure 2 shows how the simulant components interact with one another at the individual level. When a simulant undergoes an event (e.g., gives birth, changes jobs, changes address), the simulant’s attributes change accordingly. Those attributes are then captured by the observers.

Input data. We informed the simulated datasets we developed for pseudopeople using open-source input data, including data released publicly by the Social Security Administration (SSA) and the USCB. We informed physical addresses from the training data of the Python package libpostal, as repackaged by the deepparse project¹⁷. In the sections that follow, we elaborate on how we used these data sources, and how our simulation could be extended to be even more realistic in future work.

Basic demographics. We initialized the simulated population’s demographic characteristics, including age and date of birth, race/ethnicity, sex, nativity (i.e., whether a simulant was born within the US), geographic area, and household structure by sampling from the 2016–2020 ACS Public Use Microdata Sample (PUMS)¹⁸. By sampling from PUMS, we were able to match the univariate distribution of each attribute as well as joint distributions of arbitrary complexity between the attributes at the Public Use Microdata Area level, while also preserving structure within sampled household units. For instance, the PUMS data capture the age distribution of people in America, where more people were born from 1945 to 1965 than from 1965 to 1985. The PUMS data are not without limitations, however. For example, the granularity of the PUMS data is limited by privacy considerations, and specific details that might be crucial for a detailed analysis are sometimes obscured, which could affect the precision of simulations based on these data, especially in socioeconomic and health-related contexts.

Age is reported in PUMS in floored integer years, but our simulation uses precise ages in fractional years. We assigned simulants a uniformly random precise age consistent with their nominal age as sampled from PUMS. For ER research and development, it was particularly important that we did not generate simulants who are much more similar to one another than would be expected in a real population, which would make linkage unrealistically difficult. Our simulated population is the size of the US population, but every simulant is initialized from a person in PUMS, which is a 5% sample of the US. Therefore, many simulants are created from the same person in PUMS, which could create unrealistic clustering. To decrease similarity without assuming total independence between attributes, we perturbed age values at sampling time. In different components of the simulation, we sampled different entity types from the PUMS: entire households, individuals living in group quarters (GQs), or individuals living in households (non-GQs). For each entity sampled, we added a random age shift taken from a standard normal distribution to that entity’s age value(s). When perturbation led to a negative age value, we flipped the negative age value’s sign. We then defined each simulant’s date of birth to be consistent with their precise age.

Sex is reported in PUMS as binary (male or female), so we initialized a sex attribute this way as well for each simulant. We mapped separate PUMS indicators of race and ethnicity to a single composite “race/ethnicity” indicator, with the following exhaustive and mutually exclusive categories: “White,” “Black,” “Latino,” “American Indian and Alaskan Native,” “Asian,” “Native Hawaiian and Other Pacific Islander,” and “Multiracial or Some Other Race.” We defined these categories in accordance with the guidelines provided by the US Office of Management and Budget (OMB)¹⁹. Nativity describes whether a simulant was born in the United States or elsewhere, and we modeled this as a binary variable in our simulation. We used this nativity attribute to inform the likelihood that the simulant had a Social Security Number (SSN). Table 1 provides a sample of the basic demographics present in the simulated population used in pseudopeople (note that these are entirely simulated data and therefore do not constitute Confidential Unclassified Information under US Code).

Household structure. Our simulants lived in either residential households or group quarters (GQ). We used the ACS PUMS data to inform the residential household structure regarding how each simulant is related to a reference person in their household. Simulants living in GQ do not have such a relationship and GQs do not have a reference person. Residential households and GQs have geographic locations as well as physical and mailing street addresses, which may be different, because some residential households receive mail at a PO box (we do not simulate other kinds of mailing-only addresses, such as rural route addresses).

PUMS data were not sufficient to identify precisely which type of GQ each simulant resided in; they only provided information on whether it was an institutional or non-institutional GQ. We subdivided institutional GQ into three mutually exclusive and collectively exhaustive categories of carceral, nursing homes, and other institutional. We also subdivided non-institutional GQ into college, military, and other non-institutional. We chose a GQ type uniformly at random for each simulant out of the three types consistent with their institutional/non-institutional status.

For simulants living in residential households, we modeled a relationship to the reference person of their household based on the relationship values in the PUMS²⁰. Possible relationship values were reference person, biological child, adopted child, stepchild, sibling, parent, grandchild, parent-in-law, child-in-law, other relative, roommate, foster child, other non-relative.

Mortality. To model mortality, we used our standard Vivarium approach, informed by data from the age- and sex-specific estimate of all-cause mortality for the US in 2019 as produced by the IHME Global Burden of Disease Study²¹. When a simulant who was the reference person in a non-GQ household died, we made the oldest remaining simulant in that household the new reference person and updated all other relationships (this produces some households with an unrealistically young simulant as the reference person). Unlike many of our past Vivarium simulations, we did not model the underlying cause for any simulant’s death. However, we could extend this simulation to model specific causes of death in future iterations of the simulation, such as to facilitate research and development in cancer registry linkage applications.

Fertility. We used our standard Vivarium approach to an age-specific fertility model in which each female simulant has a probability of having a birth event at each time step, derived from the age-specific fertility rate for the USA. In the current version of our model, only one female parent is identified, representing the simulant who gave birth. The birth event is considered to occur at a randomly chosen time during the 28-day time step, which informs the date of birth and age of the simulants born. We select a random 4% of birth events to be the birth of twins (two newborn simulants), and for the other birth events we add a single newborn simulant. We expect that the inclusion of twins will create some particularly challenging ER data, where simulants have the same last name, address, and date of birth. We do not include adoption or any other complexities of family structure.

The newborn simulant inherits certain attributes from their mother simulant, including household, race/ethnicity, and last name (recall that the simulation associates a newborn with only a single parent, so these attributes are inherited from this individual unambiguously). These simplifying assumptions allowed us to avoid modeling the complex dynamics of relationships but precluded us from following the dominant patriarchal naming pattern present in the US. The nativity of children born in the simulation is set to reflect that they were born in the US; therefore, all children born in the simulation are assigned an SSN. Additionally, we assigned newborns a relationship to the reference person in their household (which is also their parent’s household) based on the relationship between their parent and the household reference person, using a set of logical business rules.

Migration. We attempted to include accurate patterns of migration in our simulation, as migration leads to changing addresses, which constitutes an important challenge in ER. As with basic demographics, all data informing migration in our simulation come from ACS PUMS. We used PUMS to calculate migration by demographics. There are a huge number of attributes that could explain moving behavior, and they may interact in complex ways in the real world. We modeled only some of this complexity and captured three types of household and individual migration events: migration within the simulation (domestic migration), migration into the simulation (in-migration), and migration out of the simulation (out-migration).

Domestic migration

We modeled domestic migration events as happening at a rate determined by age, sex, and race/ethnicity; we held these rates constant across time in the simulation. Individual domestic migration caused a single simulant to move and might reflect an individual moving out of their current living situation (i.e., GQ or residential household) and establishing a new one-person household, moving into GQ, or joining an existing residential household as a non-reference person. For individual migrations in which a simulant establishes a new household, we always classified the simulant as the reference person. For individual migrations in which a simulant joins an existing household, the simulant is always classified as an “Other nonrelative.” We assumed that simulants have at most a single individual migration event per time step.

When a simulant who was the reference person in a non-GQ household moved, we assigned the oldest remaining simulant in their household to be the reference person and updated all other relationships in the household according to logical business rules.

Household domestic migration caused an entire household of more than one simulant to move as a unit. As with individual migration, we calculated the rate of household migration per household-year and stratified by demographics. Because households do not have overall demographic characteristics, we used the demographics of the reference person for this stratification. Unlike individual migration, we did not change any relationships in household migrations.

We used a simplifying assumption that all simulants who moved and were of working age (which we define as age 18 and older) changed employment.

International immigration

We modeled immigration by adding new simulants to our simulation to represent individuals moving into the US from other countries. We sampled simulants immigrating to the US from the subset of the 2016–2020 ACS PUMS who had immigrated to the US in the year before they were surveyed and did not perturb their age. This approach relies on our assumption that the number-per-year and demographic characteristics of recent immigrants in the 2016–2020 ACS PUMS will not change substantially for all future years of the simulation.

We modeled three kinds of immigration events in our simulation: household moves, GQ person moves, and non-reference-person moves. As with domestic migration, a household move is when an entire non-GQ household enters from outside the country as a unit, preserving relationships within the unit. A GQ person move is when a simulant enters from outside the country and joins group quarters. Because simulants who reside in GQs do not have tracked relationships in PUMS or our simulation, these moves have no relationship structure. Lastly, a non-reference-person move is when an individual simulant enters from outside the country and joins an existing non-GQ household with some relationship other than “reference person.”

We used the weighted number of last-year immigration events of each type from the ACS PUMS to inform the yearly rate at which immigration events of each move type occurred in our simulation. We simulated constant rates over time and did not model seasonal or temporal fluctuation in immigration.

International emigration

Emigration occurs when a simulant leaves the US to live in another country. We used the Net International Migration (NIM) estimates from the Census Bureau’s Population Housing Unit (PopEst) program to determine the number of emigrants per year, by subtracting immigration numbers from ACS to isolate emigration. The NIM estimates are made by the PopEst team by combining information about immigration from ACS with information about emigration from demographic analysis (for those born outside the US) and analysis of foreign censuses (for those born in the US)²².

There are three types of emigration events that can occur in our simulation: household moves, GQ-person moves, and non-reference-person moves. These cause an entire household, a GQ person, or a household member who is not a reference person to leave the US, respectively. We stratified emigration rates by age group, sex, race/ethnicity, nativity, and US state of residence, and we assumed that these stratified rates were constant over time, without a long-term trend or seasonal variation. We stopped tracking households and individuals after an emigration event and assumed that they would not return to the US or appear in any pseudopeople data after they had emigrated.

Employment dynamics. We consider all simulants aged 18 years or older to be working age; all such simulants either have an employer or are considered unemployed. We only allow a single employer at a time for each simulant. We initialized the working-age simulants to be unemployed, employed in the military, or employed otherwise, and we considered the military to be a single employer. To employ the rest of the simulants (those with non-military jobs), we generated employers with an initial size attribute chosen from a skewed distribution to ensure that there are a few large employers and many small employers. In order to assign individual simulants to employers such that the size attribute is (roughly) accurate at the population level, we selected each simulant’s employer from the categorical distribution where the probability of each employer is proportional to its initial size attribute.

Working-age simulants (including those who are unemployed) change employment randomly at a rate of 50 changes per 100 person-years (a rate we selected subjectively to provide an appropriate challenge in record linkage). When a simulant changes employment, we sample a new employer with the same procedure used at initialization. This approach to selecting a new employer ensures that at the population level, the number of simulants employed by any given employer will remain roughly proportional to the initial size attribute sampled for that employer.

We also simulate income, which affected which datasets a simulant appeared in. For instance, the WIC dataset only recorded simulants with household income below a certain threshold. We approximated the income distribution with a log-normal distribution for each age group, sex, and race/ethnicity combination, fit to the ACS PUMS. See Appendix 1 for more detail on the distribution parameters used for each demographic group. We simulated only income earned through wages; unemployed simulants had no income. To simplify our model, we assumed statistical independence between wages and employer for employed simulants.

Post-processing

We added some elements to the simulated data after the simulation ran. This includes features that would require additional computing resources to track for the simulation’s duration, such as simulant names (first and last), employer names, and government-issued identification numbers (i.e., SSNs and ITINs).

We developed simulant first and last names based on two distinct data sources: first and middle names are sourced from SSA data, which allowed use to match name frequencies to age and sex; while last names are generated from Census data with hyphens and spaces added to make linking tasks realistically more challenging, which allowed use to match the frequencies by race/ethnicity^23,24. We generated SSNs in accordance with the current algorithm used to issue unique SSNs. We selected the first three digits uniformly at random from 001 to 899, excluding 666; the next two digits from 01 to 99; and the final four digits from 0001 to 9999²⁵. We generated Individual Taxpayer Identification Numbers (ITINs) for simulated 1040 filings by simulants without an SSN using a similar process²⁶.

We based our simulated employer names on a database of 5,321,506 “location names” from the SafeGraph “Core Places of Interest USA” dataset released in June 2020²⁷. To create a representation of bigrams from this dataset, we constructed a directed multigraph. Each word in a location name was treated as a node, and we included special <start> and <end> nodes. We included a directed multi-edge for each occurrence of a word pair in sequence in each location name. To generate simulated employer names, we performed a random walk through the bigram graph. Starting from the <start> node, we traversed directed edges selected uniformly at random until we reached the <end> node or exceeded a predetermined maximum path length. We then combined the words associated with each node that was encountered along the path to form the simulated employer name. This approach resulted in a diverse range of names that maintained a realistic quality. In the sample data included with the pseudopeople package, the W2 and 1099 employer names in 2020 include 212 distinct names and the three most commons are “San Benito Martinez Landscape Supply”, “Tony's Family Practice Inc”, and “Pikes Creek Campground”.

Dataset limitations

There are a variety of limitations to our simulation strategy which may affect its ability to reflect real-world dynamics, including but not limited to those regarding migration, employment, physical or mailing address, guardianship, household structure and simulant relationships, and simulant identification.

For instance, to make possible the simulation of complex migration dynamics, there are a series of assumptions we made regarding how simulants and households move around, into, and out of the simulation. We assumed that domestic and international migration do not change over time, but rather remain at the average rate from 2016 to 2020 in each future year of the simulation. When a simulant moves, we assume that their mailing address, physical address, and employer all change. In addition, complexities particular to household sub-structure interacting with migration are largely not captured in our simulation. For instance, a child can move out of their household without a parent, or a simulant could move without their spouse into a different household. Additionally, we assume that relationship does not affect emigration rates and that all household types are equally likely to have a simulant move out of or into them. Furthermore, for any individual migration of a simulant from one household into another, we assign the relationship “other nonrelative” in their new household. Thus, as time passes within the simulation, the proportion of households with irregular relationship structures grows. In the sample data included with the pseudopeople package, the 2020 decennial census has 4% of rows have “Other nonrelative” as relationship to the reference person, and in 2030 this rises to 16%. Even in the early years of our simulation, it is possible that there are rare, but challenging, household structures which are not sampled in ACS and therefore not represented in our data either, for example a very small fraction of very large households might present a problem in real-work linkage work that would not be identified when testing with pseudopeople data.

Similarly, there are several assumptions that we made to simplify our model of employment dynamics in our simulation. We do not model retirement, and each simulant can only have one employer at a time. There is a myriad of business dynamics that we currently do not model, including new businesses opening, existing businesses closing, business name changes, or business mergers and acquisitions. As with household physical and mailing addresses, when a business address is vacated, it is not reused. In effect, this likely makes business record linkage with these data easier than it will be in practice.

Our age-specific fertility and mortality models do not account for variations related to income or race/ethnicity, and in future iterations of this work, we wish to address more complicated dynamics between the various elements of our simulation.

There are also limitations in simulant identification because of privacy protections in the name data we have used. The data on first names excludes names with fewer than five occurrences while the data on last names included only names with at least 100 occurrences. Furthermore, we did not model the correlation between first and last names explicitly. We hope to address these limitations in future refinements of our model.