Methods · Technical deep-dive · ← Back to project overview

How the study
was conducted.

A three-stage analytical pipeline — data integration, latent attitudinal measurement, and predictive modeling — applied to a linked dataset of 1,844 survey responses and ~1.4 million operational rentals.

Below: the design of the survey instrument, the statistical machinery, the seven latent attitudinal factors, and the quality controls.

PipelineEFA → GMM → MNL
Linked N1,743
Validation5-fold CV + RF
Pseudo-R²0.128
Methodological note

The linkage matters as much as the model.

Pairing each survey respondent with their actual ride history lets temperature variance, multimodal fallback patterns, and revealed weather behaviour stand in for self-reports — and outperform them as predictors of intensive use.

— POTEBS · methodology page
Methods

How the study was conducted.

Three stages connect raw data to inference: attitudinal measurement (EFA), data-driven segmentation (GMM), and predictive modeling (multinomial logit with random forest validation).

01
Stage 1 — Data integration

Linking survey and operational records.

Two parallel surveys were distributed to ~32,000 registered users. After quality screening, 1,844 responses were retained; 1,743 were linked at individual level to the full operational trip history — combining stated preferences with revealed behavior.

02
Stage 2 — Measurement & segmentation

EFA factors and GMM intensity clusters.

EFA reduced 32 Likert items to seven attitudinal constructs — including weather tolerance, environmental self-identity, and multimodal complementarity. GMM clustering then identified three data-driven usage segments (low, medium, high intensity) without imposing arbitrary thresholds.

03
Stage 3 — Predictive modeling

Multinomial logit with random forest validation.

Multinomial logit predicted segment membership from demographic, geographic, and attitudinal predictors via backward elimination. Results were validated through 5-fold cross-validation and random forest feature importance rankings, which broadly corroborated the MNL findings.

Key methodological advantage

Linking survey responses to trip histories enables triangulation between stated and revealed preferences. The key behavioral predictor — temperature variance across a user's ride history — captures demonstrated weather resilience independently of self-reports (r=0.05 with trip volume), confirming it as a genuine behavioral trait rather than a usage artifact.

Survey instrument

How we asked, what we asked.

Two parallel surveys (FFEBSS / DBEBSS), structurally identical, fielded in April 2025 in German, French, and English via Qualtrics. Six sections, ~80 items, with thirty-two Likert-scaled items reduced to seven latent attitudinal factors via exploratory factor analysis. Full instrument in Appendix A3 of the manuscript.

[01] Field period
Apr 2025
Fielded after the operators distributed email invitations
[02] Invitations
~32k
~25,000 FFEBSS · ~7,000 DBEBSS users contacted
[03] Responses
2,001
Total · 1,844 retained after quality screening
[04] Incentive
CHF 10
Voucher (Pick-e-Bike or Galaxus/Digitec)
Six-part structure
A.
Part 1 · Mobility context & EBSS use

How users move, and where EBSS fits.

Driver's licences, PT subscriptions (GA, Half-fare, U-Abo, …), vehicle ownership, main mode by trip purpose, shared-mobility usage frequency, EBSS-PT integration patterns, mode substitution, weather likelihood, price sensitivity.

  • Q"I use e-bike sharing… for complete trips A→B / with public transport / with car / with walking / as a fallback."5-pt freq
  • Q"How often do you use e-bike sharing as a substitute for: car / private bicycle / PT / walking…"5-pt freq
  • Q"How likely would you use EBSS under: high temp / low temp / light rain / heavy rain / wind / darkness?"5-pt likelihood
B.
Part 2 · Environment, policy & TPB

Beliefs, attitudes, policy preferences.

Theory-of-Planned-Behaviour-style items on environmental responsibility, EBSS as sustainable transport, hedonic evaluation, social support; plus policy preferences on infrastructure investment, subsidies, PT-app integration, private financing.

  • Q"I have a personal responsibility to reduce my environmental impact."Likert 1–5
  • Q"E-bike sharing is a good way to alleviate pollution in the city."Likert 1–5
  • Q"The public sector should monetarily subsidise EBSS at the same level as public transport."Likert 1–5
C.
Part 3 · Barriers & investment

What gets in the way; what comes next.

Perceived barriers — availability variance, price, convenience compared to private (e-)bike — plus geographic coverage feedback (open text) and intent to purchase a private e-bike within twelve months.

  • Q"E-bike sharing availability varies too much."Likert 1–5
  • Q"Did/does EBSS use inspire you to buy your own e-bike?"Yes / No
  • Q"Does the current Pick-e-Bike service area cover your needs? If no, which regions?"Yes/No + open
D.
Part 4 · Sociodemographics

Standard background variables.

Year of birth, gender, country → postal code → tenure in Basel, education, employment status (workload + home-office share), nearest PT stop at work/education, household composition, income.

  • QAge, gender, postal code, education, employmentclosed/numeric
  • Q"Distance from your home / workplace to nearest PT stop"categorical
  • QNet household income (10 brackets)closed
E.
Part 5 · Consent & reward

Data linkage and follow-up.

Explicit consent for anonymous linkage of survey responses to operational rental records, voucher choice, optional follow-up interview consent, optional comments on the survey itself.

  • Q"Do you consent to anonymous data linking (link your survey responses to rental records)?"Yes / No
  • Q"Which survey reward would you like via e-mail?"2 vouchers
  • Q"Do you allow us to contact you for a short follow-up interview/focus group?"Yes / No
F.
Part 6 · Quality controls

Filters before linkage.

Three response-quality filters were applied before retaining cases: a minimum-completion-time threshold (5 minutes), straight-line detection across Likert blocks, and a reCAPTCHA score floor. Net retention: 1,844 of 2,001 (92.2%).

  • ×Speeders < 5 min completion timeexcluded
  • ×Zero-variance Likert response patternsexcluded
  • ×Low reCAPTCHA scoreexcluded
From 32 Likert items to 7 latent factors

An exploratory factor analysis (principal axis factoring, promax rotation κ=4) reduced the 32 attitudinal Likert items to seven interpretable latent constructs. Suitability confirmed by KMO=0.820 and a significant Bartlett's test (p<0.001). Eigenvalue cutoff 1.10 + reliability analysis retained the seven factors below.

FactorConstructItemsCronbach α
F1Weather-Independent Usage Propensity — stated willingness to ride in adverse conditions60.813
F2Public Sector Intervention Attitudes — support for subsidies, infrastructure, policy integration50.532
F3Hedonic Motivation & Social Support — enjoyment, comfort, social endorsement40.665
F4Perceived Functional Superiority over PT — faster, more convenient than transit50.728
F5Multimodal Complementarity Behavior — combining EBSS with PT, walking, fallback use40.598
F6Service Reliability & Affordability Concerns — cost and availability barriers ⚠ marginal reliability40.401
F7Environmental Self-Identity — environmental responsibility, EBSS sustainability beliefs20.633

Standardized factor scores (M=0, SD=1) computed via the regression method and merged into the analytical dataset. Factor 6 results interpreted with caution.

Continue reading

This page covers methodology and instrument design only. For findings, effect sizes, recommendations and the interactive data explorer, head back to the project overview — or jump straight into the key findings, the forest plot of effect sizes, or the spatial heatmap.