Memo: Multiply-imputed vs complete-case head-to-head

Research memorandum — does MI restore demos/CA or attenuate Q28’s lead?

Author

Jack J. Burleson

Published

Invalid Date

Research question: Would a multiply-imputed analysis restore demographics/CA while preserving the Q28 lead, or would imputation attenuate rideshare’s advantage?

Parent: common_n_head_to_head.qmd · Companions: q28_conditioned_on_car.qmd · demographics_predict_transit.qmd · ca_scores_predict_transit.qmd

1 Answer, Response, + Summary of Results

We compare each predictor family under listwise complete-case versus \(M = 20\) multivariate imputations of the missing-prone items (Q20, Q21, Student status), using chained equations (IterativeImputer, sample_posterior=True (Pedregosa et al., 2011)) with complete auxiliaries (Age, Sex, CA (McCroskey, 1970), Q28, employment, country, lat/long). Outcome-labeled base n = 241 (seed 42, stratified CV). Q28 and CA are never imputed—they appear in both arms as observed.

Short answer: Q28’s singleton lead is preserved under MI (AUC ≈ 0.762). Demographics and CA are not restored to competitive levels (≈ 0.617 and 0.590); imputation does not attenuate rideshare’s advantage among single-family models. Joint models that include Q28 remain strongest (CA+Q28+car MI ≈ 0.808; Figure Figure 1).

Figure 1.. Complete-case vs MI AUCs

1.1 Complete-case vs multiply-imputed ranking (n_full = 241)

Table Table 1 reports complete-case and multiply-imputed AUCs for each feature family.

Table 1.. Multiple imputation vs complete-case comparison.
Feature family CC n CC AUC MI n MI AUC Δ (MI − CC)
CA + Q28 + car (joint) 149 0.736 241 0.808 +0.072
Demos + Q28 (joint) 224 0.773 241 0.771 −0.002
Q28 days 241 0.762 241 0.762 0.000
Age + Sex + Student 224 0.618 241 0.617 −0.001
Group + interpersonal CA 241 0.590 241 0.590 0.000
Car license & access 149 0.607 241 0.562 −0.045

Singleton MI order: Q28 ≫ demographics ≫ CA ≳ car.

1.2 Interpretation

  1. Rideshare advantage is not an artifact of complete-case N. On the full analytic frame with MI-filled car/student items, Q28 still outranks demographics, CA, and car as a singleton family—the same ordinal conclusion as the common-N memo.
  2. MI does not “restore” demos/CA. Expanding Student-missing rows to n = 241 leaves demographics ≈ 0.617 (flat vs their own CC). CA is unchanged because it was already complete. Neither approaches Q28.
  3. Imputation mainly helps joint mobility models. Filling Q20/Q21 lets CA+Q28+car train on all 241 rows and lifts AUC (0.736 → 0.808). Standalone car weakens under MI (0.607 → 0.562)—imputed car is a noisier cue than car observed on the selected complete-case subset.
  4. Joint demos+Q28 stays high (~0.77) but still tracks Q28 rather than a demography rebound.

Sources: ca-personas followup-experiments --experiments mi_head_to_head --n-imputations 20 · outputs/followup_experiments/mi_head_to_head/ · Posit full-cohort artifacts (n = 241) · formal write-up docs/secondary_rq_followup_experiments.md

2 What questions or uncertainties remain?

Would a richer imputation model (e.g. including interaction terms, or imputing under a MAR vs MNAR sensitivity design) change the car attenuation pattern? Would ordinal / multilevel MI of Q28 strata alter residual-CA contrasts within ride-share days?

References

McCroskey, J. C. (1970). Measures of communication-bound anxiety. Speech Monographs, 37(4), 269–277. https://doi.org/10.1080/03637757009375677
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., & Duchesnay, É. (2011). Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12, 2825–2830. https://www.jmlr.org/papers/volume12/pedregosa11a/pedregosa11a.pdf