- see [[250513_154015 canada - strategies|canada strategies]]
# strategy use
"crowd" model: `mean(gpt4.1, deepseekv3-0324)`
strategies (see prompts [here](https://docs.google.com/document/d/1IUgFimHZY24hkoMKtoYzeRBJmzSfh5ZXtOz6ydwM43U/edit?tab=t.0#heading=h.529uhyfnhdpz))
- baseline prompt (harris/trump study prompt)
- generic (don't personalize)
- nofacts (don't use facts)
- noinstr (don't tell LLMs how to persuade - just persuade however it wants)
![[strategy_use 1.png]]
baseline vs noinstr
![[1747798486.png]]
correlation between baseline and noinstr
```r
Parameter1 | Parameter2 | r | 95% CI | t(25) | p
-----------------------------------------------------------------
baseline | noinstr | 0.91 | [0.80, 0.96] | 10.71 | < .001***
```
![[1747798862.png]]
# lasso
```r
# model formula
persuasion~active_listening_and_empathy + aggressive_and_explicit_directives +
audience_adaptation + build_rapport_and_common_ground + cognitive_elaboration +
contrastive_tone2 + encourage_action_with_clear_calls + evidence_or_factbased_arguments +
explicit_call_vote + localized_focus + negative_associations +
negative_namecalling + negative_tone2 + politeness_and_civil_tone +
positive_framing_and_value_alignment + positive_namecalling +
positive_tone2 + preemptive_objections + reciprocity_and_mutual_benefit +
relatable_hypotheticals + social_proof_and_normative_influence +
stimulate_anger + stimulate_enthusiasm + storytelling_and_relatable_anecdotes +
transfer_of_association + use_of_negative_testimonials + use_of_positive_testimonials +
conditionZ * preZ * (strategy_nofactsZ + strategy_genericZ + strategy_noinstrZ) +
accuracyS
```
compare with [[250513_154015 canada - strategies#lasso|canada lasso results]]
![[1747776548.png]]
all coefficients below
- `b`: lasso coef
- `presence`: mean strategy use/presence
```r
strategy b presence
<char> <num> <num>
1: preZ 17.29355229 NA
2: conditionZ 12.17501893 NA
3: Optimism & value alignment 1.75279043 0.92056185
4: Contrast positives & negatives 1.25382997 0.33120869
5: Relatable hypotheticals 0.76313792 0.43425401
6: Localized focus 0.74994704 0.30169972
7: Highlight own side positives 0.65262406 0.61732767
8: Personalization 0.59309923 0.68106704
9: Positive associations 0.55681410 0.71883853
10: Evidence & facts 0.46117920 0.51463645
11: conditionZ:preZ:strategy_nofactsZ 0.43772285 NA
12: Reciprocity 0.38236143 0.20384797
13: Positive name-calling 0.35628873 0.75590179
14: Negative name-calling 0.30958714 0.11779981
15: preZ:strategy_nofactsZ 0.21285561 NA
16: Build rapport 0.01017216 0.56303116
17: Being pushy 0.00000000 0.04567989
18: Cognitive elaboration 0.00000000 0.45432011
19: conditionZ:preZ:strategy_genericZ 0.00000000 NA
20: conditionZ:preZ:strategy_noinstrZ 0.00000000 NA
21: conditionZ:strategy_genericZ 0.00000000 NA
22: Explicit call to vote 0.00000000 0.12924929
23: Negative associations 0.00000000 0.31314920
24: Highlight opposition negatives 0.00000000 0.01923985
25: Politeness & civility 0.00000000 0.98677998
26: preZ:strategy_genericZ 0.00000000 NA
27: preZ:strategy_noinstrZ 0.00000000 NA
28: Stimulate anger 0.00000000 0.09183192
29: strategy_genericZ 0.00000000 NA
30: strategy_noinstrZ 0.00000000 NA
31: Negative testimonials -0.04559935 0.02289896
32: strategy_nofactsZ -0.04821379 NA
33: conditionZ:strategy_noinstrZ -0.10206487 NA
34: Positive testimonials -0.13061619 0.03399433
35: Social influence -0.25236084 0.18508026
36: Stimulate enthusiasm -0.41778458 0.59926818
37: Emphatic listening -0.45489906 0.76133144
38: accuracyS -0.62143067 NA
39: conditionZ:strategy_nofactsZ -0.71814604 NA
40: Facilitate action -0.79163058 0.39719075
41: Storytelling -1.27169980 0.42481114
42: Preemptive counter-arguments -1.89952118 0.56857885
43: conditionZ:preZ -16.35187862 NA
strategy b presence
```