Note: one participant wasn’t getting a partner, so they opened up a new tab, fiddled with the id and played themselves. So we exclude the game with player id YSB8RYRgF2tQjym2e, which is game AQXAv4FKrZxBrTQgE
Each game has:
4 “spiked” BoS trials where one of the rewards is high (25-30) and the other is normal (3-7)
~16 normal BoS trials where both rewards are from [1-9]
~10 normal (“easy”) PD trials where cooperating is welfare maximizing ( 2 * coop payoff greater than or equal to defect payoff). All rewards are chosen from [1-12] respecting this.
~10 “sacrifice” (“hard”) PD trials where having one person defect and the other take the sucker payoff is welfare maximizing (the defect payoff is > 2 * coop payoff). All rewards are chosen from [1-12] respecting this.
53 games in each condition
## # A tibble: 2 × 2
## chat_cond n
## <chr> <int>
## 1 chat 53
## 2 no_chat 53
## # A tibble: 2 × 6
## # Groups: game_cond, chat_cond [2]
## game_cond chat_cond no yes `NA` pct
## <chr> <chr> <int> <int> <int> <dbl>
## 1 oct2024 chat 15 90 1 0.857
## 2 oct2024 no_chat 19 86 1 0.819
mostly think partner is human, oh good.
People do talk to each other a little in the pre-game chat time.
Note that BoS has a lower points range than PD because of what range of random numbers is selected.
Looks like chat does better in spike, and a little bit on normal BoS?
Not a lot of talking.
Looks like mostly the actually using the chat is what’s helping.
Looks like using language is correlated with better outcome with BoS and slightly better outcome for PD. But if this language helping or “people who avail themselves of the option to use language are more competent”.
Is there a dose-response relationship, or is one word enough?
Especially where we have more data, looks like one word is enough. Indicative of coordination rather than negotiation, probably?
One idea is that talking on some trials may set up coordination strategies that can then effectively be used on later trials without talking on those trials.
So we want to look at overall volume of talking (in words or in # of trial talked) as a predictor for performance, controlling for talk on that round?
Looks like talking on other trials might help in BoS if you didn’t talk on this specific trial? But might just be fitting to outliers? Will need models.
In BoS: P1 prefers AA to BB, P2 prefers BB to AA. AB and BA are 0 for both.
Near chance if you can’t talk, or if you don’t talk, far above chance if you do coordinate.
In easyPD: P1 prefers BA > AA > BB > AB and P2 prefers AB > AA > BB > BA. AA is welfare maximizing.
Can get a reasonable option no matter what.
In hard PD: P1 prefers BA > AA > BB > AB and P2 prefers AB > AA > BB > BA. AB/BA is welfare maximizing.
So if you do use the chat, you tend to get the best option. (Using chat
means that easy and hard PD look different)
Whereas if you don’t chat, they look more similar at least.
How much language?
Filter only for games that talked at least a little.
Second graph filters for trials that talked.
Even in games that talk, there aren’t that many trials where both people talk?
As an exploratory thing, what if we look at the people who talked a lot or a moderate amount
high = 40%+ of trials (10 games)
med = 10% - 39% of trials (10 games)
minimal = < 10% of trials (33 games)
no = couldn’t chat (53 games)
In easyPD: P1 prefers BA > AA > BB > AB and P2 prefers AB > AA > BB > BA. AA is welfare maximizing.
In hard PD: P1 prefers BA > AA > BB > AB and P2 prefers AB > AA > BB > BA. AB/BA is welfare maximizing.
So if you do use the chat, you tend to get the best option. (Using chat
means that easy and hard PD look different)