The economics of politics: patronage and political selection in Italy

Abstract

This article investigates patronage in the Second Italian Republic by considering patronage a fundamental device able to guarantee a party presence in the governance of public bodies. The study sheds light on a particular area of party patronage, namely political appointments concerning legislators; it analyzes the factors which could determine whether a member of Parliament will be appointed to a state-owned enterprise’s board of directors after a legislature, seeking to gain a better understanding of how political actors exploit this opportunity. Direct political connections can be conceptualized as instruments to control and reward politicians and/or strategies to enhance political control over the bureaucracy. The empirical investigation suggests that legislators’ efforts in Parliament play a role in the likelihood of patronage appointments. Education does not seem to significantly increase the probability of receiving a nomination for a seat on public firms’ boards, moreover our result casts doubt on the merits or competencies of the appointed politicians.

Appendix A: Technical details

We start from a linear probability model (LPM) using a pooled ordinary least squares estimation

$$\begin{aligned} p(y=1|X_i) = \alpha + \beta X_i + \gamma lex \end{aligned}$$

where the vector of individual characteristics, $X_i$, and legislatures dummies (lex) are used to explain the phenomenon.²² In the linear probability model, the estimated parameters are the change in probability given a one-unit increase of the regressors. If the explanatory variable is binary, the related parameter is just the difference in the probability of success when it is equal to one with respect to the case where it is equal to zero, holding the other regressors fixed. We consider the linear probability model as only the starting point of the analysis. Since the OLS fitted value is an estimate of the conditional probability $P(y=1|x_i)$, it is troublesome if the predicted probability is negative or above unity. Aside from fitted values being outside the unit interval, the linear probability model implies that a ceteris paribus unit increase in the regressors always changes $P(y=1|x_i)$ by the same amount, regardless of the initial value of the covariate. Hence, increasing one of the explanatory variables would eventually drive $P(y=1|x_i)$ to be less than zero or greater than one. The fact that some predicted probabilities are outside the unit interval and that the linear probability model does not provide good estimates for extreme values of the regressors leads us to consider a nonlinear model as our primary specification (Wooldridge 2010).

Table 6 Linear predictions

Nonlinearity is considered more appropriate, given that about 25% of the values predicted by ordinary least squares estimation are negative (see Table 6). In doing so, we study our binary response model in the form $P(y=1|x)=G(x \beta ) \equiv p(x)$, where, therefore, the marginal effect of $x_i$ depends on x through the index $x \beta$, and where the G function maps the index into the response probability. We will show that LPM, Probit and Logit models give very similar results in terms of average marginal effects (Table 7). As shown in Table 8, Probit and Logit models also display almost equal measures of goodness of fit.²³ For this reason we present in the main text a Probit model as our primary specification.

Table 7 Marginal effects comparison

Table 8 Goodness of fit

Table 9 Twostage Heckman’s procedure

Although the dataset is strongly unbalanced we try to exploit its panel structure, even though 799 MPs out of 1407 are present only once in the sample legislatures. In LPM, the null hypothesis that all of the fixed effect intercepts are zero has to be rejected. In line with Wooldridge (2010), we test whether there is a statistical significant difference between the fixed effect estimator and the random effect estimator. In LPM random effect coefficients are not significantly different from fixed effect (Hausman’s error probability to reject null hypothesis when it is true is 0.379), thus there is an efficiency gain in using random effect. Allowing for heteroskedasticity and autocorrelation, we implemented another type of cluster-robust Hausman test based on bootstrapping, following Cameron and Trivedi (2005). The covariance matrix was estimated by bootstrap resampling over the individual identification variable (id) Using 100 bootstrap repetitions the p value of the test converges to 0.0157, suggesting the statistically significance of individual fixed effects. Finally, the Breusch and Pagan tests accept the null hypothesis that the variance of the random effect is zero, suggesting to use the OLS specification ($p=0.406$). Thus, there is no conclusive indication on which model has to be preferred with LPM.

Moreover the correct specification should be non linear, therefore we test for the presence of fixed effects following Mundlak (1978). We run an auxiliary regression including the individual means of all the time-varying covariates as additional explanatory variables, and then we run the model as a random effect Probit. Then we test the significance of the individual means through a Wald test. Assuming that fixed effects follow a normal distribution, and that they are linear combinations of the individual means times the related coefficients, in the nonlinear version of the model we would exclude fixed effects. In this case the p value is 0.9999.²⁴

Once fixed effects are excluded through the Mundlak approach, we consider the non-linear pooled analysis to be more appropriate than the random effects model. The fraction of variance due to the individual unobserved effect ($\rho$) is significantly equal to zero²⁵, therefore the possible gain in exploiting the dataset panel form results irrelevant, also considering the cost of the assumptions about the error term.²⁶

Finally, we also considered endogeneity issues. Endogeneity arose in our model in all of the three usual ways: omitted variables due to data unavailability, measurement error due to imperfect measures of some regressors, and simultaneity due to the possible reverse causality between the dependent variable and (one of) the regressors. What the dataset allows us to do is test the exogeneity of the retire variable through an instrumental variable approach, using the age_exit²⁷ variable as an instrument in the linear specification of the model. The reasoning behind testing the exogeneity of retire is quite straightforward. For the defeat case we assumed that it is highly unlikely to observe a reverse causality possibility: we assumed that the returns from being a legislator always outweigh those of becoming a public manager and, therefore, no legislator would lose an election in order to gain an appointment to a PPE board, once selected from the party to run again for Parliament. In the retire case, things could be different: a party may prefer to replace a legislator in the next electoral race, inducing her to withdraw in order to make room for a new candidate. In this event, the legislator and the party could agree in an “exit strategy”, therefore the appointment would be the cause and not the consequence of the legislator’s withdrawal. Using the age_exit variable as an instrument for retire seemed quite a straightforward solution, considering variables available in the dataset. Both the Durbin’s and Wu-Hausman’s tests accept the null hypothesis of considering the retire variable as exogenous (Durbin $p=0.5170$ and Wu-Hausman $p=0.5205$).

When we run the probit models on the subsamples, a selection bias problem could arise. When estimating the subsamples²⁸ with a two stage Heckman procedure (Table 9), the coefficient of inverse Mills ratio is not significant, thus no selection bias is present in subsample ones. Moreover the results are stronger in the retire sub sample.