Interest Group Communication Strategies

Abstract

In this chapter, I test the formal model empirically. I analyze the decision to mobilize and the decision to send public or private messages. I can demonstrate that the decision to send public messages depends on the distance to the expected policy outcome, while the decision to send private messages depends on the distance to the constraining actor. This is based on an identification strategy which uses the fact that the German political system functions as if the relevant decision-makers are exogenously assigned to issues based on the constitution.

Appendix

Multilevel Models

Standard regression models assume that the errors of individual cases are independent from each other (e.g. Wooldridge 2002). In the dataset at hand, two measurements for a single interest group are potentially not fully independent from each other. The interest group may have a general predisposition to use one activity over the other—independent from the context. This assumption is the basis for studies who attribute the choice of activities to interest group characteristics (e.g. Gais and Walker 2001). The error terms for individual actors may therefore be correlated, thereby violating a major assumption of many statistical models, although this would be mitigated by the fact that the panel is unbalanced.

In addition, more than one interest group participates in the events. It is unlikely that the cases within an event are fully independent from each other. Certain contexts may cause actors to use specific actions so the correlation of interest groups tactics and therefore error terms for a given event is a possibility. For example, coalitions are a factor which influences interest groups’ choice at a given point in time. The resulting correlation of errors may be problematic for the empirical modeling of strategies as well.

Using a standard regression model which assumes independent errors on each observation is therefore not necessarily appropriate. A data structure where a serious correlation of errors for specific groups and across events is possible could be problematic. The structure of (potential) error correlation calls for a non-nested hierarchical model with varying intercepts. This type of model accounts for the interdependence of the error terms (Gelman and Hill 2007).

The multilevel modeling approach has another advantage. For the application of the model the data need not be balanced (Hox 2010, 79). This is particularly important as the dataset is an unbalanced panel. Not all interest groups participate in all legislative events. Biases caused by nonattendance are therefore not a problem in the multilevel models.

The general formula for the estimated models is

$$\begin{aligned} y_i \sim logistic (X_i\beta + \eta _{j[i]} + \mu _{k[i]}) \end{aligned}$$

$$\begin{aligned} \eta _{j[i]} \sim N(0,\sigma _{\eta }^2) \end{aligned}$$

$$\begin{aligned} \mu _{k[i]} \sim N(0,\sigma _{\mu }^2) \end{aligned}$$

where $\eta _{j[i]}$ denotes the effects for interest group j and $\mu _{k[i]}$ represents the effect for event k. I estimate these models using the glmer function of the lme4 package in R (Bates et al. 2013). Note that because I am now estimating the effect of individual events, I can no longer include the majority dummy in the model. Non-varying variables at the individual level can also no longer be included.

The effects are qualitatively comparable to the results of the standard logistic regressions. The distance to the expected policy outcome is negative and significant. The costs and the uncertainty have the expected effects.

Explicitly modeling the interdependencies in the error structure seems not to dramatically change the results. This suggests that the simple logistic regression models capture the essence of the situation and are sufficient for analyzing interest group behavior. This also suggests that strategy choice is neither driven by group characteristics (type) nor by context alone, but by an interaction of the two components (Table 6.5).

Table 6.5 Results of cross-classified hierarchical logistic regression

Table 6.6 Results of cross-classified hierarchical logistic regression

The results for the multilevel regressions on private messages are similar to the standard logistic regression. The model also predicts a positive effect of the distance to the compromising actor. The costs have the expected sign and also the uncertainty measure has a comparable effect. One interesting difference is the insignificance of the access term. This is most likely driven by the decomposition into individual events (Table 6.6).

Alternative Operationalization of the Dependent Variable

As a last robustness check, the strength of the signal is accounted for in the dependent variable. The variation I report here is the simple count of the activities. The model of choice for this type of dependent variable is a negative binomial regression model. Table 6.7 shows the results for a simple count of public alternatives used. The results for the count of private messages can be found in Table 6.8.

Table 6.7 Results of negative binomial regression for count of public activities

Table 6.8 Results of negative binomial regression for count of private activities

The results are very similar to the results of the standard logistic regressions. Particularly interesting is that the actors for which the distance is larger are more likely to send a strong (and thereby costly) public message. This is in line with the predictions of the model, that the range of costs for which an equilibrium exists is larger, if the distance to the expected policy outcome is large. The results for the private messages are also in line with the results of the simple logistic regression. The simple binary operationalization captures all relevant aspects of the strategic interaction.