1. Behavior in blind environmental dilemmas - An experimental study Martin Beckenkamp Max-Planck-Institute for the Research on Collective Goods Bonn – Germany beckenk@coll.mpg.de

2. Overview Introduction: –Many environmental problems are environmental dilemmas –Environmental dilemmas often are blind dilemmas My hypothesis: Blind dilemmas are the most tragic dilemmas Experimental setup and results Discussion and policy implication

3. Introduction: Most environmental dilemmas are blind dilemmas Many environmental problems are environmental dilemmas:  Hardin (1968) “Tragedy of the commons”.  Ostrom et al. (2002) “Drama of the commons”.

8. Introduction: Most environmental dilemmas are blind dilemmas Above that: Stakeholders in an environmental dilemma often are not aware of their social interdependencies.  Minimal social situation  Environmental dilemmas are blind dilemmas

9. Minimal social situations A game with incomplete information Players do only know their own strategy sets and payoffs. In its most extreme form, players are even oblivious of the fact that their decisions are choices in a game or strategy cf. Coleman, 2005, p. 217

10. Minimal social situations Many argue that due to Pavlov-strategies or “win stay, lose shift” subjects learn cooperation in prisoners’ dilemmas with minimal information (Colman, 2005, p. 222). My general hypothesis is contrary to that: Minimal information leads to high defection-rates.

11. Cooperation in minimal social situations? But how can Pavlov-strategy results be integrated with my hypothesis?  Other design: information about the payoff-matrix (i.e., not really blind).  Confounding of games: mutual fate game (Kelley, 1968).  Negative payoffs or even shocks.

12. Minimal social situations Minimal social situations: Sidowski, Wyckoff, & Tabory, 1956, Sidowski & Smith, 1961, Sidowski, 1957 Sidowski (1957) posed the question whether social behavior can be explained by conditioning processes. Ss were divided into two major groups, an informed and an uninformed group. The Informed group was told that another subject was serving in the experiment. The uninformed group was not told that subjects were in a social situation at all. The two major groups were subdivided into three groups: (a) shock only, (b) score only, and (c) both shocks and scores. Learning under these conditions occurred regardless of whether or not the subject was told that another subject was serving in the experiment. Furthermore, the results indicate that scores are necessary for learning.

13. Minimal social situations Minimal social situations: Sidowski, Wyckoff, & Tabory, 1956, Sidowski & Smith, 1961, Sidowski, 1957 In Sidowski & Smith (1961), a “paced” exchange of rewards and punishments was introduced, so that these results can directly be compared with other results concerning the mutual fate game. Here again, learning of giving scores could be observed.

14. Minimal social situations Mutal fate control and mutual behavior control: Kelley (1968) Kelley (1968) claims that following a simple strategy called win- stay-lose-change (i.e., Pavlov) leads to mutual cooperation in the mutual fate control-game, if the game is played simultaneously, and not if the game proceeds in an alternating sequence. He claims that social processes proceed without communication and information about social interdependencies. He also points out that it is not enough to reduce all incentives to positive and negative values, because the accommodations vary as a function of positive or negative consequences involved.

15. Minimal social situations Mutal fate control and mutual behavior control: Kelley (1968) With Mintz’s (1951) panic task, a rather concrete transformation of the prisoners’ dilemma, he demonstrated that subjects do create jams without there being any danger, but that varying the magnitude of the concern about the negative result increases the amount of incoordination. However, theses conclusions refer to a prisoners’ dilemma with full information and partial communication, so that they do not support expectations in my experiment.

16. Minimal social situations Nicklisch, 2006 Experiment on learning with minimal information. Subjects get information about their own payoff, but not about the structure of the game - the mutual fate control game and the fate control/behavioral control game. In contrast to Oechsler & Schipper (2003), subjects were even not informed about the structure of the payoff matrix, but in the treatment with rich information they received a feedback about the others’ choice after each round, whereas in the treatment with minimal information such a feedback was not given. Due to the very simple structure – games with payoffs of either 1 or 0 – Nicklisch’s results cannot be generalized to my experiment.

17. To summarize the main issue of the experiment: Does it make a difference whether participants know that they are playing with another person? In many environmental dilemmas, the agents are unaware of the interdependence of their actions. It would be expected that their decisions change once they know about the actual social interdependencies in the situation. Method and Design

18. The experiment consists of four treatment groups to clarify the issue: The impact of reducing information about social interdependencies in a prisoners’ dilemma. Method and Design

19. Method and Design: Entscheidungssituation der anderen Person 19 groups, 38 subjects Control Group

20. Method and Design: Entscheidungssituation der anderen Person 19 groups, 38 subjects Treatment group 3 124 I choose A I choose B AB 80

21. Minimal social situations Oechsler & Schipper, 2003 They compare three different 2x2 games: stag-hunt, unique mixed- strategy game and prisoners’ dilemma. With respect to the prisoners’ dilemma: 36 subjects, partner design, 15 “explorative” and 5 final rounds (with higher stakes). The subjects received the common payoff matrix for prisoners’ dilemmas, but without any information about the payoff of the matched partner. 8 of 18 pairs perfectly played the Nash-equilibrium in the final 5 rounds, i.e., instead of cooperating, these pairs of subjects defected completely. 14 of the 18 pairs played the Nash-equilibrium for at least 4 of the 5 final rounds. Only one pair managed to cooperate for 3 rounds with cooperation breaking down in the last two rounds – probably due to endgame effects.

22. either 12 or 4 I choose A I choose B either 8 or 0 Method and Design: Entscheidungssituation der anderen Person 19 groups, 38 subjects Treatment group 2

23. Method and Design: I choose A and get 8 or 0 Taler. I choose B and get 12 or 4 Taler. 19 groups, 38 subjects Treatment Group 2

24. Method and Design: Entscheidungssituation der anderen Person 18 groups, 36 subjects Treatment group 1 I choose A I choose B

25. Method and Design: I choose A I choose B 18 groups, 36 subjects Treatment Group 1

26. Method and Design control group treatment group 2 40 periods with partner design in each treatment Ring-Measure of Social Values afterwards Experiment programmed in z-tree (Fischbacher 2007) Subjects recruited with ORSEE (Greiner 2003) treatment group 1 treatment group 3

27. Results control group treatment group 2 Cooperation rates 33.0% 14.7% 48.8% 87.4% N 1520 1440 1520 1440 treatment group 1 treatment group 3

28. Typical progessions

29. treatment 1 treatment 2treatment 3Control Results

30. Random-effects logistic regression Number of obs = 6000 Group variable: ID Number of groups = 150 Random effects u_i ~ Gaussian Obs per group: min = 40 avg = 40.0 max = 40 Wald chi2(4) = 194.50 Log likelihood = -2320.0605 Prob > chi2 = 0.0000 ------------------------------------------------------------------------------ decision | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- treat_1 | 4.103983.4528463 9.06 0.000 3.216421 4.991546 treat_3 | 5.590593.4573615 12.22 0.000 4.694181 6.487005 treat_2 | 3.315607.4509062 7.35 0.000 2.431847 4.199367 period |.0223191.0033893 6.59 0.000.0156761.0289621 _cons | -3.707409.3460007 -10.72 0.000 -4.385558 -3.02926 -------------+---------------------------------------------------------------- /lnsig2u | 1.139228.1426774.8595859 1.418871 -------------+---------------------------------------------------------------- sigma_u | 1.767585.1260972 1.536939 2.032843 rho |.4870981.0356456.4179337.5567602 ------------------------------------------------------------------------------ Likelihood-ratio test of rho=0: chibar2(01) = 1680.94 Prob >= chibar2 = 0.000 Fitting comparison model: Iteration 0: log likelihood = -4140.9238 Iteration 1: log likelihood = -3191.8705 Iteration 2: log likelihood = -3160.9759 Iteration 3: log likelihood = -3160.53 Iteration 4: log likelihood = -3160.5298 Fitting full model: tau = 0.0 log likelihood = -3160.5298 tau = 0.1 log likelihood = -2727.979 tau = 0.2 log likelihood = -2558.3806 tau = 0.3 log likelihood = -2467.3481 tau = 0.4 log likelihood = -2410.6021 tau = 0.5 log likelihood = -2372.45 tau = 0.6 log likelihood = -2349.339 tau = 0.7 log likelihood = -2338.3992 tau = 0.8 log likelihood = -2358.9435 Iteration 0: log likelihood = -2335.23 Iteration 1: log likelihood = -2320.5963 Iteration 2: log likelihood = -2320.0618 Iteration 3: log likelihood = -2320.0605 Iteration 4: log likelihood = -2320.0605

31. Results Random-effects probit regression Number of obs = 6000 Group variable: ID Number of groups = 150 Random effects u_i ~ Gaussian Obs per group: min = 40 avg = 40.0 max = 40 Wald chi2(4) = 203.62 Log likelihood = -2323.9209 Prob > chi2 = 0.0000 ------------------------------------------------------------------------------ decision | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- treat_1 | 2.265402.246947 9.17 0.000 1.781394 2.749409 treat_3 | 3.064809.2471928 12.40 0.000 2.58032 3.549298 treat_2 | 1.809095.2454892 7.37 0.000 1.327945 2.290245 period |.01288.0018634 6.91 0.000.0092278.0165321 _cons | -2.021292.1856346 -10.89 0.000 -2.385129 -1.657455 -------------+---------------------------------------------------------------- /lnsig2u | -.0323099.1379008 -.3025904.2379707 -------------+---------------------------------------------------------------- sigma_u |.9839749.0678454.8595939 1.126353 rho |.4919232.0344662.4249244.5592135 ------------------------------------------------------------------------------ Likelihood-ratio test of rho=0: chibar2(01) = 1672.28 Prob >= chibar2 = 0.000 Fitting comparison model: Iteration 0: log likelihood = -4140.9238 Iteration 1: log likelihood = -3185.2442 Iteration 2: log likelihood = -3160.1457 Iteration 3: log likelihood = -3160.0605 Fitting full model: rho = 0.0 log likelihood = -3160.0605 rho = 0.1 log likelihood = -2512.8503 rho = 0.2 log likelihood = -2398.2231 rho = 0.3 log likelihood = -2355.4634 rho = 0.4 log likelihood = -2337.9334 rho = 0.5 log likelihood = -2341.9181 Iteration 0: log likelihood = -2338.1462 Iteration 1: log likelihood = -2324.2204 Iteration 2: log likelihood = -2323.9209 Iteration 3: log likelihood = -2323.9209

32. Results Random-effects GLS regression Number of obs = 6000 Group variable: ID Number of groups = 150 R-sq: within = 0.0076 Obs per group: min = 40 between = 0.5197 avg = 40.0 overall = 0.2955 max = 40 Random effects u_i ~ Gaussian Wald chi2(4) = 202.59 corr(u_i, X) = 0 (assumed) Prob > chi2 = 0.0000 ------------------------------------------------------------------------------ decision | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- treat_1 |.544481.0610636 8.92 0.000.4247985.6641635 treat_3 |.7276316.0602328 12.08 0.000.6095775.8456856 treat_2 |.3868421.0602328 6.42 0.000.268788.5048962 period |.0024853.0003722 6.68 0.000.0017558.0032148 _cons |.0747092.0432691 1.73 0.084 -.0100967.1595151 -------------+---------------------------------------------------------------- sigma_u |.25722097 sigma_e |.33281305 rho |.37395407 (fraction of variance due to u_i) ------------------------------------------------------------------------------

33. Results Random-effects GLS regression Number of obs = 6000 Group variable: ID Number of groups = 150 R-sq: within = 0.0076 Obs per group: min = 40 between = 0.5197 avg = 40.0 overall = 0.2955 max = 40 Random effects u_i ~ Gaussian Wald chi2(5) = 1019.68 corr(u_i, X) = 0 (assumed) Prob > chi2 = 0.0000 (Std. Err. adjusted for clustering on ID) ------------------------------------------------------------------------------ | Robust decision | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- treat_1 |.544481.0625486 8.70 0.000.421888.667074 treat_3 |.7276316.0460953 15.79 0.000.6372864.8179768 treat_2 |.3868421.0540628 7.16 0.000.280881.4928032 period |.0024853.0004062 6.12 0.000.0016891.0032815 _cons |.0747092.0286694 2.61 0.009.0185181.1309002 -------------+---------------------------------------------------------------- sigma_u |.25722097 sigma_e |.33281305 rho |.37395407 (fraction of variance due to u_i) ------------------------------------------------------------------------------

34. Results Random-effects GLS regression Number of obs = 6000 Group variable: ID Number of groups = 150 R-sq: within = 0.0076 Obs per group: min = 40 between = 0.5197 avg = 40.0 overall = 0.2955 max = 40 Random effects u_i ~ Gaussian Wald chi2(5) = 692.31 corr(u_i, X) = 0 (assumed) Prob > chi2 = 0.0000 (Std. Err. adjusted for 75 clusters in group_ID) ------------------------------------------------------------------------------ | Robust decision | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- treat_1 |.544481.0728862 7.47 0.000.4016267.6873353 treat_3 |.7276316.0756018 9.62 0.000.5794548.8758084 treat_2 |.3868421.1008136 3.84 0.000.1892511.5844331 period |.0024853.0009502 2.62 0.009.000623.0043476 _cons |.0747092.0605195 1.23 0.217 -.0439069.1933253 -------------+---------------------------------------------------------------- sigma_u |.25722097 sigma_e |.33281305 rho |.37395407 (fraction of variance due to u_i) ------------------------------------------------------------------------------

35. Results Random-effects GLS regression Number of obs = 6000 Group variable: ID Number of groups = 150 R-sq: within = 0.0076 Obs per group: min = 40 between = 0.5197 avg = 40.0 overall = 0.2955 max = 40 Random effects u_i ~ Gaussian Wald chi2(5) = 692.31 corr(u_i, X) = 0 (assumed) Prob > chi2 = 0.0000 (Std. Err. adjusted for 75 clusters in group_ID) ------------------------------------------------------------------------------ | Robust decision | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- treat_1 |.544481.0728862 7.47 0.000.4016267.6873353 treat_3 |.7276316.0756018 9.62 0.000.5794548.8758084 treat_2 |.3868421.1008136 3.84 0.000.1892511.5844331 period |.0024853.0009502 2.62 0.009.000623.0043476 _cons |.0747092.0605195 1.23 0.217 -.0439069.1933253 -------------+---------------------------------------------------------------- sigma_u |.25722097 sigma_e |.33281305 rho |.37395407 (fraction of variance due to u_i) ------------------------------------------------------------------------------

36. Results Random-effects GLS regression Number of obs = 6000 Number of groups = 150 (Std. Err. adjusted for 75 clusters in group_ID) ------------------------------------------------------------------------------ decision | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- treat_1 |.544481.0728862 7.47 0.000.4016267.6873353 treat_3 |.7276316.0756018 9.62 0.000.5794548.8758084 treat_2 |.3868421.1008136 3.84 0.000.1892511.5844331 period |.0024853.0009502 2.62 0.009.000623.0043476 _cons |.0747092.0605195 1.23 0.217 -.0439069.1933253 -------------+---------------------------------------------------------------- R-sq: within = 0.0076 between = 0.5197

37. Discussion In social dilemmas (!): –Information matters! Information about social interdependencies may lead to higher cooperation rates. –Information about expected payoffs without knowledge about social interdependencies may reduce cooperation. –Political implications: Get people out of their veil of ignorance.