Limitations of market share forecasts with SP data

[DrWho]

Hi Stephane and Team,

I have estimated a model for new subscription products (including an opt-out) using a Stated Preference Survey.
I would now like to estimate potential market shares for different price scenarios. For this purpose, I have opted for an disaggregate approach by generating a population dataset with covariates from census statistics.

A very experienced researcher in choice modelling has told me that estimating market shares with Stated Preference data is simply not possible and invalid. Not only because of the hypothetical bias, but also because the ASCs are heavily influenced by the attribute levels in the experiment design.

While I am fully aware of these biases and limitations, I am wondering whether these forecasts are really "garbage" or still have some value. In your book "Handbook of Choice Modelling" it states in chapter 25.6 (New Alternatives) that SC forecasts are sometimes just the best available alternative and such forecasts are not described as fully invalid.

Therefore I would like to ask:

Under what circumstances (if ever) is it valid to report hypothetical market share forecasts based on the SC data? (Of course while still acknowleding the limitations)
Or would you recommend just reporting relative changes / elasticities when only SC data is available?

[stephanehess]

Hi

SP alone is indeed problematic for forecasting. But these issues affect not just market shares, but even more so (in my view) elasticities/marginal effects, as responses in SP are quite likely overstated. Of course, if you have no other data, you can still produce such forecasts, but make sure you discuss the limitations thereof. Are you at least able to correct the scale of the model using some real-world cost elasticity?

Stephane

[DrWho]

Thanks a lot for your response, despite going beyond the core of Apollo.

Unfortunately, I only have SP data. I will be able to collect some RP on the subscription products, but I will only observe absolute purchase numbers for the different products (and no active opt-out) at one specific price point, without the possibility of directly linking this back to individual survey participants. So I am not sure whether this can help in any way?

Despite these limitations I see a lot of papers using SP data only (for example in Transportation Research Part A/D). So as far as I understand, the results still have scientific value, but you just need to be very careful in framing the implications?

[stephanehess]

Yes, the results should still have value.

If you have absolute purchase numbers, then you can calculate market shares and correct the ASCs. You could use the SP market share for the opt-out

[Julián]

Dear Professor,

I am using SP data to estimate a mode choice model (road vs SSS) with MDCEV. I have the market shares revealed by the decision-makers to make predictions. We would like to use them to correct the ASCs from SP estimates.

How can we do that in MDCEV in Apollo? I do not know how to proceed with the code.

I would appreciate your help!

Thank you in advance
Julián

[stephanehess]

Hi

standard ASC correction approaches exist for discrete choice (see Train's book), where you do asc_v2 = asc_v1 + log(real_marketshare/model_market_share)

but this is for discrete choice, not mdcev. What you could do with mdcev is to make a first prediction, and then adjust the values for the constants inside model$estimate and predict again and see how much you need to change them to align the market shares with your expectations

Stephane