University of Maryland

Data Literacy & Evidence Building

NYU Wagner | Accenture | University Of Maryland | KYStats | Coleridge Initiative


“Data Science emerged from three fields.. Statistics is the mathematical field that interprets and presents numerical data, making inferences and describing properties of the data.. Operations research, .. focused on understanding systems and taking optimal actions in the real world..(and) Computing ..the design, development, and deployment of software and hardware to manage data and complete tasks” Data Science in Context p 22. (1)


  1. Understand the role of inference in evidence-based policy
  2. Learn the importance of understanding the data generating process in inference
  3. Review the sources of missingness for sub-groups

Inference and evidence-based policy

From data in the world, we build a model of some aspects of it, reason about the model to draw conclusions, and check that these conclusions correspond to what happens in the world. The better the model, the better the correspondence between the model’s conclusions and the real world. Dashed arrows denote the mapping between world and model, and solid arrows are within the world or model.                                                                                                                     Figure 1.1 Models and the World                                                                                                                     Source: Data Science in Context

This class is intended to train participants in the basics of using administrative data to generate evidence. There is not enough time to go into the details of developing a causal model, although there are some excellent books that should form part of an analyst’s reading list(1-6). Most of the focus has been on making sure that what is reported is “close” to the conceptual measure of interest. However, this section is provides a primer on, and further pointers to, the importance of understanding the data generating process in making the causal statements that are necessary to make good policy. In terms of the scoping section at the beginning of the class, the data and evidence have to be in place before Step VIII – Action – is contemplated. Figure 1 provides a nice summary- we observe some data, develop a model, draw conclusions, and ideally generalize those conclusions to the population of interest to inform that action.  But there are many ways in which errors can be made.   As Xiao-Li Meng has pointed out, total error is the combination of defects in data quality, data scarcity and the inherent difficulty of the problem!(7). Creating high quality evidence about education to workforce transitions so that decisions can be made requires the same inferential validity as is required in almost all social science. Do students who graduate with major x get better earnings than those who graduate with major y? Would students who graduate in major x have got better earnings if they had graduated in major y just because they are better students? (adapted from Morgan and Winship (2)). The focus in this class has been on getting the answer to the first question as close to right as possible by carefully defining terms – what does it mean to be a  student; to graduate; to have a major; to have earnings – and the time period that those terms cover.  Getting the data right is necessary but not sufficient to answer the second question. READ MORE

Inference and the data generating process

The structure of the Evidence Act shows the importance  that Congress placed on inference in the process of generating evidence. The first Title set up the organizational infrastructure – the people who would do the work, like Chief Data Officers, Chief Evaluation Officers, and Statistical Officials. The second Title established the importance of data access. Title III highlighted the importance of statistical agencies in producing evidence. Data literacy, in other words, is not just a matter of operations and coding.  Statistical thinking is required, and that means an understanding of the data generating process and how data came to be.  Errors can show up in many ways. Selection bias is particularly prevalent in the data generating process in education to workforce transitions. Clearly students (and their parents) will take possible future earnings into consideration when they choose a major. As a result, we only observe earnings data on those who have self selected into the major in which they think they are going to get the highest earnings.  One reason for Jim Heckman’s Nobel Prize in 2000 was the development of a technical approach correcting for selection bias – which was particularly helpful in understanding why historical analysis of the link between women’s education and earnings outcomes might be biased due to women’s higher reservation wage. Another common challenge is in data processing prior to analysis. In the Census Bureau’s Post Secondary Employment Outcomes data series, for example, outcome measures are based on the calendar year post graduation (which will commonly blend two academic years) and the earnings measures impose two labor market attachment restrictions (an earnings threshold and a number of quarters worked) that are likely to bias estimates upward for marginally attached workers. It pays to be very careful not just with your own decisions, but to examine decisions that others have made. READ MORE

Missing data for population sub-groups

In many ways, good inference requires that the analyst thinks about the data generating process not just in aggregate but also for sub groups. Take a look, for example, at the earnings and employment outcomes for students graduating from Northern Kentucky University in the MSPSR using just Kentucky data. Compare that to the results if Ohio data are included.  Missing data, in other words, can bias the results, and thinking about the missingness mechanisms is critical. It’s particularly important with administrative data, where changes in the rules for administering government programs can include or exclude particular groups for arbitrary reasons that are not related to the underlying data generating process. Failure to adjust for local conditions is another way in which geographically and demographically based errors can creep in. This is particularly true with data collection shocks like COVID, where survey-based measures were generating incorrect information to governors about their local labor markets. From January-September 2021, for example, the unemployment rate reported for the state of Michigan was unrealistically low both because of data errors in Detroit and because of an incorrect statistical correction by the federal government. As a result, the Michigan October rate, for example, had to be revised from 4.6 percent to 6.3 percent. Because unemployment is not directly measured at the local level, but rather estimated for an entire region and then allocated to other states, the revision affected other major states – Illinois, Indiana, Ohio, and Wisconsin. The Illinois governor was making decisions based on wrong data. Illinois’ unemployment was much lower than was reported. As the November 18, 2021 Illinois press release pointed out, the errors were identified when “Illinois and another East North Central Division state raised concerns about their monthly 2021 statewide labor force estimates”. This is one of many reasons that data access (which is also an important thread in the Evidence Act) is so important, and why data science is a team sport.  Every data science team should include someone who understands how the program is administered, and should have some input from people who can be affected by the decisions that are made (Step 1 in the Scoping process). READ MORE



  1. Spector AZ, Norvig P, Wiggins C, Wing JM. Data science in context: Foundations, challenges, opportunities. 2022.
  2. Morgan SL, Winship C. Counterfactuals and causal inference: Cambridge University Press; 2015.
  3. Mealli F. Answering causal questions: Angrist, Imbens and the Nobel prize. Significance. 2021;18(6):4-5.
  4. Imbens GW, Rubin DB. Causal inference in statistics, social, and biomedical sciences: Cambridge University Press; 2015.
  5. Pearl J. Causality: Cambridge university press; 2009.
  6. Pearl J. Causal inference in statistics: An overview. 2009.
  7. Meng X-L. Statistical paradises and paradoxes in big data (i) law of large populations, big data paradox, and the 2016 us presidential election. The Annals of Applied Statistics. 2018;12(2):685-726.