The U.S. Foster Care System: Parent-Child Matching


Abigail Lindner

Abigail Lindner
Worcester Polytechnic Institute


In the Fall/Winter 2022 issue of OR/MS Tomorrow, the article titled “The U.S. Foster Care System: Resource Allocation” discussed how operations research practitioners could play an important role in facilitating the distribution of finite resources among children in foster care and their respective families. (For those who have not perused the initial article of this two-part series, it is readily accessible here.)
In ”Resource Allocation”, we highlighted how the U.S. foster care system primarily exists to provide protection and support for children who have, due to an array of circumstances, come under its care; and to eventually settle these children in permanent placements. The projects detailed in the article demonstrated how the efficient allocation of resources could accomplish both these objectives, by ensuring necessary services are dispensed to bolster the well-being of children, and by reducing the duration of their stay in foster care.

Children and youth in the United States typically exit foster care through one of three main ways: reunification with their families of origin, adoption, or emancipation, also called “aging out”. Adoption ranks as the second most prevalent exit strategy after reunification. Children who have permanency plans for adoption are matched with prospective adoptive families in a process that, over the past few decades, has been followed with the assistance of modern algorithmic technologies. 

Parent-Child Matching

According to the Kids Count Data Center of the Annie E. Casey Foundation, a charitable organization in the United States dedicated to the well-being of children, over 117,000 children were awaiting adoption in the year 2020 (Center, 2021b). The most substantial age group in need of placement comprised children aged 1 to 5, followed by those aged 6 to 10 and aged 11 to 15. The criteria for determining a “fit” between a child with a permanency plan for adoption and a prospective adoptive family has seen various shifts over the past century. These evolving metrics have included parental ability based on IQ, educational attainment, and/or social class; racial or cultural alignment between the child and the prospective family1; the prospective family’s ability to accommodate certain special needs; as well as parenting and attachment styles (Haysom et al., 2020).

Another consequential factor in matching children with families is the family preferences of the child (for example, if they want a family with both a mother and a father or with no siblings)  and the child preferences of the family (for example, if they want a boy aged between 5 and 11). Algorithms designed to expedite and simplify the matching process are usually geared toward optimizing the match between the characteristics and background of the child and the characteristics and preferences of the prospective family. 

Adoptive Placement Matching and Online Dating 

As far as process goes, parent-child matching algorithms fall into two categories: family-driven and caseworker-driven. In the family-driven model, prospective families are notified via email about children who have been released for adoption, and they respond if they believe that a specific child could be a fitting match. In a caseworker-driven model, which involves more advanced software tools, caseworkers for children who have been released for adoption browse all available, registered prospective parents.

Olberg et al. (2021) drew comparisons between parent-child matching in the foster care system and online dating; both operate as two-sided markets without a centralized clearing- house, relying on purely randomized, decentralized matching models Olberg et al. (2021). An essential distinction, however, is that parent-child matching possesses a centralized protocol and middlemen, the caseworkers. Despite this, approaching parent-child matching from a two-sided market perspective proves useful; viewing the process through an economic lens enables child welfare researchers to apply a game theory framework to their analysis, with the objective of determining which type of algorithmic model best serves children and families.

Both family-driven and caseworker-driven searches are inherently dynamic, involving uncertainties about compatibility and the heterogeneous preferences of agents. Based on their research about the impact of each approach on average utility in equilibrium, Olberg et al. (2021) concluded that caseworker-driven searches are often more preferable. Preference for caseworker-driven searches over family-driven searches arises because the former, by avoiding wasted search efforts, can Pareto-dominate the latter in equilibrium, while the latter can never Pareto-dominate the former. This implies that caseworker-driven searches either outperform or are on par with family-driven searches in every dimension of inter- est (Dreeves, 2019).

Improving Caseworker-Driven Searches

An example of a caseworker-driven search method can be found in the work of Slaugh et al. (2016). The Pennsylvania Adoption Exchange (PAE), established in 1979, aims to support the adoptive placement of hard-to-place children by drawing from comprehensive data about the children and the preferences of prospective families. The tool in use at the time of their writing contained seventy-eight pairs of child-attribute values and

family preferences. A compatibility score, ranging from 0 to 100, assessed the fit between a child with a permanency plan of adoption and a prospective family being considered for the placement.

Despite its promise, the PAE was regarded as ineffective by the caseworkers who used it, as they did not trust the match recommendations generated by the system. Part of the issue was linked to implementation challenges: child welfare agencies faced difficulties coordinating algorithm design with an IT contractor and managing temporal data collection across Pennsylvania’s sixty-seven counties.

Operations researchers took up the challenge to devise a more efficient system. Expanding on the PAE concept, Slaugh et al. (2016) developed a spreadsheet-based algorithm (Slaugh et al., 2016). This model included flexible weights and geographic constraints, enabling child welfare professionals to determine on a case-by-case basis which parent-child factors were more or less pressing to match and how far to look. The overall result was a computerized matching tool that ranked families accord- ing to compatibility with a given child. It was based on a discrete-event simulation of the adoption network.

The researchers conducted simulations of the new spreadsheet-based algorithm using real data from 1,853 PAE-registered children and 2,194 prospective adoptive families. These simulations revealed that the more information “matchmakers” had on families’ preferences, the better the algorithm performed, increasing the adoption rate and reducing the average number of match attempts until a successful adoption. In essence, having more information about prospective families led to more successful matches and fewer failed ones.

  • The simulations also revealed that broadening the search area increased the adoption rate. Given this, it is even more critical for caseworkers to have reliable algorithmic tools to assist their decision-making. Assessing the files of all the families in one county is already a substantial task; imagine the workload when expanding the scope to five counties, or ten, or all sixty-seven across Pennsylvania!

Remember the Children (and Other Humans)!

The primary takeaway from Slaugh et al. (2016) is the crucial role that quality of family information plays in improving the performance of matching algorithms and consequently, achieving more adoptive placements for children in foster care. Meanwhile, Saxena et al. (2020), after conducting an extensive literature review of fifty child welfare system decision-making algorithms, asserts that these algorithms require, alongside high-quality data, a basis in more human-centered design (Saxena et al., 2020). 

As the field of Artificial Intelligence has evolved, human-centric approaches to algorithm design have gained prominence. The principle is that algorithms should “reflect realistic conceptions of user needs and human psychology” (Guszcza, 2018). In their study, Saxena et al. (2020) found that many of the algorithms ignored system factors that influenced caseworkers’ decisions, such as current policies and scarce resources. The majority focused on minimizing the risk of future maltreatment for children in foster care. While undoubtedly important, this focus does not account for the overall improvement in life quality, a primary objective of caseworkers that is relatively challenging to quantify.

To bring the human element into these algorithm-influenced decisions, active engagement with domain experts, children in foster care, and parents of children in foster care is critical. Inputs from these stakeholders would expand understanding of specific values, needs, cultural and parental expectations, and socio-technical challenges that impact parent-child matching decisions.

Saxena et al. (2020) explain, “Child-welfare workers are generally not trained in statistical thinking and make decisions based on experience, intuition, and individual heuristics.” Therefore, applying a human-centered theoretical approach to algorithm design would place the meaning-making process at the center of the process, contextualizing how child welfare workers perceive quantified metrics.

Said otherwise, human-centered approaches would better ac- count for how child welfare agents navigate the foster care-to-adoption process and prospective family relationships, as well as their interaction with the numerical data that is the bread and butter of analytics. Operations researchers and management professionals have the opportunity to refine algorithmic processes to better reflect these human realities.

Conclusion & Future Research Directions

While a slim majority of children in foster care reunite with their families - between 47% and 52% from 2011 through 2020 (Center, 2021a)- reunification is not a viable option for every child. When reunification is not viable or advisable, the second most common exit type from foster care is adoption, accounting for between one-fifth and one-fourth of exits from foster care annually. Over the past few decades, algorithmic tools, as studied by Olberg et al. (2021) and Slaugh et al. (2016), have helped caseworkers “optimally” match children with permanency plans of adoption to prospective adoptive families.

While operations research (OR) cannot eliminate the necessity for foster care, it has the potential to alleviate some of the associated stress by improving data-based processes through which caseworkers and other child welfare workers make decisions for the children under their responsibility. OR practitioners bring to the table a wealth of experience in integrating quantitative and qualitative data into models, experimenting with tools to expand data capabilities, and contemplating the human-centered design of algorithms and AI, the concern for child welfare re- searchers like Saxena et al. (2020). This knowledge could have a substantial impact on the lives of these children and families.

In 1970, Arthur Spindler, who was affiliated with the Department of Health, Education, and Welfare (now the Department of Health and Human Services), underlined the potential of OR in child welfare, stating “The need is there. The challenge is now,” (Spindler, 1970). More than five decades later, in 2023, the need persists, and the challenge remains as pertinent as ever. 

Since the enactment of the Multiethnic Placement Act in 1994, race, color, or national origin, whether of the parent(s) or the child, can no longer serve as reasons for delaying foster care or adoptive placements in the United States. 



Center, K.C.D., 2021a. Children exiting foster care by exit reason. URL: children-exiting-foster-care-by-exit-reason. accessed: June 10, 2023.

Center, K.C.D., 2021b. Children in foster care waiting for adoption by age group in the united states. URL: https://datacenter. accessed: June 10, 2023.

Dreeves, D., 2019. The pareto dominance principle for apps and websites. URL:

Guszcza, J., 2018. Ai needs human-centered design. URL:

Haysom, Z., McKibbin, G., Shlonsky, A., Hamilton, B., 2020. Changing considerations of matching foster carers and children: A scoping review of the research and evidence. Children and Youth Services Review 118. doi:10.1016/j.childyouth.2020. 105409.

Olberg, N., Dierks, L., Seuken, S., Slaugh, V., Ünver, M., 2021. Search and matching for adoption from foster care. URL:

Saxena, D., Badillo-Urquiola, K., Wisniewski, P., Guha, S., 2020. A human-centered review of the algorithms used within the u.s. child welfare system, in: CHI 2020, Honolulu, HI, USA. URL:

Slaugh, V., Akan, M., Kesten, O., Ünver, M., 2016. The pennsylvania adoption exchange improves its matching process. Interfaces 46, 133–153. URL:

Spindler, A., 1970. Social and rehabilitation services: A challenge to operations research. Operations Research 18, 1112–1124.doi:10.1287/opre.18.6.1112.