Scenario prediction with apollo_prediction and respondent-specific BAU: shares vs max(p) vs simulation

[Username22]

Hello Apollo team,

I am estimating labelled mixed logit models in Apollo on household biowaste management. Each choice set includes a respondent-specific
business as usual (BAU) / status quo alternative: “I prefer to keep my current practice.”

My goal is to predict behaviour under policy scenarios, and also to derive the expected number of households participating in each option.

My question concerns post-estimation prediction and, more specifically, how to handle the BAU alternative when reporting scenario results and when comparing predictions with observed real-world behaviour.

If I want an individual-level predicted behaviour, should I use:

- the maximum probability rule (max(p)), i.e. assign each individual to the alternative with the highest predicted probability,
- or simulated choice draw from the predicted probabilities?

How should I handle the BAU alternative?

Because BAU is respondent-specific (“keep my current practice”), it is not a unique concrete behaviour category.
In my case, respondents’ current practices may overlap with the labelled alternatives used in the experiment (for example: individual composting, curbside collection, recycling centre), rather than representing a separate behavioural category.

Would it therefore be methodologically acceptable to:

- first report predictions in the original choice space (Alternative 1 / Alternative 2 / Alternative 3 / BAU),
- and then recode BAU into the respondent’s actual current behaviour when comparing predicted versus observed real-world practices?

This seems necessary to me for external validation, and behavioural prediction because otherwise BAU remains an abstract choice category rather than an observable behaviour.

My objective is not only to predict choices within the experimental setting, but to approximate future household behaviour under policy scenarios. In particular, for my PhD research, these predictions will feed into a broader framework assessing the feasibility of alternative biowaste treatment chains. I fully understand the limitations of stated preference data, so I interpret these results as ex ante behavioural projections, not exact forecasts.

Any guidance on recommended practice would be greatly appreciated.

Thank you very much.

[stephanehess]

Hi

max(p) is always the incorrect approach as it ignores the probabilistic nature of the model. If you need predictions at the person level, you indeed need to sample according to the probabilities. Have a look at ?apollo_prediction and the simChoice setting

Re BAU, I think your approach sounds fine

Stephane

[Username22]

Hello Stéphane,

Thank you very much for your reply.

I tried using simChoice = TRUE as suggested, but I am still struggling to fully understand how to interpret the results. I apologise for the follow-up question.

With simChoice = TRUE, I do get an additional column with simulated choices. However, I obtain potentially different predicted choices for each row, while in my data one individual corresponds to 5 rows (i.e. 5 choice tasks).
In my current application, I am testing a reference scenario where all 5 choice tasks shown to a given individual are identical. As a result, the predicted probabilities for each alternative are also identical across the 5 rows for that individual.

Is it expected that simChoice produces different simulated choices across identical rows ?
If so, how should I derive a single predicted behaviour at the individual level? Should I select one of the rows ?

In addition, I noticed that when I rerun apollo_prediction, i always obtain exactly the same simChoice results. This surprises me, as I would expect different simulated outcomes if the process is stochastic.

My initial idea was to simulate individual choices multiple times, recode BAU into actual behaviours, and then compute average participation levels and standard deviations. However, given the issues above, I am not sure whether this is the correct approach or how to implement it properly.

What would be the recommended way to obtain individual-level predictions in this context?

Thank you again for your help.

Best regards,

[stephanehess]

Hi

you are getting different predictions in different rows given the probabilistic nature of the model, so a separate choice is sampled in each row

Stephane