Founded in 1920, the NBER is a private, non-profit, non-partisan organization dedicated to conducting economic research and to disseminating research findings among academics, public policy makers, and business professionals.
Anonymous evaluation of "Does the Squeaky Wheel Get More Grease? The Direct and Indirect Effects of Citizen Participation on Environmental Governance in China" (Buntaine et al)
90% CI: (71, 86)
Quality scale rating
“On a ‘scale of journals’, what ‘quality of journal’ should this be published in?: Note: 0= lowest/none, 5= highest/best”
90% CI: (3.7, 4.8)
See the evaluation summary for a more detailed breakdown of the evaluators’ ratings and predictions.
This paper presents the results of a randomised control trial examining the impact of citizens reporting violations of pollution standard (water and air) by firms in China. Specifically, if a firm is allocated to a treatment arm and goes on to violate emissions standards, the violation is reported to the regulator by a citizen acting on behalf of the experimenters. They broadly find that these reports lead to the regulator intervening more often, firms polluting less, and these effects being particularly strong when the citizen report is made public through social media (T2 vs private reports, T1). Moreover, they are able to begin to elucidate the mechanisms by which these effects are bought about (various arms within T1 which vary who the report is sent to), and show that (at least within prefecture) this is not a zero-sum game (ie there is no evidence for substantial local leakage). Clearly, issues regarding air and water pollution are incredibly important, and this paper may offer a way for citizens to reduce the damage these cause by increasing compliance. In the sense that this paper answers a large and important question, with a well-thought through and implemented large-scale RCT of a low-cost intervention it is a potentially very important contribution.
While I think that this paper will publish very well even in its current draft, there are points where I think clarification - and less bold claims - are advisable in the writing and interpretation of the results. Similarly, the econometric approach is broadly well implemented, but I think recognition of the pitfalls (primarily regarding the SUTVA assumption) might be necessary. As noted in the additional comments regarding open science, it is a shame that the pre-registration included only details regarding treatments and nothing on how the data were to be analysed, but I recognise this obviously cannot be changed at this point. I organise my comments by theme (econometrics, generalisability, interpretation of results, ease of understanding) and within these themes order from more to less significant suggested changes.
[Evaluation manager: I copy the treatment abbreviations and descriptions from the original paper for clarification below, as the evaluator refers to these in their response]
Control Group (C): “When the CEMS data indicated that the firm violated its emission standards, we did not intervene in any way. About 1/7 of the CEMS firms were assigned to this group.”
Private Appeals Group (T1): “... a citizen volunteer filed a private appeal against that violation that was not observable by the public. About 5/7 of the CEMS firms …”
Public Appeals Group (T2): “ … wrote a post on Weibo … and “@” the official Weibo account of the corresponding local EPA. … We assigned 1/7 ….”
Private Appeals to Regulator via Direct Message on Social Media Group (T1A): “... sent a private message to the corresponding local EPA’s official Weibo account, notifying them about the pollution violation and requesting that they investigate ….”
Private Appeals to Regulator on Government Website Group (T1B): “... filed a private appeal via the 12369 website to the corresponding local EPA …
Private Appeals to Regulator through Government Hotline Group (T1C): “... called the 12369 hotline to privately appeal to the corresponding local EPA. In the phone call, she notified the local EPA …”
Private Appeals to Firm through Phone Call (T1D): “… called the violating firm to privately appeal the violation. In the phone call, she notified the firm about its violation and requested that they check the issue”
All prefectures contain some treated firms, but the intensity of treatment (70% or 95% of firms assigned to treatment) varies so that they can assess the “general equilibrium” effect of the treatment on non-treated firms in prefectures with a higher (95%) or lower (75%) intensity of treatment. This is clearly a very neat experimental design. However, the motivation for high/low intensity (“indirect effect”) clearly means that the SUTVA assumption across firms does not hold, such that the difference in outcomes between treated firms and control firms (those in the remaining 30/5%) captures the sum of the direct effect on treated firms and the indirect effects on control firms. I therefore find the comparisons within treatment arms (public vs private, the way private treatments are implemented), and comparisons of high vs low intensity prefectures more compelling. To better understand the impact of this SUTVA violation, it would be nice to see a plot of how control firm violations vary through time - ie does the onset of treatment lead to changes in control firms’ violations?
Relatedly, the claim that general equilibrium effects are estimated needs caveating as just local general equilibrium effects are mediated (at least as discussed in your paper) by the regulator’s capacity constraint. Of course, even including solely these local general equilibrium effects is still rather novel. But, a key problem in environmental pollution and policy (eg carbon markets) is the impact of (global) leakage (policy reduces output but relatively inelastic demand simply means that this shifts elsewhere). I don’t think that your estimates capture across prefecture leakage, and certainly wouldn’t include leakage beyond China’s border. Data on firm-level output might help mitigate this concern (ie if it shows supply actually is not constrained), and perhaps the evidence regarding the firms being fully operational is sufficient (but would need more discussion).
Much more minor comments that are easily addressed: 1) I think estimates are in effect intent to treat - the ongoing violation reporting outside of the experiment means control firms experience the “treatment” just less intensely - and this could be explicitly recognised. I think you could therefore consider using the treatment as an instrument for intensity of treatment (receiving a report conditional on a violation) in a 2SLS approach. 2) I think clustering of the standard errors in the firm-level analyses should account for the possibility errors are correlated within the citizen doing the reporting across firms/prefectures (ie some citizens may [randomly] have more or less impact across all the reports they send) 3) I think standard errors in table 6 need to be clustered at the firm level to account for multiple observations through time from different violations by the same firm.
On P1, you give evidence that lots of different countries have made it possible for citizens to report violations, yet present little evidence that citizens then actually do use this tool. You present evidence that in China citizens do this later (P2) but understanding if your results are relevant globally would be useful - ie do citizens outside of China regularly engage in such violation reporting? I note that even if the paper is only relevant to China, the potential impact/importance of the paper is still very large. Similarly, given your evidence and the relatively low cost of implementation, it brings into question why so much cash is being left on the table, and I think this could be discussed in the concluding remarks. Perhaps part of the reason for limited wider take-up is simply how rare the hourly and near real-time data China releases is. Alternatively, is it because reporting violations is privately costly but publicly beneficial, similar to punishment in PGG?
It seems to me that it is not possible to ascertain what the effect of citizen participation is through this experiment. Rather, the effects of the treatments compared with control conditions could be through some salience-of-information channel, and there is good evidence that information does matter in other environmental contexts (eg Saavedra 2023). Indeed, in the pre-registration document it reads as if there will be an information treatment separate from citizen reports (“Firms will be assigned to complaint, information, or control conditions”). This matters only for understanding what drives your results, and therefore what range of interventions might have led to similar outcomes. The impact of the Weibo likes could be 1) public pressure and therefore unrest or 2) information contained within the idea lots of people think this matters. On p27/28 you present evidence which suggests social media isn’t useful because of it involving more people per se conditional upon the severity of the violation. It would be nice to see in this observational data whether more non-treatment reports tend to be made when the violation is worse (I assume they are), in which case this would support the possibility that greater public engagement (likes, number of appeals) normally contains some information (even though in your treatment it obviously does not). Similarly, you claim that your results show “social media is an especially effective way to deliver public appeals” yet you have no comparison for delivering public appeals with anything other than social media.
Building on the idea of identifying the channels through which effects occur, my view is that you generally do a really good job of isolating channels, but this work could be explained better. As well as being better explained, I would have much appreciated you flagging that you will deal with the potential concerns when you mention the main effects. First, I was concerned about the potential for data manipulation, yet the paragraph on p26/27 details a range of robust evidence that that is not what is happening. You might also want to look at whether the data from treated and non-treated firms follows similar leading-digit distributions, applying Benford’s Law as per Cole et al. 2019 Climatic Change.
Second, the public vs private effect could operate through several channels:
(1) firm knowledge;
(2) firms feeling consumer pressure;
(3) public pressure on regulators;
(4) speed through the system to the local EPA rather than through the central report body;
(5) novelty of social media reports;
(6) the role of central government.
(1) can be explored through comparison to the effect observed in T1D and T1C*T1D and would appear to not be driving the result (but see comment in the understanding section RE common knowledge); (2) is dealt with through T1C and the evidence regarding whether the business is a final product producer; (4) seems unlikely given the comparator of results under T1A; (6) is dealt with by the subset of T1A which receive a threat of central government follow-up. Which leaves (3) and (5). The former is what you argue to be driving the effect - supported by the “likes” treatment, while I discuss the data needed to show if there is extra “novelty” of social media in the next paragraph.
At the moment, it is unclear how much the treatment changes the probability that a violation is reported by a citizen. In text, the paper mentions ~300k reports during the treatment period but later (P18) suggests just ~5.5k of those are actually applicable (identifying a specific violation etc). Understanding how these relevant non-treatment induced reports are split would be good: what number are private vs public? Individual citizen or NGO made? The same individuals making lots of complaints or many individuals infrequently? Finally, how are these 5478 valid appeals distributed across the 5366 real violations that occur during the treatment period - perhaps a histogram of the number of violations by the number of non-treatment valid appeals that they get would be useful.
You claim that the fact that effects persist (and if anything appears stronger) in the medium-term (ie at the end of the 8-month treatment period) suggests that if the treatments were implemented in the long term the effect would remain. This seems a little challenging. One could imagine it getting stronger if there are long-run adaptations they make, or imagine it weaken if the recipients simply get used to experiencing a high volume of appeals.
It would perhaps be good to discuss how your results fit in the wider literature regarding the impact of mandatory disclosure laws to citizens on firm behaviour (eg Bennear and Olmstead 2008 JEEM). Perhaps they would also benefit from additional consideration of how a party might reasonably increase the number of appeals - eg could making the algorithm that you developed to identify the cases of violations be useful? (which links into my previous comment RE discussing why so much cash might be being left on the table).
These comments are much more minor, in order in which they appear in the paper, and I think in general personal taste issues:
When you first use the term “appeals” early in the paper (abstract, early intro) it is unclear what this means - I would define as “appeals in the forms of reports made to the regulator if a violation of an emission standard occurs”.
MEE and MEP are confused in the first parts of the paper before Footnote 13 comes in and clarifies MEE replaced MEP
Unclear early on why there are multiple regulators - at that stage we don’t know there is a central monitor and many local regulators
I would like to see a diagram of the different actors in this space and their roles (MEE/MEP recording and publishing data, operating the hotline; local EPAs beneath this policing firms; firms; citizens). Perhaps this diagram could also include the incentives they face (official and unofficial) - eg are the local EPA legally required to investigate every violation? And defined over what timescale (hourly or daily violations etc)
How you define daily violations given the data and standards seem to be at the hourly rate is not quite clear (I assume that it is if they had any violation in a 24 hour period)
Figure 1 - would be good to have an additional plot which is a histogram plot of counts of the number of stacks/firms by number of violations that they commit. Fig 1 horizontal axis labels could also be better formatted as “1st Jan 2018” etc rather than year end and YY/MM/DD
Figure 4 - I think this would be better done as a Sankey plot. Using the same bins for where firms start, then track where the individual firms are in the distribution at the end. (At present, it is unclear if all firms are reducing their emissions a bit, or if the violation firms are reducing their emissions loads and the other firms changing very little).
Common knowledge is mentioned, but it is not clear if you’re using this in the precise definition (the firm knows, the regulator knows, and each know that the other party knows) or in a looser way (they each know, but are not specifically informed that the other party knows). If it is the former, please explain that each party was additionally informed that the other party knew (in T1C*T1D) or if the latter please use a different term.