CLM | Face Tracking and Warping Demos | Computer Vision library

One option would be to do two splits, but only one is actually necessary, as the second split only needs the first value.

So we can split on |, then search for the first ,

I don't have a complete solution but hopefully enough feedback to get you started:

• It is important to distinguish so-called "conditional logit" models with alternative-specific variables (which you are interested in) and classic "multinomial logit models" with only subject-specific (or individual-specific) variables. mlogit::mlogit() can fit both (and also mixed versions) while nnet::multinom() only supports the latter.

• For fitting conditional logit models you could have the data in long vs. wide form. You specify your data in long form and also have a splitting variable z1 that is alternative-specific. This means that data from the same individual could end up in different nodes of the tree which would be rather awkward.

• Instead it would be better to have the data in wide form so that each row corresponds to an individual and then you could only consider individual-specific variables for splitting. This would also match the view of the $gradient element of a fittedmlogit object which provides the individual-specific gradient contributions. (This is what sandwich::estfun() extracts which in turns is the essential information for partykit::mob().)

• It might also be possible to do sensible recursive partitioning based on alternative-specific variables but I find it hard to see what kind of models this would yield and what these would mean. In any case, you then would have to write your own code to extract the estfun from the fitted-model object that provides the fully disaggregated alternative-specific gradient contributions.

• The loret package is somewhat unfinished and hasn't been updated in quite a while. Hence, I would not recommend using it "in production" at the moment. Also nnet::multinom() (underlying loret::multinomtree()) does not fit the model you need (as mentioned above) and ordinal::clm() (underlying loret::clmtree()) is for a completely different model.

• One specific aspect that we wanted to build into loret but did not finish yet, is automatic detection of (quasi-)complete separation in logistic models and handling it appropriately in the tree.

• Your mlogit + partykit::mob() approach does not work because the fitting function does not have the right interface (as you are correctly informed). See vignette("mob", package = "partykit") for the two supported interfaces.

• To write an appropriate interface you need to make sure that you have all the variables you need in each subset. Note that the response variable plus regressor matrix is not enough for this but you need the index variables as well! I would recommend to include these variables either through the y variables or the x variables of the formula specified for partykit::mob(). In mob_control() you can then set ytype = "data.frame" and xtype = "data.frame". Then both y and x are provided as data.frame objects and can be combined again prior to calling mlogit::mlogit(). The formula and idx arguments for mlogit() have to be provided in some way then. In the example below I have hard-coded them.

Illustration based on your example: You can set up a model-fitting function my_fit() that expects y and x to be data frames and then uses the formula = y ~ x1 + x2 and idx = c("id_choice", "altern") to fit the mlogit() model.

That is the work of tukeyHSD fucntion:

You are missing " in your Anchor tag

U are using:

I would do it this way:

If the values are constants and no formulas you don't need to loop through the data and can just use SpecialCells(xlCellTypeConstants) on the range to get all constant values (without the blank cells).

That is possible using CSS. Once we have overwritten table-layout: fixed; of the HTML element table the header cells respond to the min-width attribute. You can adapt the minimum width to your needs.

Make sure you use the correct inputID to select the HTML table. In CSS the ID is preceded by #.

Disadvantage (of course): at some point you get a horizontal scroll bar.

The following should work fine.