mixlogit with clustered standard errors: missing values in estimates table

[Julia Jones]

Hi All,

I am using Stata IC 16 to estimate a discrete choice model using the user written command mixlogit written by Arne Hole. Mixlogit generates missing values for cluster-robust standard errors, z values, p values, and the confidence interval when I cluster standard errors on choice set order (t).

The dataset includes choice decisions made by respondents during a single online survey. Respondents were presented with a series of four choice sets and were asked to choose one alternative from each. Each choice set has 3 generic (unlabeled) alternatives (alternative 1, alternative 2, and a status-quo alternative 3). Each alternative is composed of five attributes (price, distance, control, frequency, days). Alternatives 1 and 2 vary freely across choice sets, while the status-quo alternative remaining constant for each individual (varying across individuals).

When I try to cluster standard errors on t (choice set order - 1, 2, 3, or 4), I generate missing values for robust standard errors, z values, p values, and the confidence intervals. I'd like to estimate this model because clustering on t would allow me to control for respondents “learning” in different ways as they progress through each choice set. I've been using the "mixlogit" command successfully thus far, even when I cluster standard errors on pid (unique identifier for each respondent) and gid (unique identifier for each choice set).

Here is my code:

 global randvars "distance control frequency"
mixlogit y price days if coop_distributor_foodhub=="Yes", rand($randvars) group(gid) id(pid) nrep(1000) cluster (t)

Here is the estimate table this code generates:

Any feedback would be greatly appreciated!

Sincerely,

Julia Jones
Department of Natural Resources and the Environment
University of New Hampshire

[Clyde Schechter]

So, there are two layers of wrongness to this approach.

First, you have only four clusters in t, which means you have only 3 degrees of freedom for estimation. So you cannot estimate standard errors for models with more than 3 predictors, because the degrees of freedom are exhausted.

Second, while there is no universal agreement on the minimum number of clusters acceptable, it is recognized that cluster robust standard errors attain their validity asymptotically, with large numbers of clusters. I think you would fail to find anyone who would consider an analysis with only 4 clusters acceptable, even if you removed enough predictors to get standard errors for the remaining ones.