The effect of regulatory complexity on individual voluntary climate action behavior: A representative online field experiment in Germany

The amount of true information supplied to subjects prior to their decision whether and when to retire an allowance from the EU ETS at a personal opportunity cost. The information describes the market stability reserve (MSR) and its implications on how the timing of retirement influences the actual amount of greenhouse gas emissions saved. The level of detail ("complexity") of the description differs across the five experimental conditions: There are two control condtions without reference to the MSR, and three treatment conditions that supply successively more information.

2021-06-11

2021-07-02

Control condition one of two (CC1): Decision between (1) a 5 EUR bonus, (2) retirement of an emission allowance from the EU ETS, and (3) a non-response option (three discrete choices).

Control condition two of two (CC2): Decision between (1) a 5 EUR bonus, retirement of an emission allowance from the EU ETS either (2a) immediately after the end of the field phase of the survey or (2b) one year later, and (3) a non-response option (four discrete choices).

Treatment condition one of three (TC1): As in control condition two of two.

Treatment condition two of three (TC2): As in control condition two of two.

Treatment condition three of three (TC3): Decision between (1) a 5 EUR bonus, retirement of an emission allowance from the EU ETS either (2a) immediately after the end of the field phase of the survey or (2b) at a yet to be determined date at least one year later, and (3) a non-response option (four discrete choices). The date of late retirement is determined by a scientific assessment of when the impact of retirement on total emission reduction within the EU ETS would be largest, given the current regulatory framework. This option is actually offered by several climate action NGOs in Germany.

The End Points are used to compute the following summary variables at the level of the experimental condition:

In all conditions: "Willingness to contribute" (WC), defined by the frequency of subjects that retire an allowance relative to all subjects in the condition (a percentage). "Decision avoidance" (DA), defined by the frequency of participants that select the non-response option relative to all subjects in the condition (a percentage).

In control condition two of two and all treatment conditions: "Total effectiveness of contributions" (TEC), defined by the frequency of subjects that opt for late retirement relative to all subjects in the condition (a percentage). "Relative effectiveness of contributions" (REC), defined by the frequency of subjects that opt for late retirement relative to the subjects in the condition that select to retire an allowance.

n/a

n/a

Subjects are sampled from the Socio-Ecological Panel (Green-SÖP, Wave 7), which is representative for the German population. Each subject completes a questionnaire, which will be published in full with the Green-SÖP Wave 7 dataset (http://fdz.rwi-essen.de/doi.html). The parts relevant for the present experiment are provided with annotations in the Docs & Materials section, both in the German original and machine translated (https://www.deepl.com/translator) form (for sake of replicability).

Subjects sampled for the experiment are randomly assigned to one of five experimental conditions: CC1 (B_Kontrolle 1 in the questionnaire), CC2 (B_Kontrolle 2), TC1 (B_Komplexität 1), TC2 (B_Komplexität 2), TC3 (B_Komplexität 3). In all conditions subjects respond to the pre-treatment questions ExpB0 and ExpB1. Depending on treatment assignment, subjects respond to question ExpB2_a (CC1), ExpB2_b (CC2), ExpB2_c (TC1), ExpB2_d (TC2), ExpB2_e (TC3). All questions involve a real decision whether - and if applicable when - to retire an allowance from the EU ETS at a personal opportunity cost. The amount of true information supplied to subjects prior to their decision is manipulated across conditions. The information describes the market stability reserve (MSR) and its implications on how the timing of retirement influences the actual amount of greenhouse gas emissions saved. The level of detail ("complexity") of the description differs across the five experimental conditions: There are two control condtions without reference to the MSR (CC1 and CC2), and three treatment conditions that supply successively more information (TC1, TC2, and TC3).

The design is motived by the following behavioral model:
Subjects have preferences for both private consumption $x$ and contributions to the public good $y$. We assume they are additive and represented by utility function $u$: $u(x,y) = x + \rho(T) \cdot v(y)$ with $v’ > 0$. $\rho(T) \geq 0$ with $T = CC1, CC2, TC1, TC2, TC3$ captures contextual aspects of the treatment specific choice situation (T) such as the degree of complexity involved. This implies that within treatment and hence subject $u(x_0,y_i) < u(x_0,y_j)$ iff $y_i < y_j$. Factually, the available $y_i$ are the same for all conditions except CC1 and TC3. However, in terms of the information available to subjects regarding the effectiveness of the options available there is also a difference between CC2 and TC1/TC2 as the latter explicitly rank the available cancellation options. Given that we hold $x$ constant across all treatments, differences in choice patterns between treatments reflect factual or perceived differences in the size of $y$ and choice context $\rho(T)$ induced by treatment variations. Note, that changes in $\rho$ only affect the choice between cancellation and private consumption, i.e. between $x$ and $y$ but not between the two cancellation options available in all conditions except CC1.

The experiment is designed to test how WC and REC are affected by revealing unexpected and potentially morally contested information on effects of voluntary climate action and the complexity of the explanations provided. To those that are willing to contribute the information provided in TC1-3 is relevant as it reveals a strict dominance in effectiveness at identical costs ($ y_1 < y_2$). Furthermore, the information is most likely surprising as the implications of the Market Stability Reserve (and indeed the mechanism’s mere existence) are known only to an expert circle. The common belief among supporters of climate actions is that (in particular at identical costs) early action is better than delaying it (UNEP 2021, Figure 3 & 6), i.e. that ($ y_1 > y_2$). Hence, at least some participants will update their beliefs and prefer late cancellations. Updating of beliefs over the relative effectiveness of early and late cancellation can be biased, because "bad news" are discounted in certain contexts (Eil & Rao 2011, Gershman 2019, Kuzmanovic et al. 2018, Yao et al. 2021). We therefore do not expect that all contributors choose the late cancellation option. The first key hypothesis is:

Hypothesis KH1: REC is larger in TC1-3 than in CC2.

However, there might be a trade-off between REC and participants’ willingness to contribute. The information might conflict with the moral belief that urgent action on climate change is a moral obligation. Hence, there might be an aversion to see this belief challenged or an urge to act against it (Bénabou & Tirole 2011). Moreover, if the dissonance or ambivalence induced is sufficiently strong, participants might be less willing to contribute at all in order to avoid having to make a choice that they are ambivalent about (Anderson 2003, Luce et al. 1997). In this case $\rho(TC1-3) < \rho(CC2 & CC1)$. The second key hypothesis is therefore:

Hypothesis KH2: WC is smaller in TC1-3 than in CC1 & CC2.

The bibliography is provided in the "Docs & Materials" section.

If both key hypotheses are confirmed, it is not clear a priori whether providing accurate but challenging and complex information improves the overall impact of voluntary climate action.

We therefore also evaluate further exploratory Hypotheses:
CC1 (binary choice) tests for the share of participants for whom $v(y_1) > x$, where $y_1$ is abating one ton of CO$_2$ and without loss of generality we normalize $\rho(CC1) = 1$. The underlying assumption is that the average participant assumes that the abatement action is taken immediately and indeed equal to one ton (note that neither is true but the ambiguities in these attributes is the same across treatments). CC2 (timing without ranking without explanation) introduces heterogeneity in $y$ by differentiating the timing of the intervention (cancellation of the emission allowance). It can be done either immediately ($y_1$) or in one year’s time ($y_2$). No explanation or information on consequences is given. Regardless of how a participant assesses $y_1$ relative to $y_2$, compared to CC1 a weakly dominated alternative is added. Hence, by the independence from irrelevant alternatives axiom:

Hypothesis EH1: WC is not significantly different in CC1 and CC2.

EH1 implicitly assumes that the change in the decision context is sufficiently small not to affect behavior, i.e. that $\rho(CC2) = \rho(CC1) = 1$. If adding complexity in the form of presenting an additional (weakly dominated) choice option changes the decision context in a relevant way $\rho(CC2) \neq 1$, then

Alternative Hypothesis to EH1: WC is significantly different between CC1 and CC2.

The extant literature supports both an increase and a decrease in WC. WC increases due to the asymmetric dominance effect if providing a similar but dominated cancellation alternative makes the dominant cancellation alternative more attractive than the outside option ($\rho(TC1)>1$). At least some explanations for the asymmetric dominance effect are based on decision maker’s desire to simplify the choice situation (Ariely & Wallsten 1995). However, if participants perceive the addition of the second cancellation option as "abatement" becoming more complex, they might be less likely to choose it ($\rho(TC1)<1$). Effectively CC2 adds a second layer to the abatement choice. First, participants need to decide whether or not to abate and in case they choose abatement they also need to decide how to abate. There is experimental evidence that increasing the number of stages in decision-making tasks reduces the attractiveness of options (Sonsino et al. 2002 ) and that participants tend to avoid cognitive demand (Kool et al. 2010). There is reason to expect that most subjects have the prior that it is better to protect the environment earlier (UNEP 2021), implying the hypothesis that $y_1 > y_2$ and hence that a majority of participants chooses to cancel early, i. e.:

Hypothesis EH2: The REC in CC2 is below 50%.

TC2 and TC3 (timing with ranking with explanation) feature an explicit ranking between abatement options $y_1 < y_2$ that is assumed to contradict the prior belief and moral inclination of the majority of participants. The explanation requires some careful reading and abstract thinking. Those that engage with the explanation (i.e. read and process it), are more likely to update their beliefs about the relative effectiveness of $y_1$ and $y_2$. As people tend to avoid cognitive demand (Kool et al. 2010), the effects of complexity are heterogeneous (Abeler & Jaeger 2015, Oprea 2020). Those with higher cognitive abilities (and hence lower marginal cost of engaging with the explanation) are more likely to read and process it. Moreover, updating of motivated belief can be impaired (Eil & Rao 2011, Gershman 2019, Kuzmanovic et al. 2018, Yao et al. 2021). Classifying subjects based on the responses to the pre-treatment questions regarding "urgency" and "moral duty to contribute" into "high" (above mid-point) and "low" (at and below mid-point) classes, we expect:

Hypothesis EH3: The size of the treatment effects TC2 vs CC2, TC3 vs CC2, TC2 vs TC1, and TC3 vs TC1 on REC differs between the "high" and the "low" classes regarding "urgency" and "moral duty to contribute".

If being confronted with a complex explanation that a participant decides to "ignore" instills disengagement (anger, frustration, feelings of inferiority, annoyance,…) or if the dissonance or ambivalence induced is sufficiently strong (Anderson 2003, Luce et al. 1997) then both cancellations options might be perceived as less appealing ($\rho(TC2 & TC3) < \rho(CC1 & CC2 & TC1)$). Classifying subjects based on the responses to the "cognitive abilities" question into "high" and "low" classes, we expect:

Hypothesis EH4: The size of the treatment effects TC1 vs CC2, TC1 vs CC2, TC3 vs CC2, TC2 vs TC1, and TC3 vs TC1 on WC differs between the "high" and the "low" classes regarding "urgency", "moral duty to contribute", and "cognitive abilities".

TC3 (complex timing with ranking with explanation) increases both complexity (in the form of adding ambiguity w.r.t. the exact timing and size of the difference in effects) and the intensity of the ranking between the two timing options. Compared to TC2 the ranking of alternatives for those that engage with the explanation is strictly stronger. The cognitive burden is also strictly stronger than in TC2. Hence, the same predictions as for TC2 & TC3 jointly apply and to be at least as strong.

The bibliography is provided in the "Docs & Materials" section.

Randomization done in office by a computer.

Individual subject randomization.

No

Since randomization is done at the individual subject level, there is a single treatment cluster (i.e. treatment is "non-clustered"). Statistical analyses will account for potential effect heterogeneity across spatial and temporal quasi-clusters defined by the region of residence and survey participation date by means of mixed effects modeling. The number of those quasi-clusters is endogenous.

4,000 individuals

Control condition one of two: 480 individuals
Control condition two of two: 640 individuals
Treatment conditions: 960 individuals each

From a cognitive pretest of conditions CC1 and CC2 with 6 subjects we have the following data: In CC1, 3 subjects retired the allowance, and the remainig 3 selected non-response. In CC2, 2 subjects retired the allowance immediately, 3 retired the allowance one year later, and 1 subject selected non-response. This yields WC = .500 and DA = .500 in CC1, and WC = .833, DA = .167, TEC = .500, and REC = .750 in CC2. Assuming REC = .750 (as in pretest CC2) and the planned sample sizes, the null corresponding to hypothesis KH1 is rejected at alpha <= .050 and beta <= .200 (power => .800) with two-tailed Fisher exact tests for inequality of proportions between independent groups at REC => 0.8097 in any of the treatment conditions. Assuming WC = .833 (as in pretest CC2) or WC = .500 (as in pretest CC1) and the planned sample sizes, the null corresponding to hypothesis KH2 is rejected at alpha <= .050 and beta <= .200 (power => .800) with a two-tailed Fisher exact tests for inequality of proportions between independent groups at WC <= 0.7756 in any of the treatment conditions vs CC2, and at WC <= 0.4218 in any of the treatment conditions vs CC1.