8

Issues and Innovations in Population Modeling and Projections

MODELING AND SIMULATION FOR FORCED MIGRATION RESEARCH

Erika Frydenlund (Old Dominion University) explained that modeling and simulation require looking at problems from a dynamic systems perspective. Human systems are complex, adaptive systems that change continually. In social systems, people can make choices and be influenced by outside policies or by changes in the environment; these are systems that adapt and grow.

Modeling such a system requires converting relationships, variables or factors, and even agents into computer code. Running the computer code in a simulation then extracts simulated data based on a set of parameter values. Frydenlund said that unlike traditional tools, which result in static pictures that reveal little about the dynamics of a system, simulations relate dynamic variables to one another. Ethnography is a rich contextual form of research that may not translate to another context, but simulation models might help determine whether the ethnographic findings hold with a wide range of ideas and theories, she said.

Frydenlund explained that simulations are powerful partly because they require specificity and precision in the use of terms. When ethnographers use terms such as “solidary” or “citizen initiative,” the terms need to be carefully operationalized to use in a simulation, said Frydenlund. Modeling and simulation thus offer an opportunity to create a visual description of what one is studying—the phenomenon—and unpack some of these terms.

Prediction is the most obvious reason to use modeling and simulation, Frydenlund observed, but it is not the most powerful. Rather, she emphasized four other uses for modeling and simulation. First, they could facilitate discussion and understanding of solutions, perspectives, and scenarios. By having a visual artifact of all the research inputs into a problem, discussions can be grounded in common assumptions and terms. Second, they could help explore solutions, perspectives, and scenarios. Rather than predicting outcomes, models can give decision makers an idea about what the boundaries of an outcome might be. Third, they can be used to structure thoughts and discussions. By structuring thoughts and discussions around topics, models can bridge disciplinary divides. Fourth, modeling and simulation can communicate assumptions, processes, and visual analogs. Models can be visual artefacts for the systems people observe.

As an example of the fourth use, Frydenlund described her work with refugees on the Greek island of Lesbos, where stakeholders on the north and south sides of the island have different ideas of their roles. Modeling has helped stakeholders to see the bigger picture and to focus the conversation.

Modeling and simulation have evolved to the point where they can integrate many types of data and expertise. These include experiential data and narratives from fieldwork, theories and research from academia, material from documents, and input from vetted organizations. For example, workers for nongovernmental organizations (NGOs), including volunteers, know a lot about system dynamics, such as who interacts with whom, which can be used to build simulation models.

Frydenlund described three main types of models (see Table 8-1). First, systems dynamics models take a macro view—for example, who is related to whom, how relationships work, or how money moves between organizations and gets dispersed. Second, discrete event models address procedural

questions, such as how a queuing system would work. Third, agent-based models simulate the interactions of people with different opinions and needs. Lesbos, for example, has many NGOs of different sizes, some of which are pursuing just a single task. Everyone wants to do something like donate food or pull people out of the water, said Frydenlund, while washing and recycling clothes is less popular. Agent-based models simulate the phenomena that result in various outcomes, she said.

Frydenlund described challenges to building simulation models. First, useful models need to be co-created in a collaborative process involving different inputs and expertise. She encouraged population scientists to get involved in model building, because “these models are dangerous without you.” Modeling software is usually in the hands of computer scientists and engineers, while the necessary knowledge is held by social scientists and stakeholders, including refugees and people in local communities, she said. In addition, computer models take time to understand and programming experience. As an example of how to overcome this problem, the York University Centre for Refugee Studies is working with the Virginia Modeling, Analysis, and Simulation Center at Old Dominion University to build protocols for aid workers, policy makers, and academics to communicate across disciplinary lines and facilitate the building of models.

The results of modeling can be so compelling that modelers are tempted to give the results immediately to policy makers, but Frydenlund urged caution. Models require context and understanding to know what theories, data, and other inputs went into them; attention should also be given to ethical issues about who owns models and how and by whom the scenarios/experimentation will be used and interpreted. Frydenlund also noted that

models can produce many different results, given the range of assumptions and data they can incorporate, and that modelers should make sure that their models stay useful for what they were intended for.

In response to questions raised during the general discussion about the need to make decisions despite the uncertainties of models, Frydenlund observed that the social sciences are still relatively young fields for modeling and simulation, and that validating social science models can be tricky. She said that models are a bit like art, and that no model represents the truth; every model can be improved upon. Agreed-upon methods for validating models do not exist, and the models are so complex and context-specific that the best way to validate them may be through other stakeholders based on their experiences, rather than through projections. Despite these uncertainties, Frydenlund continued, models have the power to ground conversations and facilitate dialogue because they help to reveal underlying assumptions and inherent biases in the data.

Frydenlund also emphasized the importance of teaching students to be systems thinkers. They need to be comfortable with holistic models that incorporate multilevel analyses and multiple relationships. Even though young people are growing up in a digital age, they are not yet ready to be systems thinkers, she said, because they have largely been trained in macrolevel political science theory, and it is still difficult for them to apply that theory at the meso- and microlevels. As capabilities increase and models become more complex, machine learning will be a way to tame complexity. However, said Frydenlund, it is important to know how to think, so that research guides the machines instead of the other way around.

NEW TOOLS FOR ESTIMATING, UNDERSTANDING, AND FORECASTING MIGRATION FLOWS

Raya Muttarak (University of East Anglia and International Institute for Applied Systems Analysis) described new tools that demographers are using to estimate, understand, and forecast migration. The first tool she described was the use of indirect estimation methods to derive bilateral migration flows to provide a more dynamic view of migration patterns. Stocks give the numbers of migrants, defined by their birthplace, who are living in a country at a point in time. Stocks are static, easy to define, and collected in censuses and available for all countries. They are useful in measuring how many people have migrated, but they do not reflect the dynamic pattern and trend of migration. Flows give movements between countries of origin and destination during a defined period. They are dynamic but are difficult to measure and compare across countries. The United Nations and Eurostat provide collections of bilateral flow data, but they are available only for some Western countries. Flows can be estimated using stock

data, Muttarak pointed out. By using stocks as the margins in flow tables, Abel (2018) indirectly estimated bilateral migration flows for all countries between 1960 and 2015 and offered an innovative and simplified way to visualize migration flows using a circularized plot. In another application of this visualization method, Abel and colleagues (2019) plotted asylum-seeking flows by world region. In the period 2006–2010, for example, many migrants went from Eastern Africa into Southern Africa, Muttarak pointed out, whereas in the period 2011–2015 the flows from Western Asia to Europe were more extensive.

Another tool Muttarak described is the use of nontraditional data sources to estimate migration flows. For example, Zagheni and Weber (2012) estimated age and gender-specific migration rates using data extracted from Yahoo! e-mail messages. The profiles of e-mail users provide age and sex, and location is available from where people log in to receive their e-mails. Migration can be tracked by capturing differences in a location where people log in over time. Though the percentage of people who use this e-mail service is small, Muttarak observed that the method can be effective in capturing migrant flows. This work also has been extended to other social media data, including messages on Twitter (Fiorio et al., 2017; Hawelka et al., 2014; Zagheni and Weber, 2012), LinkedIn profiles (State et al., 2013), and the Facebook advertising platform (Zagheni et al., 2017). One difficulty with this tool is the amount of data that must be analyzed—millions of records, for example, in the case of Twitter data. Also, people may have more than one account, many are nonregular users, and their use of a given platform changes over time. This represents a significant amount of data management that has to be dealt with, she commented.

These new data sources can be combined with traditional data to arrive at better conclusions, Muttarak observed. In addition, expert-based judgments on features of each data source (accuracy, coverage, and undercount) can be incorporated into the estimation method.

The third tool Muttarak discussed involves better understanding of the drivers of migration. A framework developed by Black and colleagues (2011) seeks to capture the complexity and interactions of these drivers. For example, environmental factors are usually considered indirect drivers in which environmental change may influence migration through affecting other factors underlying migration such as income or agriculture. Black’s model provides a conceptual framework that describes the direct and indirect pathways through which environmental change drives migration. For example, severe droughts can reduce crop yields, which might affect migration in countries with poor resource management. People might move from rural to urban areas, or from one country to another due to conflict arising from poor management of scarce resources as exemplified in the case of Syria, which experienced 4 years of intensive drought from 2007 to 2010.

Muttarak noted that Abel and colleagues (2019) provide an example of how Black’s framework can be modeled empirically. First, she explained, they estimated the probability of conflict given climate conditions, political conditions, economic conditions, and so on. In the second step, they estimated the extent to which climate-induced conflict affects the probability of sending out asylum seekers. In the final step, the number of people who migrate can be estimated. Though Abel and colleagues (2019) did find some evidence of climate conditions and conflict on migration, the results were quite context specific—there were stronger effects in Western Asia and Northern Africa compared to other regions, and only in the period between 2010 and 2012.

Finally, Muttarak proposed the application of existing demographic tools to estimate and predict forced migration, as has been done in her work on forecasting societies’ adaptive capacity. Empirical work can first determine how vulnerability to climate change varies by population characteristics. These results can be combined with the macrotheory of “demographic metabolism,” which is a process of social change through cohort replacement by which one cohort changes as it replaces another—for example, younger people replacing older people. Then, multidimensional population methods can yield multistate population projections.

As an example, Muttarak briefly described the process of adding demographic characteristics beyond age and sex, such as education attainment, religion, place of residence, or ethnicity, into a population pyramid. The application of this approach to the Republic of Korea yields projections of populations and education levels that in turn have effects on population size, population structure, and even such factors as disaster resilience, since work in 174 countries has shown that those with the highest proportion of women with at least secondary education tend to have lower disaster mortality (Striessnig and Lochinger, 2016).

An extension of this work is to forecast future vulnerability using scenarios from the Shared Socioeconomic Pathways (SSPs), which provide alternative narratives for different levels of socioeconomic development, population, and educational attainment. In SSP1 with rapid increases in life expectancy, faster fertility declines in high-fertility countries, and educational expansion, the predicted numbers of deaths from natural disasters are substantially lower than in SSP3, a pathway with stalled educational expansion and continued high fertility and mortality (Lutz et al., 2014). This approach can also be used to develop scenarios for population size and structure in Europe based on assumed migration rates. Muttarak added that although demographers normally shy away from making assumptions for future migration scenarios, work on translating socioeconomic development pathways into future migration is an active area of study (Abel, 2018).

In response to questions raised during the general discussion about how models will be used in the future as data (including “big data”) become increasingly plentiful, Muttarak suggested looking beyond regressions and using machine learning. To some extent, this will require involving younger generations of researchers who are more steeped in these techniques, she said. It also will pose issues in terms of checking the robustness of results, since data will often be context specific.

This page intentionally left blank.