axhausen, k.w. (2015) survey challenges, modelling...
TRANSCRIPT
Preferred citation style!
!!!!Axhausen, K.W. (2015) Survey challenges, modelling challenges,
keynote presentation at the 3rd IWATS (International Workshop on Advanced Transport Studies) “International workshop on context and social interactions in activity and travel decisions”, University of Hiroshima, March 2015.!
.!
IDEC 2015
Survey challenges, modelling challenges
KW Axhausen IVT ETH Zürich March 2015
Survey challenges
IDEC 2015
Do we know the numbers? e.g. daily activities in Switzerland
0.0
0.5
1.0
1.5
2.0
2.5
3.0
0 10 20 30 40 50 60 70 80 90 100 110
Num
ber
of A
ctiv
ities
Average Cohort Age [Years]
before 1910!
1910-19!
1920-29!
1930-39!
1940-49!
1950-59!1960-69!1970-79!
1980-89!
1990-1999!
after 1999!
IDEC 2015
What do we want to know?
Who travels
when ?
where ?
with whom ?
how ?
for how long (space and time) ?
for what purpose ?
and spends how much ?
IDEC 2015
Protocols and response
IDEC 2015
Surveys, observations are „talk“
Two speakers
managing their „image“ staying within the rules of talking staying within their socially allocated/identified role fulfilling social expectations
talk and report with/to each other
=>
„Maintaing the willingness of the respondent to report“
IDEC 2015
Response as a function of response burden @IVT, 2013
0"
10"
20"
30"
40"
50"
60"
70"
80"
90"
0" 200" 400" 600" 800" 1000" 1200" 1400" 1600"
AAPO
R"respon
se"ra
te""[%]"
A;priori"es=mate"of"the"response"burden"
None"Only"mo=va=on"call"Prior"recruitment"Prior"recruitment"&"incen=ve"Predicted"
IDEC 2015
Response is a non-random process
0"
200"
400"
600"
800"
0.0"
100.0"
200.0"
300.0"
400.0"
Wave"1" Wave"2" Wave"3" Non1response"survey"
Num
ber""of"
Target"value
"
Observed"True"mean"
Sample"Respondents"
IDEC 2015
Known „error“ generating processes
IDEC 2015
Activities, movement and traces: A full example record
6 12 18
Home Out of home Movement Phone/SMS Email Smart card GPS Bluetooth WiFi IDEC 2015
Active/passive tracing: Many owners, locations, quality levels
6 12 18
Phone/SMS Email Smart card GPS Bluetooth WiFi CCTV
IDEC 2015
Filters imposed/suggested by the study: „Trips“
6 12 18
Home Out of home Movement After “trip” filter:
Home Out of home Movement
IDEC 2015
Filters due to the respondent: Forgetting
6 12 18
Home Out of home Movement After forgetting
Home Out of home Movement
IDEC 2015
Filters imposed by the respondent: Soft non-response
6 12 18
Home Out of home Movement After soft non-response
Home Out of home Movement
IDEC 2015
Filters due to the respondent: Rounding
6 12 18
Home Out of home Movement After rounding
Home Out of home Movement
IDEC 2015
What is left ?
After all processes 3 at home 2 Out of home 4 trips, 2 tours
True 5 at home 9 Out of home 26 Stages, 11 trips, 1 subtour, 2 tours
IDEC 2015
What happens next ?
IDEC 2015
Geocoding addresses
Ideal Street addresses identifying the entry to the
network Best-case Unambiguous street addresses State of the art Street address State of practice Street address/mid-street block/street corners;
missing conversion of facility names Still seen in practice Arbitrary zonal centroid, e,g post offices
IDEC 2015
Calculating distances & travel time
Ideal Complete GPS track for distance and times
with pedestrian-networks added Best-case Minimal gaps, and state-of-the-art imputation
of GPS tracks and modes State of the art SUE derived travel times and distances
(navigation network) State of practice DUE derived travel times and distances
(planning networks) Still seen in practice Shortest path on empty planning networks IDEC 2015
What should we do ?
IDEC 2015
Next steps
• Query what we really need for
• Cost-benefit analysis • Planning of prices and services • Planning for the slow modes • Social accounting
• High-quality multi-modal surveys to establish the measurement errors (add bluetooth and wifi senders, noise profile)
• Error correction models • Cross check against third party sources
• Treat survey data as indicators in a measurement model • Treat traces as indicators in a measurement model
IDEC 2015
, but especially
• Treat respondents as partners in a talk, discussion:
• Frame your request in a way which addresses them in a clearly defined social role (citizen, driver, customer, etc.)
• Account for their constraints (readability of text, full guidance through the forms, require no calculations – unless necessary, speak their ‘language’)
• Be as complex, as the topic warrants, requires, but not more so
• Don’t surprise them with unannounced requests • Don’t ask them to do work you can do
• If appropriate, provide an incentive, acknowledgement
IDEC 2015
Modelling challenges
IDEC 2015
Modelling challenges: The usual worries
Error heterogenity Is it always checked ? Spatial correlations Are they always checked ? Temporal correlations Are they always checked ? Independence Do we check the correlations of the
independent variables (sample) thoroughly enough?
Endogenity Do we fully account for it ? (sample selection) Error of the second Do you calculate it ? kind Validation How often do we ask for out-of-sample tests? Substance or do we talk about t-tests ? IDEC 2015
Modelling challenges: Substance or t-tests ?
Effect size
t-stats 1.96
small
large
Adap
ted
from
Zill
iak
and
McC
losk
ey (2
008)
IDEC 2015
Modelling challenges: Substance or t-tests ?
Effect size
t-stats 1.96
small
large
Adap
ted
from
Zill
iak
and
McC
losk
ey (2
008)
Yes, but anything new ?
IDEC 2015
Modelling challenges: Substance or t-tests ?
Effect size
t-stats 1.96
small
large
Adap
ted
from
Zill
iak
and
McC
losk
ey (2
008)
Yes, but anything new ?
Fair enough, but do we know why?
IDEC 2015
Modelling challenges: Substance or t-tests ?
Effect size
t-stats 1.96
small
large
Adap
ted
from
Zill
iak
and
McC
losk
ey (2
008)
Yes, but anything new ?
Fair enough, but do we know why?
OK, but are there surprises?
IDEC 2015
Modelling challenges: Substance or t-tests ?
Effect size
t-stats 1.96
small
large
Adap
ted
from
Zill
iak
and
McC
losk
ey (2
008)
Yes, but anything new ?
Fair enough, but do we know why?
OK, but are there surprises?
Now, this is interesting
IDEC 2015
Modelling challenges: Substance or t-tests ?
Effect size
t-stats 1.96
small
large
Adap
ted
from
Zill
iak
and
McC
losk
ey (2
008)
Yes, but anything new ?
Fair enough, but do we know why?
OK, but are there surprises?
Now, this is interesting
IDEC 2015
Choice modelling challenges
IDEC 2015
Choice modelling challenges: The usual worries
Error heterogenity Is it always checked ? Spatial correlations Are they it always checked ? Independence Do we check the correlations of the
independent variables (sample) thoroughly enough?
Endogenity Do we fully account for it ? (sample selection) Error of the second Do you calculate it ? kind Validation How often do we ask for out-of-sample tests? Substance or do we talk about t-tests ? IDEC 2015
Choice modelling challenges: less usual concerns
Error heterogenity Why don’t we check them ? Number of non-chosen How much leverage do they have alternatives for your problem? Number of choice sets How stable are our estimates? Capacity constraints Do we check for their impact on the
parameters? (attribute values of the known (non)chosen alternatives)
Unit of analysis Do we have a MAUP problem?
IDEC 2015
Residuals: False positives of a membership model So
urce
: Kop
p (F
orth
com
ng)
IDEC 2015
Residuals: MCDEV model of fleet choice So
urce
: Jäg
gi (F
orth
com
ng)
IDEC 2015
Number of non-chosen alternatives: routes So
urce
: Sch
üssl
er (2
010)
IDEC 2015
Number of choice sets: residential choice Ad
apte
d fr
om: S
chirm
er (F
orth
com
ing)
IDEC 2015
Accounting for consistency
IDEC 2015
Learning approach of the generic one-day transport model
Competition for slots on networks and in facilities
Activity scheduling
k(t,r,j)i,n
qi ≡ (t,r,j)i,n
Mental map
IDEC 2015
Model estimation: betai,0 = betai,n? betai,n-1 = betai,n?
Competition for slots on networks and in facilities
Activity scheduling
k(t,r,j)i,n
qi ≡ (t,r,j)i,n
Mental map
IDEC 2015
Parameter estimation
betai,n
betai,0
IDEC 2015!
Model estimation: betai,0 = betai,n? Route and mode
-1.5
-1
-0.5
0
0.5
1
Konstante Skalierungsparam. Zeit Preis alfa f. Gem.F.
Par
amet
er
Kalibrationsschritt
Sour
ce: V
frtic
(200
3)
Do we have a MAUP problem ?
IDEC 2015
Do we have a MAUP-like problem for DCM?
• Location choice, obviously
• But also, mode choice
• Stage • Trip • Sub-tour • Tour • Daily schedule
IDEC 2015
Do we have a MAUP-like problem for DCM?
Stage Trip Subtour Tour Value of Time Walking CHF/h 152 28 26 24 Value of Time Bike CHF/h 194 39 43 40 Value of Time Car CHF/h 135 25 30 27 Value of Time PT CHF/h -30 2 7 6 Value of Time PT access CHF/h 819 15 22 22
TT PT / TT Car - -4.46 12.33 4.07 4.16 TT Walk / Access time PT - 0.19 1.83 1.19 1.09 Transfer / TT PT min -220.43 107.00 31.28 32.92 Interval / TT PT - 0.96 7.00 3.47 6.33 Access time / TT PT - -27.10 7.67 3.02 3.35
Sour
ce: S
chm
ujtz
(201
5)
IDEC 2015
Do we have a MAUP-like problem for DCM? So
urce
: Sch
muj
tz (2
015)
0!
20!
40!
60!
80!
100!
120!
140!
160!
180!
0!
0.1!
0.2!
0.3!
0.4!
0.5!
0.6!
Stage! Trip! Subtour! Tour! Dayplan!R
un ti
me
[s]!
Mod
el fi
t [-] !
ρ2!MNL with route-specific parameters!
ρ2!MNL with mobility tool ownership!
Run time in [s]!MNL with route-specific parameters!
Run time in [s]!MNL with mobility tool ownership!
IDEC 2015
What should we do ?
IDEC 2015
Next steps
• Become more systematic • Test for choice set size effects • Test for the stability of the estimates wrt choice set • Test for the stability wrt imputation of the attribute values
• Check for the right unit of analysis
• Check for the right set of explanatory variables
IDEC 2015
Questions ?
www.ivt.ethz.ch
IDEC 2015
References
Jäggi, B. (Forthcoming) Decision modeling on the household level for energy, fleet choice and expenditure, , Dissertation, ETH Zürich, Zürich. Kopp, J. (Forthcoming) GPS-gestützte Evaluation des Mobilitätsverhaltens von free-floating CarSharing-Nutzern, Dissertation, ETH Zürich, Zürich. Schmutz, Simon (2015) Auswirkung von analytische Einheiten und Aggregationsregeln auf die Verkehrsmittelwahlmodellierung, MSc thesis, Zürich, January 2015. Schuessler, N. (2010) Accounting for similarities between alternatives in discrete choice models based on high-resolution observations of transport behaviour, ETH Zürich, Zürich. Vrtic, M. (2003) Simultanes Routen- und Verkehrsmittelwahlmodell, PhD Dissertation, Fakultät für Verkehrswissenschaften, TU Dresden, Dresden. Zilliak, S. and D. McCloskey (2008) The Cult of Statistical Significance: How the Standard Error Costs Us Jobs, Justice, and Lives, University of Michigan Press, Ann Arbor.
IDEC 2015