The popularity function: a spurious regression? The case of Austria

Estimating vote and popularity functions is one of the most frequent activities in empirical public choice analysis. In their influential survey of this work, Nannestad and Paldam (1994) referred to a large number of studies estimating links between economic variables and voters’ evaluations of political parties and governments in many countries. One of their main conclusions was that a clear link exists in particular between unemployment and voters’ evaluations of governments and ruling parties as expressed by election results or voting intentions embodied in popularity data from opinion polls and surveys. Weaker but frequently also significant influences were obtained for inflation rates and sometimes (real) income or income growth rates. Estimations of vote and popularity functions were most successful using time series data, although some progress was also made with micro-studies. A more recent example of successfully using vote and popularity functions to predict election results is Fair (2012). However, as many authors note, such functions often lack stability across time and across countries. These observations were confirmed by the survey by Lewis-Beck and Stegmaier (2013), who estimate that about 500 such investigations have been carried out altogether. There even appears to have been a renewed wave of interest in such activities since that survey.

The lack of stability of vote and popularity functions is somewhat disturbing and justifies additional econometric work. For instance, Kirchgässner (2009) has shown that in Germany the popularity function disappeared in recent years or at least did not result in significant and economically easily interpretable estimates for the influence of even the most frequently detected determinants on voters’ opinions about the ruling parties, namely unemployment and inflation. He attributed this partly to changes in the European institutional framework, where the rate of inflation is no longer determined on a national level but is the main agenda of the European Central Bank. On the other hand, the lack of significance of unemployment and labor market variables is less easily explained as these are still at least partly under the control of government policies. These and similar concerns call for a reconsideration of the vote and popularity function for other countries as well.

One of the most astonishing facts is the small number of econometric studies on vote and popularity functions using the econometric techniques developed in the context of the unit root and cointegration analysis initiated by the seminal work of Granger and coauthors, in particular Engle and Granger (1987). This is surprising in view of the large number of applications of this methodology to other, more traditional economic functional relations (e.g. the consumption function or the money demand function) yielding results which often modified earlier estimations in a fundamental way and gave way to new theories about the behavior of economic agents. It could be expected that similar modifications would result from the application of such techniques to the vote and popularity function. One notable exception to the lack of unit root and cointegration techniques is the replication and extension by Harrison and Marsh (1998) of estimations for Ireland by Borooah and Borooah (1990) which, while confirming some of the results of the earlier study, also showed important modifications of others. In particular, they found that the short-run impact of the economic variables is weaker than, and different from, that suggested by the earlier study; economic influences seem to be stronger on the general level of government approval than on short-term changes in popularity as manifested in opinion polls. Only a few other investigations of this type can be found in the economics literature (cf. Boya & Malizard, 2015; Chang et al., 2009; Davidson, 2005; Davidson et al., 2006).

More generally, it seems possible that the high coefficients of determination obtained in some vote and popularity functions combined with the lack of stability are the result of spurious regressions in the sense of Granger and Newbold (1977). In such a case, the presence of stochastic or deterministic trends preventing the stationarity of the time series involved may lead to seemingly strong correlations between economic variables which are, in fact, not related at all. Such may well be the case for some vote and popularity functions as well, possibly contributing to the lack of stability and the often weak performance of these functions in forecasting election results. However, it has to be borne in mind that variables like the popularity of parties or governments as well as some economic variables like the rate of unemployment cannot be strictly driven by a stochastic trend because they are bounded from above and below. Hence, the unit-root property cannot hold outside these boundaries, which may lead to inaccurate results with the cointegration methodology if the values of these variables hit them or are close to them.

In this paper, we perform a first examination of popularity functions for a particular country, namely Austria, using some of the techniques of unit root and cointegration analysis as well as other methods of econometric time series analysis, including Granger causality. Due to the lack of sufficiently long time series for election results, we confine ourselves to those popularity functions where we have consistent annual data, at least for the period 1976 to 2010. Unfortunately, in Austria, popularity data not made available to the public or even academics, except when they are published (often in a biased way) by the political parties who commission such surveys. So far, Austria has been the focus of studies by several authors (Neck, 1979a, 1979b, 1988, 1996; Hofreither, 1988; Neck & Karbuz, 1995, 1997; Schneider et al., 2017). Most of them found at least plausible and sometimes significant effects of unemployment on the voters’ evaluation of the governing parties. However, these results too may suffer from the spurious regression problem and therefore need to be reexamined in the light of the unit root and cointegration methodology. As it turns out, our findings cause severe doubts to arise about the appropriateness of those earlier results, leading us to conclude that it cannot be ruled out that at least some of them were examples of spurious regressions. To obtain more insights into this question, we conducted additional econometric investigations which show that the earlier results seem to be more robust than the unit root and cointegration methodology apparently suggests.

The following parts of the paper are organized as follows. Section 2 provides some theoretical considerations about popularity functions as well as a brief literature review. In Sect. 3, we give an introduction to the political system in Austria insofar as it is of relevance to the estimation of popularity functions. Section 4 reports the results for the Social Democrats (SP), who were in charge of economic policies from the beginning of the period under consideration until 1999, while Sect. 5 presents the results for the People’s Party (VP, Christian Democrats or Conservatives), who were part of the federal government from 1987 to the end of the period. Section 6 investigates the determinants of popularity of the incumbent, i.e., the party (in one-party governments) or ruling coalitions over the entire period considered, establishing a more stable impact of unemployment on voters’ evaluations of the federal government. Section 7 summarizes the main results and gives some reasons for the lack of links between some other macroeconomic variables and the popularity of the parties in Austrian federal government.

In this section, we provide some theoretical considerations and give a brief literature review focusing on more recent papers.² Vote and popularity models analyze the relationship between economic and political variables and support for a government. According to Nannestad and Paldam (1994) and Paldam (2004), about 300 papers on vote and popularity functions were written between 1974 and 2004.³ Most of this research is empirical in nature. If the theoretical reasoning is summarized, starting with Downs (1957), Davis et al. (1970), Mueller (1970), Stigler (1973), Frey and Schneider (1978a, 1978b, 1979), Fair (1978), Hibbs (1977), Hibbs and Vasilatos (1981) and Kirchgässner (1986), one general finding of these authors⁴ is that, according to Downs’s theory, the behavior of selfish politicians and voters can be reduced to the operational idea known as the responsibility hypothesis: Voters hold the government responsible for the past development of the economy. This hypothesis predicts that if the economy is going well, voters will approve of this and the popularity or the election outcome of the governing party (parties) will increase; if the economy is in a bad shape, in contrast, the popularity and the election results of these parties will deteriorate.

Most authors choose a linear functional relation to model the vote/popularity function, and in most cases the economic variables unemployment rate, inflation rate and growth rate of personal disposable income have been used. This is also the case in this paper; hence, we model the popularity function in the following way:

The theoretically expected signs are UR < 0, IR < 0, YR > 0.⁵

Following the contribution by Lewis-Beck and Paldam (2000), Table 1 stylizes and summarizes the facts about the empirical results of the vote and popularity functions in the period 1970 to 1990. From Table 1 we summarize first that most of the empirical findings on vote and popularity functions derive from the estimates built on the responsibility hypothesis. This hypothesis offers a simple and reduced link between the economy and the voter. A second important finding is that the vote and popularity functions do not produce very stable results. Third, it is difficult to compare the results of the various authors and of the different countries because almost every author has their own specifications.

In a study by Kirchgässner (2009) with the provocative title “The lost popularity function: Are unemployment and inflation no longer relevant for the behavior of German voters?”, the author questioned the relevance of the popularity function in Germany for economic and institutional arguments. As we have seen in the discussion above, unemployment and inflation (and in most cases the growth of real disposable income, too) were the variables which quite often showed up to be statistically highly significant in vote and popularity functions in most countries. Kirchgässner correctly argues that up to that point there had been a general empirical finding that rising unemployment and inflation have a negative impact on a government’s popularity. This was true for Germany as well: for the governments of Adenauer, Erhard, Brandt, Schmidt and Kohl, rising unemployment and inflation had the negative influence on the popularity of these governments predicted by the theory and were statistically highly significant. However, this result does not hold for the Schröder government: when Kirchgässner estimated the popularity for the Schröder government, neither unemployment nor inflation had a statistically significant influence on its popularity. Kirchgässner argues that the lack of an impact for unemployment might be for statistical reasons, namely the short observation period and the low variance of the explanatory variables. In terms of inflation, the voters might have realized that they cannot hold the government responsible for this phenomenon any longer as the European Central Bank (ECB) has been in charge of monetary policy in Europe since 1999. Although Kirchgässner found no significant influence of macroeconomic variables on the popularity of the Schröder government, he thought that it was much too early to draw the general conclusion that voters do not hold a government responsible for economic development because more and better data are needed to undertake further investigations to confirm these results.

The only studies which, to our knowledge, deal with the Austrian situation are the ones by Neck (1979a, 1979b, 1988, 1996), Hofreither (1988) and Neck and Karbuz (1995, 1997). In these papers, econometric evidence is presented for the influence of macroeconomic variables on the popularity of political parties in Austria. They use popularity data provided by the Institute for Market and Social Analyses (IMAS), Linz, and quarterly data over the period 1975 to 1993. The main results are as follows: the rate of unemployment, the growth rate of disposable income and the rate of inflation were identified as economic determinants of voters’ evaluation of political parties. The papers showed that there is evidence for a structural break in the popularity functions relating to the change from a one-party and/or small coalition government to a “grand” coalition (a coalition of the big parties SP and VP). For a one-party government or a small coalition, the predictions of the responsibility hypothesis for the popularity functions were confirmed, meaning that the unemployment rate and inflation rate have the theoretically predicted negative sign and are statistically significant in most cases. However, these results are not very stable; in some cases, they had the predicted statistically significant coefficient and theoretically expected sign, in other cases they did not.

As we are considering popularity functions for Austria for the period 1976 to 2010, it is appropriate to briefly introduce the political system and some economic policy issues in the country in that period. From 1970 to 1983, Austria was governed by a one-party government led by the Social Democratic Party (SP) under federal chancellor (prime minister) Bruno Kreisky. During this period, the effects of the first and second oil price shocks resulted in a deceleration of growth and (in 1975) even in a (mild) recession. This prompted the government to launch a policy later dubbed “Austro-Keynesianism”, which consisted in expansionary fiscal policy measures to combat unemployment (mainly by increasing government expenditures, financed by public debt) and monetary policy following a strict peg to the deutschmark (the so-called hard-currency policy) aiming at “importing” low inflation rates from Germany, Austria’s main trading partner. The institutional background was the system of “social partnership”, a voluntary cooperation of the main interest organizations for employers and employees, which guaranteed moderate wage growth and the virtual absence of strikes but also included extensive interventions by these interest groups in the Austrian economy.

When the SP lost their absolute majority in Parliament in 1983, they formed a “small coalition” with the third largest party, the Freedom Party (FP), which at that time stood for a right-wing liberal policy stance. This coalition came to an end in 1986 when the charismatic Jörg Haider became leader of the FP, turning the party toward right-wing populist positions. After an election that year, in which the SP kept its relative majority in Parliament in spite of considerable losses, the SP formed a so-called grand coalition with the second largest party, the Austrian People’s Party (VP), a Christian Democratic or conservative party. In 1986, following successful campaigns resulting in a ban on nuclear and hydro power plants, a fourth party was elected to Parliament for the first time in nearly 30 years, namely the Greens (environmentalists), the Communists having lost their last seats in 1959, never to regain them since.

The SP-FP’s “small coalition” did not manage to keep the public budget deficit under control, hence a primary goal of the SP-VP’s “grand coalition” was to consolidate the federal budget. Another challenge came from political and economic developments in the late 1980s, with the eventual demise of the Eastern bloc and the centrally planned economies of Central and Eastern Europe and the Soviet Union. Sooner or later, the Austrian government became convinced that it would be desirable to join the European Union to find an appropriate status in the new European political and economic architecture. After a positive referendum, Austria became a member of the EU, together with Sweden and Finland, in 1995. Cooperation between the two biggest parties (and, what was very important, the organizations in the “social partnership”) to manage Austrian entry into the EU is generally regarded as the main achievement of and justification for the “grand coalition” during the 1990s.

The coalition government was less successful at keeping the federal budget under control, however, and after joining the EU tensions increased between the two ruling partners. This was amplified by the rise of the FP, the main opposition party, which increased its voting share from less than 5 percent in 1983 to nearly 27 percent in 1999. Growing frustration about political deadlock due to the divergent positions of the two coalition parties, especially within the VP, which was tired of being number two in the government, prompted the end of the “grand coalition” (which was not really “grand” anymore, as the FP had replaced the VP as second largest party in the 1999 election) and the formation of another “small coalition”, this time between the FP and the VP, led by VP politician Wolfgang Schüssel.

The new government aimed at major changes (in their language: a “turnaround”) in several policy fields, especially economic policy. In spite of strong criticism, both international and national, of the participation of the right-wing populist and nationalist FP, the “small coalition” managed to implement several reforms aiming at securing the sustainability of public finances. However, a (near-) balanced budget was achieved only once and the coalition came under stress from internal differences and a split within the FP as well as from several scandals shaking its credibility. Thus, after a period of dominance by the VP, which became the strongest party in the 2002 elections, in 2006 the SP regained its former position as number one. This led again to the formation of a “grand coalition” government by the SP and VP, which had to cope with the “Great Recession” 2007–2010 (which it did fairly well) and which suffered from a steady decrease in its popularity due to the lack of a long-term project justifying its existence, until a new “small coalition” of the VP and FP was formed in 2017. This government collapsed after a political scandal involving then-FP leader Heinz-Christian Strache in 2019, giving way (after an interim expert cabinet) to a coalition between the VP and Greens, which have been in office since 2020.

The SP led the Austrian government during the period 1976 to 1999, with the federal chancellor and minister of finance coming from this party; hence, we will first conduct an analysis for this party. Although for the larger part of this period it had to share power with another party, it is reasonable to assume that voters attributed successes and failures in economic policy largely to the SP, if at all. Figure 1 shows that the rate of unemployment (according to a national measurement, the only series available over the entire time horizon) began to rise a few years after the first oil price shock from around 2 percent to 7 percent at the end of the millennium. The rate of inflation (measured by the growth rate of the consumer price index), on the other hand, decreased toward the end, after some oscillations in the first few years, eventually resulting in virtual price stability. Figure 1 also shows the development of the popularity of the SP during these years, which, after 1983, decreased more or less steadily. It may be conjectured (and this was also the conclusion of the studies mentioned above) that this fall in the SP’s popularity was related to the simultaneous increase in unemployment, a phenomenon especially relevant for the potential electorate of this party, which at least in the past was a typical workers’ and employees’ party.

Equation (1) in Table 2 gives the result of a traditional popularity function with the arguments unemployment rate (UR) and inflation rate (IR). It indicates a seemingly clear negative effect of the unemployment rate on the popularity of the SP as predicted by the theory. The inflation rate’s effect has the wrong sign and is insignificant. The latter also accords with earlier studies and is easily explainable: SP voters are by far more affected by the threat of becoming unemployed and by social and political troubles from high unemployment than by inflation, which poses a higher threat to wealthier people, who overwhelmingly show no potential sympathy to the SP anyway. Moreover, even before the introduction of the euro and the shift of monetary policy to the EU, it was generally recognized that a small open economy like Austria depended to a large extent on foreign (global, European and German) developments with respect to inflation; hence, the government and ruling parties were rarely made responsible for increases in price levels.

Equation (2) seems to suggest that the negative effect of the unemployment rate is even stronger when the inflation variable is omitted. Entering an income growth rate variable did not result in significant effects in either of our estimations and so they are not reported here. Interestingly, using the inflation rate as the only regressor results in a significant (but positive) effect, which points toward a spurious regression (Eq. (3)). Thus, an initial conjecture might be that the increase in the unemployment rate contributed to the decrease in the popularity of the Social Democrats in Austria during the 1980s and 1990s. However, we have to check whether this apparent effect is not due to a spurious regression as well.

For this purpose, we examined the stationarity properties of the time series involved. The usual procedure is to execute unit root tests on them. Table 3 presents the results of the Augmented Dickey-Fuller unit root tests (ADF test) for the unemployment rate and its first difference. The null hypothesis is always the presence of a unit root, hence non-stationarity. The “Probability” column gives the MacKinnon one-sided p-values. The results clearly cannot reject a unit root for the unemployment rate at any reasonable level of significance. For the first difference, the unit root hypothesis can be rejected at the 10, 5 and 1 percent levels in the three versions of the ADF test reported, which we accept as evidence for the stationarity of the first difference. Hence the unemployment rate can be regarded as being integrated of order one, an I(1) variable. The Phillips-Perron (Phillips & Perron, 1988) (PP) tests confirm this result. For the sake of completeness, we mention that we also conducted ADF tests for the (insignificant) inflation rate and found that it is an I(2) variable during this period, which may or may not have a deterministic trend.

Table 3 also contains the results of the ADF test for the popularity of the SP (variable SP). Here the situation is more difficult to interpret. The ADF tests for SP accept the unit root hypothesis but only at the 10 percent level when including a deterministic trend. This (negative) trend is highly significant in the respective version of the test (t value:  − 3.21). Moreover, the test equations for the first difference of SP (not reported here) show that specifications without the trend have coefficients of determination near zero, which shows that the trend must not be omitted. Nevertheless, we also ran ADF tests for the first difference of SP, which unambiguously reject the unit root hypothesis. The results of the PP test are also ambiguous, with one strong rejection, one weak one and one acceptance of the unit root null hypothesis. Hence, SP is I(1) or its movement is governed by a deterministic trend.

Accepting for the moment the assumption that SP is an I(1) variable, we can attempt to check whether the two series SP and UR are cointegrated. Remember that this is a necessary condition for running a regression like Eq. (2) as both variables are in any case non-stationary and, in the absence of a cointegration relation, the regression will be spurious. Table 4 presents the results of the Engle-Granger and the Phillips-Ouliaris cointegration tests. The null hypothesis is always that the series are not cointegrated. In no instance can the null hypothesis be rejected at the 5 percent level. Next, Table 5 gives the results of various versions of the Johansen cointegration test. They are mixed, rejecting cointegration in only one case (with no intercept and no trend). Therefore, the hypothesis of cointegration cannot be regarded as disproven but needs further investigation, especially in view of the possibility of the presence of a deterministic trend. One argument against the cointegration hypothesis comes from the fact that we were not successful in estimating an error correction equation for SP (under the Engle-Granger approach, the residuals of the cointegration equation seem to have a unit root), which would be a necessary and sufficient condition for cointegration between SP and UR.

Let us now consider the evidence for the deterministic trend. First, we included the trend in the estimated equation for SP. Equations (4) and (5) in Table 2 show that it is highly significant when included both as the only regressor and together with the unemployment rate. Moreover, when the unemployment rate is included in addition to the trend, it becomes completely insignificant (Eq. (5)) and the adjusted R² even deteriorates. This implies that the deterministic trend must be included when explaining SP and gives a strong indication that the unemployment rate does not contribute to the explanation of the decrease in popularity of the Social Democrats while in power in the 1980s and 1990s.

In order to make the SP series stationary, it is necessary to remove the deterministic trend. This can be done either by taking first differences or by regressing the series on the trend and taking the residuals of that equation instead of the series itself. For the first possibility, we note that the first difference of SP does not contain a stochastic trend and can therefore be regressed on the first difference of the unemployment rate in order to check whether at least a short-run link exists between the change in the unemployment rate and the Social Democrats’ popularity. This regression results in an insignificant positive (“wrong”) coefficient of UR and an adjusted R² of 0.03; hence such a short-run effect is not present either.

The same result is obtained if we remove the deterministic trend by taking the trend-free component of SP instead of the variable itself. First, we took the residual of Eq. (4), to be called SPS, as the systematic part and examined its properties. ADF tests confirm that SPS does not contain a unit root. When regressing it on the change in the unemployment rate, this variable again has a positive coefficient which is significant at the 10 percent level only, confirming the presumption of no cointegration. Just to complete this task, we note that the second difference of the inflation rate used as the regressor turned out to have a positive and completely insignificant coefficient as well.

To summarize: The development of SP popularity is mainly driven by a downward deterministic trend, irrespective of the development of the rate of unemployment. Although it cannot be excluded completely, only very weak evidence exists for a systematic relation between unemployment and the popularity of the SP. The decline in the SP’s popularity is, therefore, overwhelmingly determined by factors other than the usual independent economic variables in the popularity function. They may include a general decline in the attractiveness of Social Democratic politics with European voters, the lack of a charismatic leader (like Kreisky in the 1970s), the rise of free-market and environmental ideologies and the rise of populism (in Austria, the FP), among others.

Next, we examined popularity functions for the VP during its participation in federal government from 1987 to 2010. Figure 2 shows the development of its popularity (VP) and of the unemployment and inflation rates during these years. It first shows the rise in the unemployment rate, which then, with short-lived downward movements, remained around 7 percent. The inflation rate first increased until 1992 and then fell to levels compatible with price stability except for a second peak in 2008 at the end of the boom before the “Great Recession”. The popularity of the VP fell until the end of the 1990s, then rose with the VP being number one during its “small coalition” with the FP but already from 2005 on it dropped again rapidly until the end of the estimation period.

For the VP, the a priori presumption of systematic economic effects are weaker than for the SP because the VP was only a junior partner from 1987 to 1999 and at most an equal partner from 2007 onwards (when it provided the finance minister but not the federal chancellor). It might be conjectured that the rate of inflation is more important than the rate of unemployment for the popularity of this Christian Democratic or conservative party as most of its potential voters (the self-employed, farmers, civil servants, managers and similar groups) are not strongly affected by higher unemployment.

We first start with “naïve” OLS regressions of VP against UR and IR; see Eq. (1) in Table 6. Both economic variables have the expected negative sign but are insignificant and the R² is rather low. Equation (2) shows that the inflation rate taken alone remains insignificant (and the coefficient even switches its sign) while the unemployment rate (Eq. (3) taken alone becomes significant at the 10 percent level at least. If anything, unemployment seems to have a stronger (but still doubtful) effect on the popularity of the VP as well during this period.

The next step consists in testing for unit roots in the three variables for this period. Table 7 gives the results of the ADF tests. According to them, the unemployment rate has a unit root while its first difference does not; hence it is I(1). For the inflation rate, we again find it to be I(2): neither for IR nor for its first difference can the null hypothesis of a unit root be rejected. For the popularity variable VP, the ADF tests give mixed results (two rejections of the unit root, one acceptance) but Phillips-Perron tests (PP) accept the null at least at the 5 percent level. For the first difference of VP, the ADF tests are also inconclusive but here the PP tests strongly reject the null hypothesis of a unit root. As the PP test was shown to be more powerful than the ADF test in the finite sample with annual data (Choi & Chung, 1995), we conclude that VP is an I(1) and IR is an I(2) variable. For none of the variables does a deterministic trend become significant.

This implies that cointegration may exist between VP, UR and the first difference of IR, all of them being I(1) variables. In view of the insignificance of the explanatory variables in the candidate for a cointegration equation and its low R², it does not seem very promising to look for a cointegration relation but we did so nevertheless. Table 8 presents the results of the Engle-Granger and the Phillips-Ouliaris cointegration tests. Neither of them rejects the hypothesis that the series are not cointegrated. The Johansen cointegration tests give mixed results for the equation with both variables but indicate at most one cointegration relation (Table 9). If we look at the Johansen test for each of the regressors separately, we see that cointegration between VP and UR is clearly rejected by one Johansen test; cointegration between VP and the first difference of IR is rejected in all but one of the tests. In addition, no error-correction mechanism equation can be established. Altogether, we can be fairly sure that no cointegration exists between this set of variables.

Although no long-run relation exists between the variables considered, it is still possible that there is some short-run influence of economic variables on the VP’s popularity. This is tested by regressing the first difference of VP on the first difference of UR and/or the second difference of IR. As expected from the previous results, none of these regressions gives any significant coefficient and the adjusted R² is even negative in all three cases. Hence, we can be confident in saying that the popularity of the VP is not affected by either of the usual economic independent variables of popularity functions. Additional support for the lack of effects of the economic variables on the popularity of the VP comes from pairwise Granger causality tests which do not reject the null hypothesis of no Granger causality between any of these economic variables and the popularity of the People’s Party either.

Next, we investigated the development of the ruling party or parties. This has the advantage of focusing on the parties in government, the only ones which may actually have a direct influence on the economy, at least at federal level; in addition, it allows us to exploit the longer series over the entire period where data are available, hence giving higher degrees of freedom for the estimations. The disadvantage lies in the fact that the questions in opinion polls refer to individual parties: The members of a coalition may be evaluated differently, with successes or failures possibly being attributed to one ruling party at the expense of the other. Moreover, the composition of governments changed several times during the period under consideration. For this reason, we introduced dummy variables for the one-party government of the SP (SPONLY, being equal to 1 in 1976–1982 and 0 else), the “small” SP-FP coalition (SPFP = 1 in 1983–1986), the “grand” coalition (GK = 1 in 1987–1999 and 2007–2010) and the “small” VP-FP coalition (VPFP = 1 in 2000–2006). The popularity of the ruling party or coalition (the incumbent) is formed by taking SP as the popularity of the one-party government of the SP and by taking the sums of the coalition parties for the other periods. This variable is dubbed INC and varies, of course, between the different regimes for reasons other than economic ones, including different numbers of party stalwarts in the various parties. The development of the incumbent’s popularity and economic variables is shown in Fig. 3.

Again, we first ran OLS regressions for popularity functions over the entire period, trying UR, IR and YR as regressors; see Eq. (1) in Table 10. None of the regressors becomes significant, including the constant; hence this is a misspecification. When omitting one of them (e.g. Equation (2)), the situation does not change. Obviously, as expected, this is due to the breaks in the series INC (the popularity of the incumbent) in years of changes in the form of government, with coalitions (especially the “grand” coalition) enjoying more popularity than a one-party government or the unpopular SP-FP coalition in the 1980s, with the latter also suffering from the (completely non-economic) Waldheim affair. Next, we split the constant into the four different regimes by taking the dummy as the only regressor. Equation (3) shows that this reduces the standard error of the regression to a little more than half of that with the economic variables, explaining more than two thirds of the variance of the dependent variable. The coefficients show average support for the respective governing parties during their period in office. If we add the economic variables to this specification (Eq. (4)), the inflation rate and the growth rate of real income remain insignificant but the unemployment rate becomes significant with the expected (negative) sign. Omitting the two insignificant regressors improves the fit (Eq. (5)); hence we accept this specification for the moment.

We cannot exclude the possibility of some endogenous effects on the unemployment rate, which may be an imperfect indicator of the state of the economy as viewed by the voters. Therefore, we performed a two-stage least squares estimation of the determinants of the popularity of the incumbent (Eq. (6) in Table 10) and the rate of unemployment, which is assumed to depend negatively on the inflation rate (a Phillips curve relation) and the growth rate of real income (an Okun-law type relation). Equations (1) and (2) in Table 11 show that this specification for the unemployment rate performs rather well, with the expected negative sign of the two economic regressors both with and without the regime dummy. Both instruments exert independent influences on the unemployment rate. Figure 4 shows that the development of UR and of its estimated value (URFIT) are rather close, justifying this choice of instruments.

Comparing Eqs. (5) and (6) in Table 10 shows that the effect of the unemployment rate on the popularity of the incumbent becomes stronger when its dependence on the inflation and growth rates is taken into consideration, even though these variables do not directly influence the dependent variable INC. The unemployment rate can therefore be regarded as a reliable indicator for voters’ evaluation of the macroeconomic performance of the governing parties. For the sake of completeness, we also provide the results of a GMM estimation of the popularity function in Eq. (7) in Table 10 to show the robustness of the influence of the unemployment rate in the chosen specification.

Equation (8) in Table 10 shows the results of an OLS estimation with lagged values of the unemployment rate added as regressors. These are not significant but point toward some possible dynamic structure of the effects of unemployment on the incumbent’s popularity. This may be due to autocorrelation in the unemployment rate series, as Eq. (3) in Table 11 indicates. The first-period lag clearly raises the suspicion of a unit root in this series, so this question has to be examined in detail, both for the unemployment rate and the popularity variable. The results of this examination are displayed in Table 12.

The results can be briefly summarized as follows: Both the Augmented Dickey-Fuller and the Phillips-Perron tests indicate that unit roots in the two series cannot be rejected at reasonable significance levels, in contrast to the first differences, which do not show a stochastic trend. With high probability, both variables can therefore be regarded as I(1). This raises the question as to the possibility of spurious correlation. To obtain more information about this question, we examined whether the two variables are cointegrated. Tables 13 and 14 present the results of cointegration tests, the Engle-Granger and the Phillips-Ouliaris test in Table 13 and the Johansen test in Table 14.

More details about the tests are available from the authors. Fortunately, these tests provide unambiguous information on the question as to whether the two series are cointegrated or not. All the tests indicate that there is no cointegration between the unemployment rate and the popularity of the incumbent. This implies that there is no long-run relation between these two variables and the influence of the rate of unemployment on government popularity is a short-run phenomenon. This is not astonishing given the breaks in the popularity series due to the changes in the composition of the government during the estimation period. Cointegration breakpoint tests (not reported here in detail) confirm breakpoints in 1986 and 2006 in some specifications.

To obtain more information about the effects of changes in the unemployment rate on the incumbent’s popularity, we also examined the question as to whether Granger causality can be shown between these two variables. To do so, we estimated a vector-autoregressive (VAR) model, which as such leaves the direction of causality open in the modeling phase and tests statistically for the presence of an increase in the explanation (in a dynamic and statistical sense) of one variable by the other. Table 15 gives the results of such an estimation.

Here the lag length 2 was determined automatically by the Schwarz criterion; the Akaike criterion and the lag exclusion Wald tests give the same result. The results for the Granger causality tests show that the null hypothesis that UR does not Granger cause INC can be rejected at a high significance level, while the reverse (INC Granger causes UR) can be accepted only at the 10 percent significance level. The Johansen cointegration test for the VAR confirms the results for the single equations: no cointegration relation exists between INC and UR.

Finally, we applied the procedure developed by Toda and Yamamoto (1995) to test for Granger causality in cases where the degree of integration of the time series involved is uncertain. For this purpose, a VAR model of the form y_t = α + β₁y_t−1 + … + β_p+dy_t-(p+d) + ε_1t has to be estimated where p is the lag length, d is the maximum integration level of the variables, α is the intercept term, β_i are coefficient matrices and ε_1t is the error term. The Toda-Yamamoto causality test adopts the null hypothesis of no Granger causality (β₁ = …. = βp = 0) with a test statistic based on the Wald statistic and an asymptotic χ² distribution with p degrees of freedom. In our case, p = d = 1, and we calculated the Wald statistic for the VAR of Table 15 for UR not to Toda-Yamamoto cause INC to be 4.85, which for a χ² distribution with one degree of freedom is significant at the 5 percent level. This confirms that there is some influence of the unemployment rate on the popularity of the governing parties, at least in the short run in the sense of Granger causality.

We can summarize our results as follows: Neither during the first period with the dominance of the Social Democrats nor during the second (overlapping) period with the participation of the People’s Party in government, can influences be established of either unemployment or inflation affecting the popularity of either party. This puts a question mark over previous results for this country (including our own), which claimed to have found such influences. There is a real possibility that those earlier studies fell into the trap of spurious regressions.

However, it would be premature to conclude from this result that the popularity function does not exist at all. First, our estimates for the ruling parties and coalitions show that there is indeed some negative influence of the unemployment rate on the incumbent’s popularity, which, when taking account of differences between regimes, can be established over the entire period. Particularly the Toda-Yamamoto approach showed that, irrespective of the integration properties of the time series involved, the unemployment rate is Granger-causal to the evaluation of the ruling party/parties during the years under consideration. Several tests showed that this effect seems robust against variations in the estimation method. So, even for Austria, with its dominance of “grand” coalitions, news about the death of the popularity function is exaggerated. The inflation rate not being significant is plausible given the status of the Austrian economy as a small open one with an essentially fixed exchange rate which, as a member of the Euro Area since 1999, no longer has a monetary policy of its own at all. Voters may rationally hold international developments responsible for financial developments instead of domestic politicians.

Moreover, Austria is far from being a two-party competitive political system, with a “grand” coalition in power for most of the period under consideration and the “social partnership” having decisive influence even when only one of the major parties is in government; hence not too many opportunities are available for partisan policies in economic policy making. Finally, the time periods investigated are rather short and the data on party popularity may not be very reliable, although the annual data for the election years are close to the election results. Nevertheless, the influence of unemployment, which is generally regarded as the main threat to the welfare of a large number of citizens in Austria, does have some influence on an evaluation of the parties in power in accordance with the responsibility hypotheses behind the popularity function. On the methodological side, we have shown that the use of the unit root and cointegration methodology can lead to new and unexpected results but they have to interpreted with care and should be rechecked using other econometric methods, for example the Toda-Yamamoto procedure of detecting Granger causality, also when the degree of integration of the time series is not beyond doubt. It remains to be seen whether applying these methods to data from other countries provides challenges to their popularity functions.

UR = unemployment rate in percent; Statistics Austria, STATAS, Vienna, various years. IR = inflation rate, consumer price index; Statistics Austria, Vienna, various years. Popularity (SP, VP) = popularity of the SP (Socialist/Social Democratic Party), VP (People’s Party/Christian Democrats or Conservatives) and FP (Freedom Party/right-wing liberal/populist) in percent, yes-shares; Fessel GfK, Vienna, 2014 and various years. All econometric estimations were made using the EViews software.