Conjoint analysis Advanced Methods and Models in Behavioral Research Conjoint analysis gt Multilevel models You have to understand What it is Which different kinds of Conjoint Analysis there are ID: 246253
Download Presentation The PPT/PDF document "Advanced MMBR" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Advanced MMBR
Conjoint analysis Slide2
Advanced Methods and Models in Behavioral Research
Conjoint analysis -> Multi-level models
You have to understand:
What it is
Which different kinds of Conjoint Analysis there areHow it can be of use in typical TIW researchHow it can lead to different kinds of statistical analyses (of the repeated measures or multi-level kind) For this, you can use these slides AND the literature online In the laptop exam, running a repeated measures analysis is part of the requirements.Slide3
The logic of the course
binary Y
logistic regressionconjoint analysis: way of data collection that might come in handy "repeated measures" / "multi-level" data We practice on self-collected data
some practice/training in survey design and execution
Advanced Methods and Models in Behavioral ResearchSlide4
Advanced Methods and Models in Behavioral Research
Conjoint Analysis
Underlying assumption: for
each user, the "utility" of a product can be written as U(x1,x2, ... , xn) = c0 + c1 x1 + ... + cn xn
10 Euro p/m
2 year minimum
free phone
...
How do you rate this
proposal? (-5 ... +5)Slide5
Advanced Methods and Models in Behavioral Research
Two kinds of research questions
Which phone do people tend to prefer?
How do different attributes of a proposition affect the utility of a proposition? (and how does that differ across different kinds of people)Slide6
Advanced Methods and Models in Behavioral Research
Why is this important?
It is an important tool in social science when you want to investigate how someone
’
s behavior or evaluations depend on circumstances and a useful tool in typical TIW Master’s Theses
It's not only of use in marketing!Slide7
Advanced Methods and Models in Behavioral Research
One of the main advantages = more data:
Example - Adoption of technology
If you ask for behavior only (
“did you adopt” etc), then you get one piece of info per person, and the rest you have to infer by comparing different persons. -> OK, and gives “real” behavioral data, but data are sparse.If you offer different scenarios and then ask whether someone would adopt, you get how adoption depends on the context for this person. -> Richer data per person, but not behavioral
Given that sample sizes of >200 are often necessary, often option 2 is more feasible than option 1.Slide8
Or, use it on ...
price of bulb color of light expected
lifetime
...Would you be interested in buyingthis light bulb
?
(-5 ... +5)
New light
bulb
strongly favor issue personally
core of party's strategy
many other parties against it
gets a lot of media attention
...
Do I submit a motion? (-5 ... +5)
Submit
a motion
in
parliament
...
...
...
...
...
??Slide9
Advanced Methods and Models in Behavioral Research
Example: properties of mobile phones
Give each person 200 of these cases (
“
vignettes”)Per person, run multiple regression on their 200 answers --> you get (personal) values for each of the dimensions.You could then:Create groups of people with the same kind of values, or …… get an estimate of the average trade-off between dimensions, or …… compare different groups of respondents
large b/w screen
long
battery
life
not
flashy
costs 6
Euro/month free monthly contract
How do
you
rate
this
option? (-5 ... +5)
Suppose you get a
phone that has …
Hmm... How would that work anyway?Slide10
Advanced Methods and Models in Behavioral Research
The data would look like this:
Y
D1
D2
D3
D4
D5
id
…
+4
-1
-1
0
1
0
1
-3
1
1
1
0
-1
1
0
0
0
1
0
-1
1
0
1
0
-1
1
0
1
+1
…
…
…
…
…
1
+1
…
…
………1-1……………1+4……………1…………………
Other considerationsHow many dimensions?How many levels per dimension?How many cases per person?Which cases from all possible cases per person?How many persons (sample size)?Judgment or choice?How do I make sure that all given cases are ok?
ISSUES
The D-values are chosen by the researcher -> experiment
200
is way too much
3
5
= 243 > 200
(
or even: 4
10
=
…)
If you look at the complete data set, this is repeated
measures
:
the data are clustered ->
standard multiple regression will not
workSlide11
- Sidestep: design of experiments literature -
Suppose you have a large set of potential vignettes (way too large to use them all)
One option: choose a random subsample of all possible vignettes... but that might not be optimal, here’s why:(X’X)-1Advanced Methods and Models in Behavioral Research
D-optimal design
Taguchi designA,C,E,T,G,I,V –
optimal designSlide12
Alternatives (1)
Ask people directly what their favorite phone is
If you could choose, which phone (and monthly plan) would you prefer? + much easier to collect you will end up with as many suggestions as you have respondentsdoes not capture trade-offs between attributes
Advanced Methods and Models in Behavioral ResearchSlide13
Alternatives (2)
Ask people directly for "their" utility model
For instance: Which of the following factors do you find important when it comes to <buying a phone> / <adopting an innovation> / <choosing a buyer on eBay> / ...? Please divide 100 points over the following factors: - price - battery life - ...+ much easier to collect asks for introspection about choice process, not for a “real” choicenonlinear effects can never occur like this
Advanced Methods and Models in Behavioral ResearchSlide14
Usually, the comparison of interest is
within persons
person 1 person 2 person 3 person 4 person 5 But if you consider only the <sold phones >, you get the comparison between persons (to get a feeling for this difference, think about what would happen to phones that are typically evaluated in 2nd place)
Advantage of
conjoint
analysis
(
subtle
)Slide15
Disadvantages of conjoint analysis
All fictituous decisions (do you like / would you buy / how would you evaluate ...)
(Often) assumes weighted average. This does not allow "all or nothing" weightsCan be quite complicated and/or boring for the respondentsCan be quite complicated for the researcher, both to implement and to analyze...Slide16
Advanced Methods and Models in Behavioral Research
Kinds of Conjoint Analysis
1. CVA
: Conjoint Value Analysis
Respondent evaluates single vignette at a time 2. ACA: Adaptive Conjoint (Value) Analysis Adapting the cases you offer based on previous answers of the respondent. 3. CBC: Choice Based Conjoint Asking for preferences between 2 (or 3 or 4) cases.
4. PP-CBC
: Partial-profile Choice Based Conjoint
Comparing only part of the attributes of the cases.
10 Euro p/m
2 year minimum
free phone
...
How do you rate this proposal? (-5 ... +5)Slide17
Advanced Methods and Models in Behavioral Research
1. CVA: Conjoint Value Analysis
Respondent gets to evaluate one vignette at a time
NoteResearcher chooses the dimensionsThe variance of the multiple regression estimator equals (X’X)-1
(with X the matrix of D
’
s)
This implies that choosing the subset of cases that you are going to use, affects how broad your confidence intervals will be
-> Experimental design literature: full-factorial design,
D
-optimal designs,
etcSlide18
Advanced Methods and Models in Behavioral Research
2. ACA: Adaptive Conjoint Analysis
Adapting the cases you offer based on previous answers of the respondent.
You give, say, 20 cases to each respondent.Which cases the respondent gets, depends on his answers to the first couple of cases. Much more efficient than just randomly choosing casesButimpossible to do off-line or by phone, and even online quite
difficult to implementSlide19
Software … is problematic
Advanced Methods and Models in Behavioral Research
Rare (part of
general survey
software mostly)…
Expensive …
Visually not very nice…
No output of raw data …Slide20
Advanced Methods and Models in Behavioral Research
Sawtooth software / SKIM Research
www.sawtoothsoftware.com
www.skimgroup.com/softwareSlide21
Advanced Methods and Models in Behavioral Research
3. CBC: Choice Based Conjoint
Asking for preferences between 2 (or 3 or 4) cases.
Which of these three offers do you prefer? Or …Rank these three offers Or …Distribute 10 points over these 3 offers
10 Euro p/m
2 year minimum
free phone
internet = per Mb
15 Euro p/m
1 year minimum
phone costs 70
internet = per Mb
30 Euro p/m
2 year minimum
free phone
internet = freeSlide22
Advanced Methods and Models in Behavioral Research
3. CBC: Choice Based Conjoint (2)
Resp
chooses preferredvignette Issues:Which of these kinds of questions works best? We don’t know.If you have only choice data (or ordinal data), how can you arrive at values for the different dimensions?
Y
D1
D2
D3
D4
D5
id
set
1
-1
-1
0
1
0
1
1
0
1
1
1
0
-1
1
1
0
0
0
1
0
-1
1
1
0
1
0
-1
1
0
1
2
1
…
…
………120……………121……………130……………13………………
…3
Y = 0/1 => Logistic regression, but …
… that does not use all the available information and if we use the data for all persons we have dependencies in the data (more cases per person) and … what if we have ordinal data?Slide23
Advanced Methods and Models in Behavioral Research
Which method to choose when?
We have rules of thumb only
… www.sawtoothsoftware.com/products/advisor/Slide24
Sample
sizes
in conjoint analysisThis is difficult ... for many reasonsQuite a lot of degrees of freedom (#people, vignettes pp, dimensions per vignette, values per dimension, ...)
Not clear up front
how different answers from the same person are going to beRules of thumb (see paper online)Generally 150 – 2000 participantsComparing groups? 200+ participants per groupTesting purposes not
comparing
groups
: 300+
Advanced Methods and Models in Behavioral ResearchSlide25
Sample
sizes
in conjoint analysis (2)For choice-based conjoint: n t a > 500 cWith n = number of respondents t = number of tasks per person
a = number
of alternatives per task c = number of “analysis cells” (main effects model: max number of levels in an attrib.)
(
with
interactions
: max
number
of
interaction levels)Advanced Methods and Models in Behavioral ResearchSlide26
Advanced Methods and Models in Behavioral Research
Statistical issues
We need to know how we can deal with
“
nested data” (for instance, more than one answer per person)
Can't the Sawtooth people take care of this?Slide27
Marketing vs other social science
Marketing ("conjoint analysis")
Try to come up with the weights for each dimension for each respondent (e.g. to then segment the population)Other social science ("vignette study") Try to come up with the average weights for groups of people (so other aim, and you need less choices per person)Slide28
Advanced Methods and Models in Behavioral Research
To Do
Read and understand the literature on the course website on Conjoint Analysis.