Follow-up from presentation

[marcdotson]

Mansky’s law of diminishing credibility?

[marcdotson]

• Current ensembles either induce randomization on part-worth estimates or induce randomization on the number of attributes included.
• Need for clever randomization in the input data based on specific data pathologies.
• Discuss using aggregate logit or hierarchical multinomial logit for the ensemble.
• Discuss the form of the meta-learner.

[marcdotson]

• Move to an ensemble of hierarchical MNL.
• Induce randomization on the training data WRT attribute non-attendance and screening.
• Consider whether or not to reference Mansky's law of diminishing credibility.
• @jeff-dotson to think about the form of the meta-learner (may be related to inducing randomization on the training data at the respondent-level).

[marcdotson]

• Move to an ensemble of hierarchical MNL. (Moved to #38.)
• Induce randomization on the training data WRT attribute non-attendance and screening. (Moved to #40.)
• Consider whether or not to reference Mansky's law of diminishing credibility.
• @jeff-dotson to think about the form of the meta-learner (may be related to inducing randomization on the training data at the respondent-level).

[marcdotson]

Superseded by current work.