Shaping EU agencies’ rulemaking

Abstract

EU agencies have become important regulatory venues. Initially established to provide expert advice, many have gained far-reaching decision-making and enforcement powers. This has attracted considerable attention from stakeholders, but the extent of their influence on EU agency conduct has remained a black box. We employ a novel dataset of 203 consultations (2007–2017) containing 26,468 attempts of stakeholders to change proposed regulatory rules by the European Union Aviation Safety Agency, as well as the agency’s response. This dataset allows for an original approach to measuring influence by linking influence attempts to rule changes. We found that business interests are far more influential than diffuse public interests. This has important implications for the legitimacy of EU agency stakeholder policies, as they are meant to make EU agencies more broadly accessible. National regulators are also influential in EU agency consultations, pointing to the unacknowledged importance of stakeholder consultations for EU-national regulator interactions.

Appendix

A: CRDs that are not in the analysis

See Table 5.

Table 5 CRDs that are not in the analysis

B: Examples of comments that were ‘noted’

In CRD 2007-01, ART indicated. ‘AMC 25.1709 point 4 Compliance summary page 51 The intent is not to examine each individual wire and its relation to other wires. Rather, it is to ensure that there are no hazardous combinations. In this sentence, please clarify if hazardous means HAZARDOUS failure condition which must be demonstrated as extremely remote. If not, change the word in order to avoid confusion.’

In CRD 2008-18, IACA (International Air Carrier Association) indicated: ‘IACA agrees with the conclusion and recommendation of the rulemaking group for EASA to discontinue further rulemaking activity and terminate the 25.045 rulemaking task.’

In CRD 2008-18, AEA (Association of European Airlines) indicated: ‘AEA would like to stress its support the work performed by the rulemaking task 25.045, as well as share the opinion delivered by this group.’

In CRD 2011-20, Gatwick Airport indicated. ‘After discussions in the rulemaking groups London Gatwick supports this proposal.’

In CRD 2013-25, Air France indicated: ‘The rulemaking group has performed a great job. The PBN integration in the European regulation was quite challenging.’

In CRD 2013-09, the Aerospace Industries Association indicated: ‘AIA would like to request more detailed information on the assumptions within the EASA assessment in order to provide substantive comments on the economic impact. It is very likely that a predictive alert as defined in this NPA, which fully accounts for all of the considerations mentioned, would cost much more than the estimated cost provided in the RIA.’

In CRD 2016-13, Airbus indicated: Significant Latent Failure Airbus would like to remark that the word more in the definition of Significant Latent Failure may lead to confusion since it is always possible to combine a latent failure with multiple other failures or events to produce a Haz or Cat FC. A review of fault tree or dependence diagram will identify all latent failures and their contribution to the top event. Even those considered as having only a small influence will become significant latent failures with this definition since, combined with an unlimited number of other failuresevents a HazCat FC will occur.

In CRD 2016-13, ENAV indicated: ‘M1 Article 3 × Provision of ATMANS, airspace structure and flight procedure design, and ATM network functions It is not clear what part of Article 3 × the GM is referred to.’

In CRD 2016-16, IAOPA Europe indicated: ‘IAOPA Europe notes that the Agency considers that the scope of NPA 2016 16 is limited to the correction of editorial errors and the addressing of non controversial issues raised by EASA itself or stakeholders. Accordingly, no Impact Assessment has been included in the NPA. We consider that the Agency should clarify the intention of GM1 FCL.735.A and GM2 FCL to make it abundantly and unambiguously clear that this is an optional alternative for enhancing an MCC course to standards and levels appropriate for CAT operation and is not to be taken as a mandatory requirement for existing MCC course providers who elect not to choose this option.’

In CRD 2017-06, UK CAA indicated: ‘Thank you for the opportunity to comment on NPA 2017 06, Loss of control or loss of flight path during go around or other flight phases. Please be advised there are no comments from the UK Civil Aviation Authority.’

C: Robustness check of operationalisation

The coded randomly selected observations, to check whether agreed changes do indeed lead to changes in regulation, are available on figshare at https://doi.org/10.6084/m9.figshare.11860785.

Table 6 reports on a robustness check to see whether including partially accepted comments in our measurement would lead to different results.

Table 6 Regression results for influence, 3 models

The first robustness check shows what the results would have looked like if partially accepted was also scored as influence in the dependent variable (model 2). A multilevel logistic regression model was used, as in the main analysis. The second robustness check was performed with influence as a scale variable (model 3). This variable ranges from 1 (high influence) to 0 (no influence). To calculate the score, accepted comments were scored with 1, partially accepted comments were scored with 0.5 and not accepted comments were scored with 0. A multilevel linear regression model was used to analyse the results. The results of both robustness checks are displayed in Table 6 and Fig. 2.

Fig. 2

Predicted probabilities of having influence for each stakeholder type from four different operationalisations. (PA partially accepted, fi firm, ba business association, bu business interests, cg citizen group (air sport), cg-dif citizen group (diffuse), ra regulatory agency, lu labour union, pa professional association, ot other)

Using the different ways to approach the influence variable does not change the substantive results of this paper. When the partially accepted category was included as influence, the only difference is the lack of significance of the regulatory agency category. However, this category was only significant at a 0.1 level. Approaching influence as a scale variable leads to the same change in significance of the regulatory agency type. The size of the effects does increase when adding partially accepted in the operationalisation of influence and again when the variable is operationalised as a scale variable (see Fig. 2). This is the result of additional observations being counted as influence. The chance of having influence therefore becomes higher overall. But the different model specifications do not change the difference between types of stakeholder within each model specification, which is key to the analysis. The results of these robustness checks are therefore that the results in the paper are robust against different operationalisations of influence.

Additionally, we show whether separating results for business associations and firms have an effect on the results (Table 6, model 4; Fig. 2). The results in terms of predicated probability are nearly identical for all but the business interest group types.

D: Merging stakeholder types

The INTEREURO codebook was followed to code stakeholder types. To better fit the aim of our research, one category was split (see methods section). Furthermore, stakeholder types were compiled into an ‘other’ category. These stakeholders are EU agencies other than EASA, intergovernmental organisations, individuals, institutions and government or related. Individual refers to an actor that responded to the consultation with only their own name. It could not unambiguously be determined for these actors whether they represented their own personal interests or the interest of an organisation they were a member of or worked for. Individuals are therefore also part of the other category. Furthermore, we merged the two types of business interests; firms and business associations. The separate influence attempt rates of these groups are shown in Table 7.

Table 7 Influence attempt frequencies in ‘other’ category

E: Descriptive statistics of the control variables

See Tables 8 and 9.

Table 8 Descriptive statistics of (scaled) density variable

Table 9 Frequency table of EC regulation variable

F: Model fit statistics and random intercept statistics

Table 10 gives the model fit statistics of the two models presented in the paper. Two issues are noteworthy. As the χ² of model 2 is statistically significant, the addition of the dummy type stakeholder variables improves the model. This is however not the case for model 3, indicating that adding control variables did not lead to more explanatory power. Additionally, the interclass correlation coefficient points out how much of the variation is accounted for by nesting the observations using the cross classified multilevel model. As the model can account for 14 per cent of the variation by nesting at the consultation level and for about 7 per cent of the variation by nesting at the stakeholder level, the use of multilevel models is statistically supported. The variables that is accounted for by nesting at the year level is minimal (0.8 per cent). We however decided to still include it because this nesting structure accounts for the structure of the data.

Table 10 Model fit statistics and random intercept of two models in main analysis

G: Robustness check with years as variable instead of nesting level

See Table 11.

Table 11 Models with different use of the year variable

H: Robustness check with Bayesian model

As a robustness check, we performed the analysis using a Bayesian model. We used naïve priors for all regression coefficients (normally distributed, with a mean of 0 and a precision of 0.01). The model ran for 9000 iterations with a burn-in period of 1000 iterations. The model does not nest per year in order to ease convergence.

As is evident from Table 12, Figs. 3 and 4, the results were very similar to our main results.

Fig. 3

Posterior estimates of the variables. Dotted line represents the lack of an effect

Fig. 4

Predicted probabilities of having influence for each stakeholder type based the Bayesian model. (bu business interests, cg citizen group (air sport), cg-dif citizen group (diffuse), ra regulatory agency, lu labour union, pa professional association, ot other)

Table 12 Bayesian model estimates

I: Statistics for most frequent commenters

See Table 13.

Table 13 Participation, influence rate and total influence of the 20 most involved stakeholders