Page 1 of 1

Runtime and convergence for large multi-select MDCEV models

Posted: 14 Sep 2026, 19:38
by a_ali
Hi all,

We are currently designing a multi-select stated choice experiment and would really appreciate some advice on whether the analysis we have in mind is practical in Apollo before we finalize the design.

Our setup is:

21 inside goods plus a no-buy/outside good
Respondents can select multiple items and report a quantity for each, so we observe both participation and continuous consumption
Approximately 3,000 respondents completing 10 choice tasks each, giving 30,000 panel observations
We are planning to estimate and compare an MDCEV model and a nested version. Ideally, we would like to consider specifications both with and without a budget constraint and allow for random heterogeneity across respondents.

Our main question is whether this scale is reasonable for these models in Apollo, particularly once we introduce continuous mixing. Are there practical limits or challenges we should anticipate in terms of runtime, number of draws, identification, or convergence?

Any guidance from those who have estimated similar models in Apollo would be greatly appreciated.

Re: Runtime and convergence for large multi-select MDCEV models

Posted: 15 Sep 2026, 06:41
by stephanehess
Hi

the version without mixing should not cause problems, especially as the forthcoming version of Apollo will include analytical gradients for MDCEV and MDCNEV. For the version with mixing, you'll want to ensure you run this on a machine with sufficient RAM and a large number of cores

Stephane

Re: Runtime and convergence for large multi-select MDCEV models

Posted: 24 Sep 2026, 05:17
by a_ali
Thank you for your response. Looking forward to the new version!

Re: Runtime and convergence for large multi-select MDCEV models

Posted: 16 Oct 2026, 10:16
by sophiedavies
Yeah the “provide full details” stuff at the top is exactly what saves time when you’re dealing with runtime/convergence headaches on big MCD*EV setups. In my experience, the version without mixing is a good sanity check, but once mixing is in, the bottleneck is usually memory plus how many draws you end up with, not some mysterious syntax issue. Also, if you’re on a machine with limited cores/RAM, you’ll get convergence failures that look like “model trouble” but are really just compute limits.