~/projects/general-conference/training
What the training shows
The leadership database records where each General Authority studied, what, and what they did for a living. This study joins that to their conference talks and asks what the training predicts once the obvious is removed: the calling held when speaking, the decade, and the country of birth. Lawyers and agency, physicians and bodies, business graduates and stewardship, educators and children are the hypotheses; a classifier that guesses the field from the words, speakers held out, is the test of the premise.
What stands out
- Mostly, the training does not show. With every speaker held out, a classifier names the field of 34% of talks, against 44% for always guessing the largest field and 32% when fields are shuffled between speakers; pooling each speaker's talks into one document gives 46% against 38% shuffled. Balanced accuracy is 21% against a chance of 17%.
- 21 of 592 adjusted comparisons survive false-discovery control at q ≤ 0.10, and most of those are small. Business 2,863 talks, Law 1,193 talks, Education 1,107 talks, Engineering and science 580 talks; every effect is measured against the calling, decade and birth-country mix of that field's own talks.
- The profession does not leak into the examples. None of the 42 domain-vocabulary tests (courtroom, clinic, firm, classroom, laboratory words, hedges, list markers) and none of the 24 quotation tests survives the threshold: lawyers do not say “court” more than their calling and era predict.
- Topic and scripture move a little, and not in the hypothesised directions. The largest standardised effects are small: Medicine: proposed sustain · favor manifest · area seventies below expectation (0.41 against 1.86, d -0.25, q 0.00); Engineering and science: bofm/1-ne below expectation (0.61 against 1.18, d -0.07, q 0.00). The physicians' zeros rest on 37 speakers and should be read as sparse, not silent.
- Style follows the field, not the degree level. Law: second below expectation (1.04 against 1.27, d -0.21, q 0.00). No trait differs by highest degree once calling and era are fixed: doctorates do not write longer sentences or hedge more.
- What the paragraphs do differs by a few points at most. Engineering and science: book, read, true, missionaries below expectation (0.06 against 0.10, d -0.13, q 0.00).
- There is no profession in the words to fade with years in the calling. Within 122 speakers with eight or more talks, the classifier's confidence in the true field goes from 0.23 in the first three talks to 0.22 in the last three; it fell for 60% of them.
- BYU-trained speakers do not sound more alike than any random group. Their within-group style distance is 4.296 against 4.323 for random groups of the same size (p = 0.468); 9 Harvard Business School graduates and 36 internationally trained speakers get their own panels.
- 7 research leads, none published. For leaders whose education is still unknown, the classifier's guess is written off-site for the leadership database's research, never shown as a label.
Loading the study…
Model card
- Join and groups
- Conference speakers resolve to leader records through the shared speaker index; 6,569 talks of 300 words or more have a classifiable field. Field is read from the highest degree's subject, degree and education group, with the career sector as fallback: law; medicine and dentistry; business, accounting, finance and economics; education and seminary teaching; engineering and the sciences; humanities and social science. Degree level is the highest degree stated. School flags: any BYU degree, a Harvard Business School degree, and the database's international-training tier.
- Adjustment
- Each talk's expected value for a measure is the mean of its cell (calling group × decade × US-born or not; cells under five talks fall back to the grand mean). A group's effect is the mean of observed minus expected over its talks. Intervals and p-values come from a 1,000-draw bootstrap that resamples speakers, not talks, so a prolific speaker cannot manufacture a finding. Benjamini–Hochberg q-values are computed within each section; the page shows q ≤ 0.10 by default.
- Measures
- Topic assignment from the topic study; scripture references by standard work and the sixteen most-cited books; quotations by kind from the quotes study's per-quotation output; five domain lexicons plus hedges and list markers per ten thousand words; the style study's traits and reading grade; the functions study's paragraph shares; story families from the stories study; voice from the prosody sample.
- Classifier and convergence
- TF-IDF over masked words and bigrams, logistic regression with balanced classes, five folds grouped by speaker. The permutation baseline shuffles fields between speakers twenty times and re-runs the same pipeline. Convergence tracks the out-of-fold probability of the true field, and the speaker's own domain lexicon, by years since first general calling and within speaker from first three talks to last three.
- Schools and frequency
- Cohesion: per-speaker mean of adjusted style (or topic) vectors for speakers with three or more talks, standardised; within-group mean pairwise distance against five hundred random groups of the same size. HBS is compared with other business-trained speakers. Speaking frequency is talks per year from first general calling to last talk for Seventies, adjusted for the decade of the call; session position is order of appearance within the session in the source.
- Known failure modes
- Field is a coarse label for a life; many leaders held two careers and the database records one. Education is sourced for most modern leaders and thin before 1980, so the early decades contribute mostly through career. Calling-by-decade cells are small for women's and auxiliary callings. The permutation baseline is twenty runs, not a thousand. Session order in the source is assumed to follow the programme. None of this is causal: training is confounded with everything that chose it.
Methods and data for the corpus as a whole.