3

Say, for example, that I want to determine the market share or relative popularity of coffee houses within a certain population through a survey. What is the best way to write a question that accurately measures this?

Some issues that I am thinking of:

I don't want to ask a single choice question ("Which coffee house do you go to?"), because the coffee houses are not mutually exclusive. I may enjoy more than one equally often.

If I ask a multiple choice question, then I can't really get a real "market share," since the sum of the proportions of people who go to each coffee house sum to over 100%. I can only say that "x% of people in this population go to this coffee house."

Is it possible to ask a series of questions ("Which coffee house do you prefer the most?" "Which coffee house do you prefer the second most?", or "Rank the following coffee houses")?

Can I ask a multiple choice question then rebase the proportion to the total number of responses? For example, if I have 100 respondents, but they selected 200 coffee houses (because each respondent said they went to two coffee houses, maybe), can I calculate a frequency table based on 200, the number of selections, instead of the number of respondents?

hgcrpd
  • 1,427

4 Answers4

3

I'd suggest several Qs along these lines:

  1. Which is the one coffee house you go to most often?
  2. What other coffee houses do you visit more than once a month (say)? [Probe to negative, i.e. keep asking ".. and which others do you visit more than once a month?" ".. and which others?" until interviewee answers "none"/"that's it" / "no others" ]
  3. What other coffee houses can you recall having ever visited? [Probe to negative]
  4. [For each coffee house mentioned in (1) or (2)]: "How many times have you visited Cafe Y in the last month?"

Tabulating responses to (1) gives "X% of people said they go to Cafe Y most often" for each Y, with sum <=100% (can be <100% as some people never visit coffee houses).

Combining responses to (1) & (2) gives "X% of people go to Cafe Y more than once a month" for each Y, with sum > 100% but still with denominator N for each coffee house, where N is number of respondents. Similarly combining (1), (2) & (3) gives "X% report ever having visited Cafe Y".

Taking the mean of (4) for each coffee house gives you "Cafe Y was visited an average of m times in the preceeding month". Denominator is N again (remember to include those respondents who didn't mention Cafe Y as zeroes).

In principle you could refine things further, but asking more Qs that this may increase your respondent quit rate.

onestop
  • 17,737
  • 2
  • 62
  • 89
2

I think there is a big difference (both practical and how you approach the problem) in popularity vis a vis market share. Since market share is more analytically challenging, I'll focus on that.

In my opinion, the best solution to this problem is going to involve a set of stated preference experiments, more formally known as discrete choice modeling. DCM has applications in a variety of contexts all involving a consumer and some choice they need to make. That choice may be what coffee shop to go to, what computer to buy, or how to get to work in the morning. If you think about the important the important attributes that may impact consumer choice when picking a coffee shop to visit, you will be able to design experiments to capture this information. This could include price, offerings, proximity to work, brand loyalty, etc. You would want to design a set of experiments that tests each of these attributes independently of one another and then use one of the DCM techniques to develop coefficient estimates for those attributes. Turning that into a market share simulator is a relatively easy task, though model validation is always an issue.

In my work, the "seminal" reference text is by Ben Akiva and Lerman.

Ken Train's book is also a great resource.

Finally, Sawtooth Software is a fairly large player in this market and they have a variety of tools that can assist in the design and analysis of DCM. I don't use their tools that often however.

Another method that you may be able to adapt is the Van Westendorp method. We recently used a set of questions to validate our DCM for a high end electronics study we performed. Van Westendorp was quick, easy, and provided instant feedback for our client. It certainly has it's drawbacks, but that's for another discussion.

User1865345
  • 8,202
Chase
  • 3,073
  • 3
  • 21
  • 29
  • I didn't fully digest the "multiple choice" aspect of the question previously. I would look further into the Van Westendorp methods if that is what you are limited to... – Chase Nov 17 '10 at 19:27
1

Frequency show the values that variables take in a sample. In other words, shows the amount of individuals said that prefer coffeehouse A. This amount would be the frequency of coffeehouse A.

axis: X:coffeehouses Y:amount of individuals

dimitris
  • 111
  • 1
    But what if an individual goes to more than one coffee house? Or what if Person A goes once per month to five different coffeehouses, and Person B goes five times per month to just one coffeehouse? – Matt Parker Nov 17 '10 at 16:03
  • I agree with Matt, and this is probably not the best measure per say to assess market share (particularly as people have less loyalty). But this approach is enlightening as to consumers preferences. It will diverge from market share depending on how much people choose multiple vendors and if frequency of coffee buying varies between preferred vendors (i.e. do people who prefer A overall drink more coffee than people who prefer B.) – Andy W Nov 17 '10 at 18:05
1

You can ask consumers a question along the following lines:

Out of every 100 visits to a coffee house how many times do you visit each one of the following coffee house? Please ensure that the total adds up to 100.

A. Option 1
B. Option 2.. etc

You can then normalize to get percentage of times each consumer goes to each one of the coffee houses. A simplistic analysis would then assume that each consumer visits a coffee house the same number of times during a year and simply compute the average of the percentages across all consumers to get an estimate of market shares. A more sophisticated analysis would compute the weighted averages with the weights being the the number of times that a consumer goes to coffee shops. You can get the weights as well via a survey question by asking them how many times they visit coffee shops in a year etc.

There are other ways to estimate market shares but the above seems to be a simple yet reasonable approach.

Community
  • 1
  • 1
    Ouch. In the real world the incidence of people who can't add correctly to 100 is going to be high. Even so, in their minds most people can't tell the difference between 90% and 95% so providing them a scale where they can do so is pointless. I'd suggest you'd be better off giving a Likert scale for each of your major competitors and one marked other. You can then look at the proportion of your likert scale over the sum of all likert scores (lower bound should have a value of zero) as an indication of market share. – russellpierce Nov 17 '10 at 15:07
  • @drknexus Adding up to 100 is an issue which is mitigated to some extent if you do online surveys where the software can check if the total adds up to 100. Not being able to tell the difference between 90% and 95% is measurement error and that can be dealt with if necessary. The question form I suggested is one of the standard question types used in market research but has not been used much in the past because of the issue about totals that you mentioned. I feel that the issue of totals is much less pertinent in online surveys. –  Nov 17 '10 at 15:24
  • I seldom think being a "standard" makes something a defensible practice. That being said, I suppose even in offline surveys I suppose participants responses can be rescaled to total to 100 if they make some sort of math error. – russellpierce Nov 17 '10 at 15:53