Plotted are the distributions of relativizers (in proportions) across different variables. Numbers above the bars indicate the number of contributing cases. This is to facilitate making sense of what look like differences from looking at the height of the bars: the greater the number of contributing cases, the more likely the difference is real.
So far I have just created visualizations of the data and haven't conducted any sort of statistical analysis, so the language describing the patterns in the plots is deliberately hedging.
There seem to be no sex differences.
The only dialect that seems to go against the trend is New England.
There may be a higher incidence of "that"s in the oldest age group for both sbj and non-sbj RCs -- will be interesting to see if this comes out in the analysis or is just an artefact of one or two idiosyncratic old speakers.
For non-sbj RCs, there seem to be more nulls / fewer "that"s with increasing education level. For sbj RCs, it's harder to see a clear pattern.
The antecedent seems to matter quite a bit for non-subject RCs, but not for subject RCs.
Matrix clause type seems to matter.
Head specificity seems to matter: for non-subject RCs, more nulls than "that"s for empty heads; reverse for nonempty heads. For subject RCs, there appears to be a greater incidence of "who"s with empty heads (maybe because there's a lot of empty animate heads like "guy"?)
There appear to be more "that"s with increasing amount of intervening material, and fewer nulls (non-sbj) / fewer "who"s (sbj).
There appear to be more "that"s & fewer nulls with increasing RC length for sbj RCs, and the opposite for non-sbj RCs: fewer "that"s and more "who"s.
RC subject type seems to matter, even within pronominal subjects
-
Run analyses:
- separately for sbj and non-sbj
- full dataset (leaving out RCSubjType, that's undefined for the sbj RCs)
-
Things to ponder/try out in analysis:
- collapse "lonehead" into "other" category (MatrixClauseType) / wait for Tom's "presentational" (discourse-new) category to throw the existential and some of the "other"s into
- include dialect/region
- throw out level of education 9 ("unknown")
- try out multinomial regression --> are there enough "wh"s to even do this?
- interactions between socio factors?
- include random slopes for all fixed effects?
- do all analyses once with "wh" collapsed into "that", and once excluding "wh". best case: doesn't make a difference. if it does make a difference, that would be further support for trying to do the multinomial regression.