80th Annual Idea Groups

AAPOR Idea Groups

After a popular first year, we are pleased to offer Idea Groups at the Annual Conference again! This innovative pre-conference program offers a unique opportunity for participants to engage in intimate, focused networking sessions through guided discussions.

Taking place from 2:00 pm – 5:00 pm on Tuesday, just before the official start of the conference on Wednesday, Idea Groups provide a platform for attendees to convene in smaller, more specialized gatherings. These gatherings aim to foster dynamic conversations around shared topics and questions that resonate within our AAPOR community.

The primary objective of Idea Groups is to create a conducive environment for informal exchanges, enabling members to delve deeper into pertinent issues and explore collaborative solutions. Whether seeking insights, sharing experiences, or simply connecting with like-minded peers, Idea Groups offer a space tailored to your professional interests.

We encourage all attendees to take advantage of this valuable opportunity to enhance your conference experience and forge meaningful connections within our vibrant AAPOR community. Spots are limited and registration will be opening soon!

If you have any questions, please email Ryan Green, Education Manager.

 

Register for an Idea Group

AI and Other Advancements in Qualitative Research: Considerations for QUALPOR and AAPOR

Organizer: Darby Steiger

Qualitative research methods have continued to rapidly evolve since the explosion of generative AI over the past few years. In last year’s inaugural Idea Group on the Future of Qualitative Research at the 79 Annual AAPOR Conference, a group of 21 researchers across diverse sectors convened at AAPOR to discuss the ways in which AI tools have the potential to help us become more efficient in all phases of qualitative research. We also discussed our concerns about bias and representation, ethics and consent, and security risks related to using AI tools that may not align with our needs and obligations as social science researchers. The outcome of this session was a collaboratively-created list of over 40 specific ways in which generative AI could support the qualitative research process, along with 11 key areas of concern to keep in mind.

There are three main questions that will be posed in this proposed follow-up to last year’s Idea Group:

  • The first question is to ask AAPOR researchers to share learnings on how our various organizations have been considering, testing, or implementing AI tools since we last met and any challenges or roadblocks we have experienced.
  • A second question is to ask about any other advancements we have been making in the way we design, conduct, analyze and report on qualitative research that are not related to AI.
  • Finally, the third question is to ask and explore considerations and implications of these new methodologies for QUALPOR and AAPOR. Specifically, we will explore the potential for QUALPOR to create a set of guidelines, best practices, or standards for the use of AI in qualitative research, and what those could be.

Teaching America’s Youth about Public Opinion Polling: What Is AAPOR’s Role in (Re)Building Democracy/Democratic Institutions through K-12 Education?

Organizer: Robyn Rapoport & Lance Holbert

The proposed idea group will consist of representatives from the AAPOR community (including those involved in both conducting public opinion research and public opinion scholarship) and K-12 education to discuss what is needed to design, disseminate, and deliver effective classroom materials to educate America’s youth about the public opinion polling process and the role of public opinion research within a democracy. Building on Jamieson et al.’s (2023) call for the creation of K-12 education materials as part of a broader effort to protect the integrity of survey research, the University of Pennsylvania’s Leonore Annenberg Institute for Civics, in coordination with industry partner SSRS, is convening experts in the polling and education sectors to consider the following questions:

  1. how do we define the democratic function of polling,
  2. how can the polling process be deconstructed into digested segments from which to build lesson plans,
  3. which polling topics are best situated in civics versus STEM course offerings,
  4. how can these materials be seamlessly integrated into existing courses,
  5. what specific education goals can be met by these educational products,
  6. how do we educate teachers about polling and supplying teacher training for the proposed materials and
  7. how can these efforts be evaluated.

Goals for the idea group also include identifying how to facilitate meaningful collaboration between AAPOR and K-12 education leaders, securing a baseline set of commitments to continue work on the project, and mapping out a work calendar for the next 12 months. Next steps may also involve leveraging learnings from the course for journalists developed through the AAPOR Education Committee.

Reference: Jamieson, K. H., Lupia, A., Amaya, A., Brady, H. E., Bautista, R., Clinton, J. D., … & McNutt, M. K. (2023). Protecting the integrity of survey research. PNAS nexus, 2(3), pgad049. https://doi.org/10.1093/pnasnexus/pgad049

Using Multiple Data Sources for AI Alignment: Bridging Survey Research and Machine Learning

Organizer: Frauke Kreuter, Richard Zemel, Shafi Gldwasser, Ramya Vinayak, Daniel Oberski, and Stephanie Eckman

AI and machine learning (ML) researchers are constantly seeking high-quality benchmark data to evaluate and refine their models. For specific skills, such benchmarks often stem from standardized tests or custom-built datasets. However, when it comes to attitudes, behaviors, and living conditions, such benchmarks are scarce. Few survey datasets have been adapted for use in the ML/AI community. One notable exception is the American Community Survey, which has been made available as the “Folktables” dataset. This idea group aims to bring together survey researchers and computer scientists to explore how existing and future human data can enhance the development and alignment of large language models (LLMs). These data could serve not only as benchmarks but also as supplementary sources in the response generation process. The panel will focus on five key questions to ignite interdisciplinary dialogue:

  • Benchmarking: Which role can upcoming surveys place in holding out new data, based on events since LLM models were trained. Doing so will help test a model’s true generalization, and also see if a model can continuously be updated over time.
  • Modeling Success with Biased Data: Given that only a handful of surveys or human data sources are of exceptionally high quality—offering strong representation and measurement—how much can an LLM effectively learn from biased datasets to approximate outcomes similar to high-quality data? What are the risks and trade-offs?
  • The Role of Sparse, High-Quality Data: How much sparse, high-quality data is needed for successful model fine-tuning or re-training? This mirrors the challenges faced in survey methodology, where there is a growing demand for robust benchmarks to calibrate adjustments. What lessons can the two fields share? Likewise holding out data in the benchmark for evaluation specific sub-groups, particularly those that have relatively few survey responses.
  • Trend Modeling with Minimal External Data: How can LLMs be guided to model trends accurately with limited external data? This question opens up opportunities to discuss parallels with small area estimation techniques in survey research, offering a fertile ground for cross-pollination of ideas.
  • Preparing Data for Benchmarking and Alignment: What steps are needed to transform existing and forthcoming datasets into effective benchmarks? How can we ensure these benchmarks are useful for improving model alignment and performance across diverse applications?

This panel — featuring experts from machine learning, theoretical computer science, statistics, psychometrics, and survey methodology — promises to spark a vibrant discussion at the intersection of survey research and machine learning. By tackling these pressing questions, the group aims to lay the groundwork for more robust, human-aligned AI systems, leveraging the untapped potential of diverse data sources.

Preparing Survey Cost Estimates – Strengths and Shortcomings

Organizers: James Wagner

Costs are a consideration in virtually every survey design decision. Although we have the total survey error perspective for discussing errors, we lack a common framework for discussing costs and their impact on design and subsequent potential errors. At the 2024 AAPOR conference, an Idea Group met to discuss survey costs. The discussion was wide-ranging and informative on multiple topics. We would like to build on that session at this year’s conference with an Idea Group to focus on the development and tracking of cost estimates.
Survey cost estimates can drive the design of a study, yet are difficult to prepare for a number of reasons, including variation in response rates, unanticipated complexity in questionnaire design or administration, unplanned increases in pay rates, postage, or materials, and many other potential sources of uncertainty. Inaccurate cost estimates can result in survey designs that are infeasible or projects that cannot meet the original objectives. How different organizations derive these estimates, manage risk and uncertainty, and track expenditures vs. budgets is an underexplored topic. Survey organizations overall would benefit from best practices and shared language around survey cost estimates, including developing budgets and regular evaluation of inputs to components of a design.
Further, differentiated costs of efforts to recruit early and late respondents, including design features that may be particularly attractive to these groups, are rarely sought after and are underused by survey designers – even though these are a critical part of adaptive and responsive designs and may be the source of cost overruns to meet a target sample size. We hope to discuss several questions related to issues with preparation of cost estimates: What are the facilitators and barriers to developing accurate cost estimates?
  • Including facilitators such as data on prior studies, vendor quotes, predictive models, and barriers such as inflation/price changes, misspecified requirements, and inaccurate specifications.
  • To what extent are formal cost evaluations conducted, including comparing cost estimates to actual costs, allocation of budgets to fixed and variable costs, identifying sources of error in estimates, or exploring variation in costs over different types of studies?
  • Which costs are most difficult to predict? Is this due to bias or variance? Is this due to lack of desire to have this information or communication, inability to track at different levels of the budget, perceived need for confidentiality even within an organization, or other reasons?
  • Which costs are measured with the least accuracy?
  • For which survey components could improved accuracy lead to the greatest improvements in the accuracy of overall cost estimates?
  • Which areas of survey design specifications that are impactful to costs appear to be least/most accurate? (e.g., clients/users routinely underestimate survey length) What areas are changing the most over time and thus require budget updates more frequently (e.g., return to response rates on modest incentive levels)?
  • What tools are needed to improve cost estimates?
  • More analysis of existing data?
  • Better tracking of components of costs?
  • To what extent and in what ways are organizations using cost models or statistical models to write cost estimates or predict changes in costs during the field period?