It's a few years late but I think I've discovered the bug that may have been your problem (or at least generates a similar error that others might find via a search):
Linear regression (specifically the fitlm() method) fails when using a table as first input AND when that table contains a variable with 3 or more dimensions. This failure occurs even when that variable is not included as a term in the fit.
Bug is reproducible via the following, showing a fit working with minimal variables, still working with an unused 2d variable, and then failing with an unused 3d variable:
X = load('carsmall')
cars = table(X.MPG,X.Weight,nominal(X.Model_Year),'Var',{'MPG','Weight','Model_Year'})
fitRunsOk = fitlm(cars,'MPG~Weight*Model_Year')
cars.unused2dVar = rand(size(cars,1),5)
fitStillOk = fitlm(cars,'MPG~Weight*Model_Year')
cars.unused3dVar = rand(size(cars,1),5,5)
fitFAILS = fitlm(cars,'MPG~Weight*Model_Year')
>> Error using classreg.regr.FitObject/assignData (line 134)
Predictor and response variables must have the same length.
The workaround until fixed (I've submitted a bug report) will be to make a temporary copy of your table that omits any variables with 3 or more dimensions.
