B.1 Robustness checks

In order to control whether our results are driven by specific coding procedures or specifications, we have explored alternative solutions. First, as far as the econometric specification is concerned, we have also run probit and linear probability models, and the results were qualitatively identical.

Second, we also proposed an alternative coding for Justice Bazy-Malaurie. In fact, while Justice Bazy-Malaurie was originally appointed by Bernard Accoyer (right-wing), she was reappointed by Claude Bartolone (left-wing) at the beginning of 2013. We have therefore computed two versions of our indicators: one in which Justice Bazy-Malaurie remains a right-wing appointed Justice and one in which she switches to a left-wing appointed Justice after her reappointment. Results were nearly identical, since the correlation coefficient was close to 1.

B.2 Autocorrelation

A potential bias could emerge in our estimations if the error terms were correlated over time. Indeed: even if our dataset cannot be seen as a true time series (some days have more than one decision), the structure is nevertheless very close. This very special structure allowed us to create variables close to lag values of the dependent variable: previous is the average censorship rate of the last decision day.

Biases could emerge in our estimation if the error terms were correlated over time. To deal with this potential issue, we now consider the data as a time series, in which the decisions’ publication order corresponds to the time component. We replace the variable previous by the lag value of the dependent variable (L.censorship). When estimating this new model, we estimate a coefficient associated with L.censorship equal to −0.4465. This coefficient is statistically different from zero at the 5 % level, and statistically not different from what we found for previous. The coefficients associated with both variables have indeed the same sign, the same significance level, and the same magnitude.

Second, to detect potential serial correlation between our error terms, we compute the residuals of this last model (without clustering). The correlation coefficient between the error term and the lag value of the error term is equal to −0.008, which suggests no serial correlation. Moreover, graph 6 plots the error terms with the lag values of the error terms. The graph detects no autocorrelation. ^[23]

Third, in order to run the Breusch-Godfrey test, we run the previous model in a linear form (OLS regression) without clusters. ^[24] The probability of rightfully rejecting the null hypothesis of the Breusch-Godfrey test is equal to 0.7129. The test fails therefore at rejecting the no-serial correlation hypothesis, which confirms results presented in the article.

B.3 Logit diagnosis

Apart from collinearity, which was addressed in the above discussion, we now present several tests to check the validity of our estimations. As a baseline, we will focus on the specification of column 4 of Table 6 (1995–2013 time period with a linear trend). To check the robustness of our specification, we proceed to several tests as mentioned by Peng et al. (2002).

First, we proceed to a link test, which aims at testing whether our model is well specified. It regresses the censorship decisions on the predicted values and the square root of the predicted values. If the predicted values turn out to be significant, this entails that the model is not entirely misspecified. If the square root of the predicted values is highly significant, this implies that the model misses some important independent variables. Running the link test on the logistic regression of column 4 yields a significant coefficient for the predicted values (at the 1 % level), and to a non-significant coefficient for the square value of the predicted values at the 10 % level. This result confirms the quality of our estimation.

Second, we propose to run Hosmer and Lemeshow’s goodness-of-fit test. This test aims to measure the match between the predicted probabilities and the binary outcomes. Failing to reject the null hypothesis supports the empirical model. In our case, the probability of rightfully rejecting the null hypothesis is equal to 0.1829, which validates our estimation.

Moreover we propose to evaluate the prediction probabilities of our logit estimation, using the ROC curve analysis. The underlying idea of this instrument is to measure the number of false positives and of false negatives that the estimation produces. The explanatory power of the regression to discriminate between censorship and validation decisions is represented by the area under the curve of graph 7 in the appendix. As we can see, the area under the curve is close to 0.8, which indicates a good quality for the predicted values.

Last but not least, we look at the possibility that a few outliers drive our results. To do so, we plot three statistics associated with each observation: the Pearson residuals (Figure 8), the deviance residuals (Figure 9), and the Pregibon leverage (Figure 10).

Figure 5:

Composition of challenged laws.

Figure 6:

Plot of residuals and lag residuals.

Figure 7:

ROC curve analysis.

Figure 8:

Standardized pearson residuals.

Figure 9:

Deviance residuals.

Figure 10:

Leverage.

As one can see from Figure 10, three observations stand out (352, 491, 492). In order to verify that these outliers do not drive our results, we run our baseline model without these observations. Results were not affected by the drop of these three observations, which supports the previous findings.