Adaptive questionnaires for polarized survey questions

Which question is worth asking next depends on the answers already given, so a questionnaire has to be designed as a sequence rather than as a set.

Schematic. A few candidates are selected from a larger pool, and the order in which they are asked is part of the decision. No data is shown.
Status
Completed
Period
2021 — 2022
Themes
Fair & reliable predictive modeling

Research question

How should a questionnaire choose both the subset and the order of its questions so that each answer carries as much information as possible?

Why it matters

A survey has a budget. Every additional question costs attention, and on socially polarized topics it can cost willingness to answer at all. That makes the choice of which questions to ask, and when, a design decision rather than an administrative one.

The challenge

Choosing the most informative subset of questions is already combinatorial. Choosing the order makes it sequential, because what is worth asking next depends on what has been answered so far. The two parts cannot be solved well in isolation, and the respondent’s remaining answers are exactly the quantities that are unknown at the moment the next question has to be chosen.

Approach

The project modeled the task as an online variant of matrix completion, and solved it with an ensemble of predictive models, probabilistic greedy policies and dynamic programming for the sequential decisions.

Status

Completed research at the Complex Engineering Systems Institute, 2021 — 2022, with Ricardo Montoya and Charles Thraves. There are no public outputs from this work.

All research