PaperlayerView exact revision on arXiv

Source-linked AI summary

A Discrete Choice Framework for Modeling and Forecasting The Adoption and Diffusion of New Transportation Services

Feras El Zarwi, Akshay Vij, Joan Walker

arXiv:1707.07379v1stat.AP

TL;DR

Current travel demand models lack a mechanism for forecasting adoption and market shares of new transportation modes. The paper integrates latent class choice and network-effect models and estimates them using carsharing enrollment data. The fitted model identifies three behaviorally consistent classes, and its final log-likelihood is higher than the MNL model’s after choice-based sampling.

  • Problem

    Current travel demand models lack a mechanism for projecting membership and market shares of new transportation modes, limiting long-range forecasts.

    Source: §1 · p. 3

  • Method

    The paper integrates a dynamic latent class choice model with a network effect model and estimates it using revealed-preference carsharing enrollment data.

    Source: §1 · p. 3, §7 · p. 29

  • Results

    Final log-likelihoods after choice-based sampling were -14.74 for the LCCM and -156.53 for the MNL model.

    Source: §5 · p. 20

  • Takeaways & Limitations

    The model identifies three latent adopter classes with significant, behaviorally consistent preferences and supports forecasting adoption under investment scenarios.

    Source: §1 · p. 3, §5 · p. 20

    The model’s predicted adopter count misses the 30th-month actual value, which the authors say may reflect unobserved competition or service issues not captured by the data.

    Source: §5 · p. 21

Abstract

from arXiv · show

Current travel demand models are unable to predict long-range trends in travel behavior as they do not entail a mechanism that projects membership and market share of new modes of transport (Uber, Lyft, etc). We propose integrating discrete choice and technology adoption models to address the aforementioned issue. In order to do so, we build on the formulation of discrete mixture models and specifically Latent Class Choice Models (LCCMs), which were integrated with a network effect model. The network effect model quantifies the impact of the spatial/network effect of the new technology on the utility of adoption. We adopted a confirmatory approach to estimating our dynamic LCCM based on findings from the technology diffusion literature that focus on defining two distinct types of adopters: innovator/early adopters and imitators. LCCMs allow for heterogeneity in the utility of adoption for the various market segments i.e. innovators/early adopters, imitators and non-adopters. We make use of revealed preference (RP) time series data from a one-way carsharing system in a major city in the United States to estimate model parameters. The data entails a complete set of member enrollment for the carsharing service for a time period of 2.5 years after being launched. Consistent with the technology diffusion literature, our model identifies three latent classes whose utility of adoption have a well-defined set of preferences that are significant and behaviorally consistent. The technology adoption model predicts the probability that a certain individual will adopt the service at a certain time period, and is explained by social influences, network effect, socio-demographics and level-of-service attributes. Finally, the model was calibrated and then used to forecast adoption of the carsharing system for potential investment strategy scenarios.

1. INTRODUCTION

Current travel demand models lack a mechanism for forecasting adoption and market shares of new transportation modes. The paper proposes an integrated adoption framework that represents distinct adopter segments and the spatial effect of the technology network.

1. INTRODUCTION

  • Current travel demand models cannot project membership or market shares for emerging modes, limiting long-range forecasts of travel behavior.The introduction situates this gap amid anticipated transportation and infrastructure changes that may affect short- and long-term travel decisions.Source: §1 · p. 3
  • The proposed framework integrates a latent class choice model with a network effect model governed by destination choice.It represents innovators/early adopters, imitators, and non-adopters, allowing adoption preferences to differ across market segments.Source: §1 · p. 3
  • The framework accounts for heterogeneity across adopter segments and measures how growth in the technology network affects adoption.The introduction contrasts this disaggregate approach with aggregate diffusion models and their limited coverage of policy variables such as spatial configuration.Source: §1 · pp. 3–4
  • The model incorporates social influences, network effects, socio-demographics, and level-of-service attributes to explain adoption across market segments.The proposed application is forecasting market shares of upcoming transportation modes for policy and investment strategies.Source: §1 · p. 4

2. LITERATURE REVIEW

The literature review contrasts aggregate diffusion approaches with disaggregate adoption models and motivates a framework that represents individual heterogeneity, social influence, and spatial effects. It also reviews prior transportation applications and the limits of stated-preference evidence for dynamic adoption forecasting.

2. LITERATURE REVIEW

  • Bass’s diffusion model distinguishes innovators from imitators and describes adoption as rising from a slow start to takeoff and eventual market saturation.The review also describes cumulative adoption as an S-shaped curve and sales over time as bell-shaped.Source: §2 · p. 5, Figure 1 · p. 6
  • Aggregate diffusion models forecast population-level adoption, but their policy sensitivity is limited and they do not capture spatial configuration as needed here.The review notes extensions incorporating price, advertising, and other marketing variables, as well as cross-country diffusion applications.Source: §2 · pp. 5–7
  • Disaggregate adoption models represent individual adoption probabilities using decision-maker and technology attributes, communication channels, and time.The review highlights their utility-maximizing basis, heterogeneity, transferability with calibration, and ability to represent policy variables.Source: §2 · p. 6
  • Threshold models represent heterogeneous adoption thresholds that decline over time, but the authors seek a framework for three distinct market segments.The proposed segments are early adopters, imitators, and non-adopters.Source: §2 · p. 7
  • Prior transportation diffusion work includes time-series ridership studies and agent-based electric-vehicle models, including models combined with discrete choice.Examples include revealed-preference rail ridership analysis and electric-vehicle simulations incorporating user differences and social interactions.Source: §2 · p. 7
  • Studies of alternative-fuel vehicle adoption often use stated-preference data, which can represent attribute sensitivities but may require calibration and validation for forecasting.The review argues that hypothetical scenarios do not capture actual market demand or the dynamic social and spatial dimensions of diffusion; earlier studies overestimated projected shares.Source: §2 · p. 8

3. METHODOLOGICAL FRAMEWORK

The framework combines a latent class choice model with a destination-choice network-effect model to represent adoption over time across innovators/early adopters, imitators, and non-adopters. Class-specific adoption behavior incorporates social influence, spatial accessibility, socio-demographics, and service attributes.

3. METHODOLOGICAL FRAMEWORK

  • Three market segments—innovators/early adopters, imitators, and non-adopters—represent distinct adoption behavior and shape early uptake, later diffusion, and the curve’s plateau.Innovators influence the curve’s early slope, imitators its later adoption rate, and non-adopters the timing of its plateau.Source: §3 · p. 9
  • The latent class choice model combines class membership probabilities with class-specific adoption-choice probabilities, accommodating unobserved segment membership and differences in choice behavior.The class-specific choice model conditions adoption on membership in one of the three segments.Source: §3 · pp. 9, 13
  • The model predicts adoption or non-adoption for each person and period, combines the destination-choice model with class-specific adoption and membership models, and is estimated with the EM algorithm.The destination-choice model dynamically supplies accessibility to the adoption model before joint estimation of adoption and class membership.Source: §3 · pp. 10, 14
  • Innovators’ adoption utility depends on decision-maker and technology attributes, imitators’ utility also reflects prior adopters, and non-adopters have an alternative-specific constant.The cumulative number of adopters in the previous period represents social influence for imitators; the non-adopter constant is intended to be highly negative.Source: §3 · pp. 10–11
  • A destination-choice model measures accessibility with a utility-based logsum, representing the expected worth of available destinations and feeding spatial network effects into adoption.Its destination utility accounts for travel distance, destination employment, socio-demographics, service attributes, and proximity to home; accessibility can change as destinations or explanatory variables change.Source: §3 · pp. 11–12
  • Early adopters and imitators differ in their sensitivities to decision-maker and technology attributes, while only imitators’ adoption behavior is influenced by social influence.A person may be an early adopter in a local context despite adopting later in calendar time, because the framework represents disaggregate decisions.Source: §3 · p. 12

4. DATASET

The dataset combines carsharing enrollment and travel records with a household travel survey sample treated as non-adopters. Its zip-code-focused analysis uses the service’s station network to construct accessibility measures and accounts for choice-based sampling.

4. DATASET

  • The service had 14 pods or stations and five on-street pick-up/drop-off locations, represented in the analysis as 16 zip-code-based stations.Some on-street locations shared a zip code with other stations.Source: §4 · p. 15
  • The dataset covers all service sign-ups over 2.5 years, with six months of trip records; analysis retained 1,847 adopters who used the service during the final six months.An adopter was defined as someone who signed up and made at least one trip during that final period.Source: §4 · p. 15
  • The dataset figures track cumulative adopters and the growth of pods/stations and on-street locations over the 2.5-year period.The cumulative-adopter figure was used to locate the system’s position on the diffusion curve.Source: Figure 4 · p. 15, §4 · pp. 15–16, Figure 5 · p. 16
  • The study added a random sample of 2,724 household travel survey observations from the same state, assuming they were non-adopters throughout the 2.5-year period.Given a prior adopter probability of 3x10-4, the sample was expected to contain approximately one adopter.Source: §4 · pp. 15–16
  • For zip codes without a station or on-street parking, accessibility was imputed from the nearest served zip code with distance friction estimated in the model.The accessibility measure was based on the spatial distribution of destinations, using the system’s stations as origins and destinations.Source: §4 · p. 16
  • Because adopters were over-represented in the choice-based sample, observations were weighted using population and sample fractions for each stratum.The passage identifies these fractions as W_g, the population fraction, and H_g, the sample fraction.Source: §4 · pp. 16–17

5. ESTIMATION RESULTS AND DISCUSSION

The estimation links destination-choice accessibility to class-specific adoption and reports distinct class patterns and hold-out forecast performance. Results also expose a month-30 miss that the authors associate possibly with unobserved market or service factors.

5. ESTIMATION RESULTS AND DISCUSSION

  • The destination-choice model used six months of carsharing trips across 16 zip-code stations, with distance, employment rate, home, and on-street parking as explanatory variables.Four station-specific alternative constants were included; constants were omitted from all 16 utility equations to avoid difficulty evaluating accessibility for new stations.Source: §5 · pp. 17–18, Table 1 · p. 17
  • Logsum accessibility values from the destination-choice model were calculated for carsharing and household travel survey observations before the datasets were integrated for adoption-model estimation.The household survey variables used were home zip code, gender, and work TAZ.Source: §5 · p. 18
  • The model’s rho-bar-squared was almost 1.0, which the authors attribute to observation weights and the very small adopter market share.The sample contained 4,571 individuals and 120,665 observations.Source: §5 · p. 18
  • The estimated population split was 0.22% innovators, 16.80% imitators, and 82.98% non-adopters; high-income groups and men were more likely to be innovators.Monthly income was represented by average zip-code income because individual income was unavailable.Source: §5 · pp. 18–19, Table 2 · p. 18
  • Adoption utilities had expected signs and were significant at 1%, except major technology firm employment for innovators; imitators also responded to prior adopters.Accessibility was associated with adoption for both adopter classes, while major-firm employment was associated with early-adopter and imitator utility.Source: Table 3 · p. 19, §5 · p. 20
  • For months 27–29, actual adopters fell inside the forecast interquartile boxes; month 26 fell within whiskers, but month 30 fell outside them.The model was calibrated on month 25 and forecast months 26–30 using 1,000 bootstrap draws.Source: §5 · pp. 20–21, Figure 6 · p. 21

6. POLICY ANALYSIS

After calibrating the adoption model, the authors forecast a base case and three investment scenarios. A station outside a major technology firm yields the most forecast adopters, while downtown station and on-street parking forecasts are similar; the Bass model cannot distinguish the strategies.

Calibration

  • The model is calibrated by adjusting innovators’ and imitators’ adoption-utility constants until predicted and actual final-month adopters match.Calibration uses sample enumeration across the city’s population.Source: §6 · p. 22

Scenario design

  • The base case assumes no new stations or on-street parking facilities, while the alternatives add facilities outside a major technology firm or in a downtown zip code.The downtown alternative compares a station/pod with on-street parking in the same zip code.Source: §6 · p. 22

Scenario forecasts

  • Stations/pods outside a major technology firm yield the most forecast adopters; downtown station and on-street parking forecasts show no significant difference.The downtown comparison is limited because facility effects were represented only by dummy variables; car counts and parking area were unavailable.Source: Figure 8 · p. 23, §6 · p. 23

Forecast uncertainty

  • The predicted cumulative-adopter confidence-interval bounds widen over time, indicating increasingly stochastic forecasts; the technology-firm location remains highest.The comparison uses 1,000 bootstrap draws of the adoption-model estimates.Source: §6 · p. 23, Figure 10 · p. 25, Figure 11 · p. 26

Bass-model comparison

  • The Bass model estimates p = 0.0051, q = 0.2108, and M = 2200, then predicts cumulative adoption will plateau at 2,200 across the investment strategies.Its forecasts are identical across strategies because it omits policy variables, spatial configuration, and socio-demographic influences.Source: §6 · p. 28, Figure 13 · p. 28

7. CONCLUSION

The framework combines a latent class choice model with a network-effect model to represent heterogeneous adoption preferences and assess spatial effects. In forecasts, a station outside a major technology firm yields the most monthly adopters, while downtown station and on-street parking outcomes show no significant difference.

7. CONCLUSION

  • The framework integrates a latent class choice model and network-effect model, representing innovators/early adopters, imitators, and non-adopters with heterogeneous adoption preferences.The analysis considers social influences, network/spatial effects, socio-demographics, and level-of-service attributes.Source: §7 · p. 29
  • The network-effect methodology quantifies how network size and access to more destinations relate to adoption across market segments, supporting location prioritization.The model also uses utility maximization and captures multiple sources of decision-making heterogeneity.Source: §7 · p. 29
  • A station/pod outside a major technology firm yields the most expected monthly adopters, while downtown station and on-street parking forecasts show no significant difference.The findings come from calibrated forecasts across potential investment scenarios.Source: §7 · p. 29

Related papers

Loading 1707.07379v1…