Currently, if pandas is installed, import lmo automatically registers the dataframe- and series accessors (pandas docs). There are several major issues with this approach:

• It significantly slows down import lmo if pandas is installed, even if the user doesn't import pandas. If I remember correctly from a benchmark I did a while back, this is responsible for the majority of the runtime of an (initial) import lmo call.
• It's impossible to statically type: Static type-checkers cannot "be told" that pd.DataFrame and pd.Series actually have the l_* methods, let alone their exact signatures. The only workaround is to manually cast the dataframe or series instances to lmo.contrib.pandas.{DataFrame,Pandas}, which is very annoying.
• There are better alternatives:
  • Series.l_{}(...) just calls lmo.l_{}(_, ...) on the instance, and in the "plural" cases, rewraps it as a Series with a new "r" index. Once Support __array_ufunc__ in lmo.l_(loc|scale|variation|skew|kurtosis)  #184 is implemented, this could (for the most part) also be done with a simple lmo.l_{}, which is a better alternative.
  • DataFrame.l_{}(...) is basically an alias for DataFrame.apply(lmo.l_{}, ...), so almost as easy to do it manually.