Proposed Alternatives to the RILE Index

As the RILE index has been the most popular use of the manifesto data and as most analysts are well aware of at least some of the criticism that has been pointed out, there is a range of measures

CEUeTDCollection

that have been proposed and that should improve over the flaws of the RILE. The following sections will introduce the most notable such alternatives.

Kim and Fording RILE (KFRILE)

Perhaps the simplest alternative was suggested by Kim and Fording (1998) and Kim and Fording (2002), who use the same subsets of categories for left and right as the RILE index, but employ a different logic for calculating the position of a manifesto. While the RILE index normalises the counts of political statements that belong either to the left or the right category with respect to the total number of political statements in a manifesto, the measure proposed by Kim and Fording does it with respect to the total number of political statements in the two subsets of left and right. The advantage of this is that it evaluates position with respect to the supposedly ideological part of the manifesto and not the manifesto as a whole. It is thus not dependent on issue categories that are ideologically irrelevant, which is one of the problems of the RILE index. The value of the index is calculated as shown in equation 3.2.

KFRILE = N_R−N_L NR+NL

(3.2)

Logit RILE (LRILE)

An improvement over the kind of measure Kim and Fording suggest is proposed in turn by Lowe et al.

(2011), who address the problem of the marginal effect of additional statements. Both the RILE and the KFRILE measures assume a fixed marginal effect for an additional political statement that is coded in the data set. Lowe et al. (ibid., p. 130) propose a decreasing marginal effect model, which entails working with proportions (of one category or set of categories to another) and a logarithmic scale. Their proposed measure, adopted into this left-right context, takes the form of a logarithm of the ratio of the number of right statements to left statements (0.5 is added to each count for methodological reasons (ibid., p. 132)). This logit scale has no predefined end points and any position on it is theoretically possible, given an extreme count in one of the categories. In their article they propose 13 different scales for various issue pairs in the manifesto dataset. This article focuses on its application to the same left and right categories that were defined for the RILE index. Thus, using the above notation, the formula for a logit RILE index would be as shown in equation 3.3.

LRILE = logN_R+ 0.5

N_L+ 0.5 (3.3)

CEUeTDCollection

Prosser left-right index (PLR)

While these two alternative approaches focused primarily on how to calculate the positions of parties on an assumed scale in terms of its content, there is also a range of approaches, which concentrate on how to determine the issue categories that should be used for estimating positions from the data, thus taking a more inductive approach to the measurement of political position. For this purpose Prosser (2014) takes the logit scale (equation 3.3) as the most valid way of constructing left-right positions from the manifesto data, but focuses on a method to select the appropriate categories form the data set. He uses exogenous correlations between issue categories and the relevant scale (correlation between an item and a scale that excludes the latter, starting with a naive initial scale and then adding and dropping issue categories) to select the appropriate categories (ibid., p. 95). In an elaboration of the Lowe et al. (2011) measure for salience, he uses the ratio of the logarithm of the total number of statements in an issue category (plus one) to logarithm of the total number of statements in a manifesto (plus one) to recalculate the raw category values in the manifesto data set (Prosser 2014, p. 96) for the purposes of evaluating these exogenous correlations. This ensures that they are comparable across manifestos of different length. Components that correlate with a scale are added to it and those that do not are removed iteratively until a stable equilibrium for the scale is reached. This results in a general left-right scale. Prosser uses the same procedure to construct separate economic and social scales. For consistency, we focus here only on the left-right dimension.

Franzmann and Kaiser left-right (FKLR)

Also focusing on how ideological issue categories should be selected out of the whole data set, Franzmann and Kaiser (2006) propose a measure for party positions on a left-right dimension, which changes across countries and time. The corresponding calculation of party positions involves the following steps (ibid., pp. 167-174). First, a linear regression is used for each of the 56 coding cate-gories, with the values of the categories as the dependent variable and party dummies as independent variables, to select the categories that most distinguish between parties. These are assumed to be the ideological categories. This is done separately for all party systems. Depending on which parties emphasise which issues, they can then either be classified as left or right. Categories that do not differentiate between parties are classified as valence issues. Thus, the whole range of categories in the manifesto data set is taken into account.

The final position on the left-right dimension is calculated similarly to the RILE index – the difference between the right position scores and the left position scores is divided by the sum of

CEUeTDCollection

the position scores plus the valence scores (Franzmann and Kaiser 2006, p. 173). This is shown in equation 3.4. R, L andV refer to right, left and valence scores across all the issue categories that have been thus identified. The scores themselves are calculated from the values of the issue categories in the data set by subtracting the lowest score for a category across the parties at an election. The ideological position of a party at an election is a moving average with the two adjacent elections taken into account. Using this method, party positions have also been calculated on separate economic and social dimensions (see Franzmann 2009).

FKLR = R−L

R+L+V (3.4)

Jahn left-right (J)

Although partially written as a critique of the Franzmann and Kaiser measure, Jahn (2010) proposes a very similar approach to estimating the positions of parties on a left-right dimension. Jahn uses Norberto Bobbio’s theoretical account of the left-right dimension (Bobbio 1996) to determinea priori the core issues that relate to each of the poles of this dimension. This is the part of the ideological dimension that he assumes. Multidimensional scaling (MDS) is then used to determine the location of these assumed core issues on the dimension. Thereafter, Jahn uses regression to select the time and country specific elements of left and right (those that correlate highly with the core dimension).

Another MDS is applied to this new set of issues to determine their location on the final form of the dimension. The positions of parties on the left-right dimension are constructed as the sum of the values of the issue categories in the manifesto data set each multiplied by their locations determined by MDS as shown in equation 3.5, where LRcore refers to the set of categories that are assumed,

LR_extra to those determined by the regressions and S to the corresponding locations determined by

MDS.

J =X

LR_core×S+X

LR_extra×S (3.5)

Elff economic left-right (EELR)

The final two measures considered here are methodologically much more intricate and thus it is not possible to concisely or formally describe the mechanism that leads from the manifesto data to the final estimate of ideological positions. The following only conveys their overall logic. Elff (2013), in his proposal for a left-right measure, starts with the fact that positions in an ideological space,

CEUeTDCollection

considered as latent variables, are different from the frequencies of words or sentences in a manifesto and proposes a measurement model that is able to take this difference into account and go from the latter to the former. His proposed model also includes a dynamic component that captures the possible change of positions over time. It assumes that parties, unless they are new parties, do not establish their positions tabula rasa, but use their previous position as a point of departure.

Elff estimates the positions of policies as well as actors in a policy field separately from the party positions that are reported in the manifesto data set. He employs this model on a rather restricted sets of coding categories (Elff 2013, pp. 224, 228) to estimate party positions in a uni-dimensional economic space and in a two dimensional liberal-authoritarian – permissiveness-traditionalism space.

For consistency, we focus here only on the left-right dimension.

K¨onig et al left-right (K)

K¨onig, Marbach, and Osnabr¨ugge (2013) also propose a dynamic latent variable model based on the manifesto data, which additionally incorporates information from expert surveys into the estimation process. One of their main concerns that had not been addressed by previous measures is cross-country comparability. They use a Bayesian framework to incorporate prior information (expert assessments) about the shape of the latent policy space and a logit transformation (Lowe et al.

2011) of the data, excluding some of the categories of the manifesto coding scheme. Furthermore, they assume that parties take “the same position in their first EP election as in the previous national election” and that “parties with the highest relative seat share gains in their country take the same position in the next election” (K¨onig, Marbach, and Osnabr¨ugge 2013, p. 9). They conclude that their method provides plausible estimates on the left-right ideological dimension with reference to convergent and construct/face validity. Their estimates seem to be similar to expert judgements, but very different from those of the RILE.