What Follows What? Relations between Economic Indicators,Economic Expectations of the Public,and News on the General Economy and Unemployment in Germany,2002-2011

Impact of PEs and EIs on News Coverage

Economic news relies on real-world cues such as EIs or public opinion of the economy. Which of the two is more potent for news coverage has not fully been clarified. ⁸ Behr and Iyengar reveal that news coverage is more likely influenced by EIs than by public opinion. ⁹ Later, Wu et al. support this result. ¹⁰ However, in periods with an economic downturn, public opinion can forecast economic news better than the real-world economy. In Blood and Phillips’s study, ¹¹ the effects of changes in public opinion on news on the general economy are reported across time. Here, during an upturn and a stable period, public opinion had an effect on news whereas leading EIs had no effect. Further, economic news is often characterized as negative. ¹² Changes in news coverage are found to be asymmetric, favoring the coverage of negative economic developments. ¹³ Nadeau et al. ¹⁴ show that EIs (change in unemployment, change in inflation, the oil shock in 1979/1980) and elite judgments have an impact on the economic outlook in the news—and that the news interprets economic conditions. The dramatic event of the oil shock had an especially strong negative impact on news outlooks in their study. Also different affectedness of employment may cause different levels of news salience. ¹⁵ In the study by Starr, ¹⁶ consumer sentiment and the unemployment rate have a stronger forecasting ability for news than do other real-world indicators such as the Federal Funds rate, consumption, inflation, and stock prices. However, Soroka ¹⁷ proves that very little unemployment news variation can be explained by the unemployment rate. He argues that a certain amount of noise, the interests of the journalists and third parties, as well as the salience of other issues may produce this result. Similarly, Nadeau et al. ¹⁸ discuss the idea that the journalistic interpretation of real-world events may lead to weak relations between economic news and EIs. The news pattern in recession periods is an excessive increase in negative news caused by real-world developments. Journalists may act in a “regular mode,” with greater interpretative autonomy, during non-recession periods that declines during recession periods. Hence, I ask the following.

Impact of News Coverage and EIs on PEs

Although many researchers have revealed that news coverage influences people’s agendas and views about the economy, ¹⁹ and traditional news media affect the public agenda, ²⁰ the public can also obtain economic information from their daily lives, experiencing real-world cues in their environment that supersede the influence of media coverage. ²¹ The obtrusive contingency hypothesis expects agenda-setting effects to decrease with personal experience of an issue. Media may set the agenda for abstract, unobtrusive topics. ²² Obtrusive topics are those that “people can have personal experience” ²³ with and that relate to “events in their daily life.” ²⁴ Previous research a priori defines unemployment as an obtrusive issue, whereas the national gross domestic product (GDP) can be defined as either obtrusive or unobtrusive. ²⁵ For national economic topics, indicators in people’s lives such as the bustle on shopping streets or the construction activity in a city can influence economic sentiment. Still, only when the news reports about economic issues can these be grasped as a social problem and expectations for the economic development in a country be developed. For unemployment, Mutz ²⁶ finds that personal experience affects the evaluation of the individual employment situation, whereas, at a national level, economic news affects the evaluation of unemployment as a social problem. Hagen ²⁷ argues that unemployment contains personal, egocentric (fear of losing one’s job), and sociotropic (national problem) aspects, and finds that the news affects the perception of unemployment as an important national problem but is not likely to engender the fear that one will personally become unemployed. Yet both the general state of the economy and the increase in unemployment can affect personal life. Therefore, I understand the general state of the economy and national unemployment development as potentially obtrusive topics and expect the following according to the obtrusive contingency hypothesis:

National unemployment development is a lagging indicator of the economy, that is, follows changes in the general economic situation. ²⁸ Hence, when the public develops national unemployment expectations as opposed to current unemployment evaluation, the general state of the economy may be considered as well. Therefore, I test the following:

According to the concept of collective media effects, media users and non-users develop similar economic expectations through follow-up conversations. ²⁹ Yet a high degree of news consonance across outlets is crucial to the measurement of collective news effects. ³⁰ Hence, I test the interaction effect:

Shehata and Strömbäck find no short-term agenda-setting effects for the perceived importance of unemployment as an “important political issue today” and suppose that chronic accessibility due to high news salience prevents an effect. ³¹ However, Soroka ³² expects that media influence on the public decreases as unemployment increases since more people then have their own experiences with unemployment, that is, obtrusiveness and the perception of an egocentric problem increase. ³³ This is why personal experience creates similar economic expectations among media users and non-users. ³⁴ Hence, I test the interaction effect:

During the financial crisis, economic news served as the dominant source for the public when it was almost impossible for people to accurately evaluate economic issues. ³⁵ Wu and Coleman ³⁶ and Hester and Gibson, ³⁷ using the examples of presidential candidates and the economy, respectively, show that negative information has a more powerful public agenda-setting effect than does positive information. Wu et al. ³⁸ reveal that, in a recession, news has a stronger effect than the real economic situation on PEs. More attention may be given to economic news during a recession to assess whether someone’s own employment situation may be affected. As the need for orientation increases, the public pays more attention to economic news, ³⁹ and issue salience increases as a result. ⁴⁰ Hence, I expect the following:

Impact of News Coverage and PEs on EIs

Grossarth-Maticek and Mayr ⁴¹ find that counts of the word “recession” in German newspapers can somewhat forecast the German GDP. This relation should be moderated by economic decision making, which is influenced by public opinion. Wu et al. ⁴² show that public sentiment from two to four months ago predicts economic performance. Therefore, both belief shocks ⁴³ and self-fulfilling expectations ⁴⁴ can move the economy. Expectation-driven business cycles explain changes in the economy through changes in economic expectations, which are caused by news shocks: “Empirically, news shocks about future productivity growth [. . .] induce an expansion in aggregate consumption, investment, employment, output and stock prices.” ⁴⁵ Hence, false announcements of the macroeconomic development can affect economic growth. ⁴⁶

Iyengar and Kinder ⁴⁷ argue that, when making decisions, recipients rely on those topics that are most salient to them. After establishing attitudes and expectations, these should affect intention and behavior according to the theories of reasoned action and planned behavior. ⁴⁸ For example, TV evening news sentiment can predict private consumption. ⁴⁹ This comprises the micro-explanation of why consumer sentiment on the economy is considered an early indicator of the economy. ⁵⁰ Therefore, I propose the following hypotheses:

The German Case

This study analyzes economic news coverage by three television news shows and a newspaper with the largest audience in Germany, thereby representing Germany’s pluralistic media structure. According to Hallin and Mancini, ⁵¹ the German media system belongs to the Northern European or democratic corporatist model, with high newspaper circulation compared with that of other countries. Nevertheless, the reach of daily newspapers has been decreasing since the 1990s, dropping to 44% in 2010 from 54% in 2000, whereas TV reached 86% (85%) of the German population in 2010 (2000). ⁵² In Germany, the quantity of economic news coverage in newspapers and TV has increased during the last two decades, ⁵³ as it has in Europe generally. ⁵⁴ Therefore, I consider German TV and newspaper media outlets a good European example for researching economic news.

The periods of the first quarter 2003 to the third quarter 2004 and the second quarter 2008 to the first quarter 2009 comprise recessions in Germany. ⁵⁵

News Outlet Selection

The news outlet with the largest audience in terms of reach (9.62 million and 9.14 million viewers in 2002 and 2011, respectively) is the major evening news show ARD Tagesschau, a public service broadcast, followed by the major evening news show ZDF heute (5.17 million and 3.97 million viewers in 2002 and 2010, respectively), the second public service broadcast. ⁵⁶ During the observation period, the reach of a major commercial broadcaster’s evening news, RTL aktuell (3.89 million and 3.91 million viewers in 2002 and 2010, respectively ⁵⁷ ), exceeded that of heute. The fourth most important news outlet in terms of circulation is the commercial daily tabloid, BILD, whose circulation is in decline (4.05 million and 2.94 million in 2002 and 2010, respectively ⁵⁸ ).

The TV news shows provide general evening news. In the public service news shows, 70% of the content is dedicated to political and economic news. ⁵⁹ About 40% of the commercial news show, RTL aktuell, consists of political and economic news, with the rest consisting of sports, human interest, or celebrity stories. BILD is not specialized in providing economic news and does not have a business news section but is a “prototypical representative of tabloid journalism.” ⁶⁰ Yet its news is found to influence economic sentiment. ⁶¹ Shehata ⁶² finds that TV news and tabloid newspapers display greater agenda-setting effects than other print media news.

Therefore, the news outlet selection covers the news repertoire of many Germans and allows the disclosure of collective agenda-setting effects on the aggregate level.

Data

PEs

PEs concerning the general state of the economy and unemployment expectations are obtained from the monthly Consumer Survey of the European Commission. This survey has been conducted in the European Union Member States and candidate countries by partner institutes on behalf of the European Commission according to a commonly agreed methodology since 1961. ⁶⁸ The consumer survey is comparable with the University of Michigan Consumer Sentiment Survey for the United States. For this survey, two thousand respondents among the German population older than fourteen years are polled through in-home, computer-assisted face-to-face interviews during the first two weeks of every month. Sampling is stratified by gender, age, occupation, household size, federal region, and community size. The sampling error has a confidence level of 95%, and the error for a 10% share is ±1.3%. ⁶⁹ The question used to determine public opinion on the expected development of the general state of the economy is, “How do you expect the general economic situation in this country to develop over the next twelve months?” ⁷⁰ The question used to determine public opinion on unemployment expectations is, “How do you expect the number of people unemployed in this country to change over the next twelve months?” ⁷¹ The answers are given on a 5-point ordinal scale (and a don’t know category [N]): “the economic situation will . . .” “get a lot better” (PP), “get a little better” (P), “stay the same” (E), “get a little worse” (M), or “get a lot worse” (MM); “employment will . . .” “increase sharply” (PP), “increase slightly” (P), “remain the same” (E), “fall slightly” (M), or “fall sharply” (MM). On the basis of the distribution of the answer options for each question, the European Commission calculates balances (B) ranging from −100 to +100 on the basis of weighted averages as B = (PP + ½P) − (½M + MM), where PP + P + E + M + MM + N = 100 (N is the percentage of respondents without any opinion). The balances are seasonally adjusted. Hence, the PE indicators are a measure of the average and the frequency of expectations.

Figures 1 and 2 plot EI, PE, NT, NC, and NV, for the general economy and for employment, respectively, against time. For equivalent display and analysis, all data standardized by z-transforming are therefore mean-centered to zero.

OPEN IN VIEWER

Figure 1.

Production index, news on the general state of the economy, and PEs for the general state of the economy.

OPEN IN VIEWER

Figure 2.

Unemployment rate, news on employment, and PEs for unemployment.

Measurement

The temporal relation between EIs, news coverage, and PEs is estimated using VAR modeling ⁷² and Granger causality. ⁷³ The VAR approach is to write a symmetrical system of equations that allows dynamic and interdependent analyses between the variables of interest. ⁷⁴ A VAR model including the past (lags) of the dependent and independent variables of the forms

Δ Y t = α 1 + β 1 Δ Y t - n + λ 1 Δ X t - n + ε 1 t , and

Δ X t = α 2 + λ 2 Δ X t - n + β 2 Δ Y t - n + ε 2 t .

is applied to the data, where α is the constant, β₁ and λ₂ are the autoregressive coefficients of the dependent variable, λ₁ and β₂ are the coefficients of the independent variable, and ε is the error term. With VAR models, a variable is described as a result of its own past and that of other variables. Granger causality is based on VAR models and tests whether a variable can be better predicted by using the past of another variable instead of just its own. This concept therefore explores lead–lag relations or forecasting qualities. After each VAR model is estimated, Granger causality tests determine whether X_t−n helps forecasting Y_t (Equation 1a). In addition, the signs (±) of the VAR coefficients reveal whether the relations are positive or negative. This VAR procedure follows recent research on the relations between the “reality,” media content, and public opinion. ⁷⁵

In time series modeling, it is crucial that the data fulfill the condition of (weak) stationarity to avoid spurious correlations appearing with a low Durbin–Watson d statistic revealing autocorrelated residuals, a high R², and high significance of the coefficients. ⁷⁶ The commonly used augmented Dickey–Fuller (ADF) test is performed to detect the unit roots, and the Kwiatkowski–Phillips–Schmidt–Shin (KPSS) test is used to assess trends and level stationarity for each time series. The variables, public opinion and EIs, contain trends and drifts, especially around 2009 during the financial crisis (see Figures 1 and 2) and are non-stationary. NT, NC, and NC are stationary according to ADF and KPSS. ⁷⁷ To eliminate the trend from the non-stationary series and achieve mean stationarity, the common method used is differencing. ⁷⁸ Differentiated time series of order 1 (Y_t − Y_t−1, Δ) for PE and EI are tested as stationary according to ADF and KPSS. Therefore, the analysis is executed based on changes (Δ) in the current month (t) compared with the previous month (t − 1) for PE and EI and based on levels in the current month (t) for NT, NC, and NV. This approach is similar to that of Soroka’s media content time series. ⁷⁹ For the interaction terms, products of non-standardized data are calculated and afterward z-transformed. For NT × NC (PE × NC), a higher consonance should reinforce and a lower consonance should weaken NT (PE) values. Therefore, NC is computed as (1 + M(SD NT)) − SD NT so that the new mean of NC is 1. For PE × NT (EI × NT), negative PE (EI) and negative NT invert PE (EI) into positive interaction values, whereas positive NT weakens negative PE (EI).

The appropriate lag order of the model is selected by using Akaike’s information criterion and Wald lag exclusion statistics. All VAR models satisfy the Eigenvalue stability condition. For testing separate effects of the recession periods 2003/2004 and 2008/2009, the variables are set to zero (mean) during the remaining periods. This procedure allows the analysis of the complete observation period without break while absorbing (non-)recession effects. As the autocorrelation or a non-white noise process of the residuals ε ^ can lead to misinterpretation of the relation between variables, Durbin–Watson test (first-order serial correlation) and a Lagrange-multiplier (LM) test (H0: no autocorrelation at lag order) are performed. The LM test value of the highest lag is reported per VAR model.

Impact of EIs and PEs on News Coverage

RQ1 asks whether real-world EIs or PEs better forecast news coverage (NT, NC, NV) in recessions or non-recession periods. Table 1 shows that NT of news on the general economy is barely set by the corresponding EI or PE but depends strongly on NT of the previous month (column 1, t − 1+***). The negative signs of the EI coefficients reveal that the more positive the past economic development was, the less consonant is NT and vice versa (column 2). The more negative PE was two months ago during the 2008/9 recession and the non-recession period (column 3, t − 2−**, t − 2−*), the more news is published. NT and NC of news on employment strongly depend on past unemployment numbers. The more unemployment there is, the more negative and consonant is NT (column 4). With decreasing unemployment, NT also becomes less consonant, as the positive signs of the EI coefficients reveal (column 5). During the 2008/2009 recession, negative PE, i.e., expecting decreasing unemployment, forecasts a more positive NT (column 4). Yet NV can be forecast neither by unemployment numbers nor by corresponding PE (column 6) but by the NV on the general economy (column 7): The more news there is on the general economy, the less employment news there is in the subsequent month (t − 1−**). The unemployment news tends to be caught up one month later (t − 2+^†). However, LM test indicates that the residuals at t − 2 of VAR models of news coverage of the general economy and NV of employment contain further information. In sum, journalistic interpretation (NT, NC) depends more on EI than on PE. Only journalistic selection (NV) of news on the general economy is partly related to previous changes in PE.

OPEN IN VIEWER

Table 1.

Impact of EI and PE on News Coverage.

	News coverage
	General situation of the economy			Employment
Dependent variable	NT	NC	NV	NT	NC	NV	NV
Past NT/NC/NV	t − 1+***, t − 2+ †	t − 1+**	t − 1+***, t − 2+**	t − 2+*	t − 1+*	t − 1+*	t − 1+*
ΔEI 2003/2004	n.s. (2.532 † )	t − 2− † (n.s.)	n.s. (n.s.)	t − 1− † (5.340**)	t − 1+* (4.582*)	t − 1+ † (n.s.)	n.s. (n.s.)
ΔEI 2008/2009	n.s. (n.s.)	t − 1−*, t − 2−* (4.617*)	n.s. (n.s.)	t − 1−* (5.362**)	t − 1+* (4.446**)	n.s. (n.s.)	n.s. (n.s.)
ΔEI no recession	t − 1+ † (n.s.)	t − 1− † , t − 2− † (3.757*)	n.s. (n.s.)	t − 1− † (8.308***)	t − 1+* (7.041**)	n.s. (n.s.)	n.s. (n.s.)
ΔPE 2003/2004	t − 1+* (n.s.)	n.s. (n.s.)	n.s. (n.s.)	n.s. (n.s.)	n.s. (n.s.)	n.s. (n.s.)	n.s. (n.s.)
ΔPE 2008/2009	n.s. (n.s.)	n.s. (n.s.)	t − 2−** (3.492*)	t − 1−* (2.80 † )	n.s. (n.s.)	n.s. (n.s.)	n.s. (n.s.)
ΔPE no recession	n.s. (n.s.)	n.s. (n.s.)	t − 2−* (2.171 † )	n.s. (n.s.)	n.s. (n.s.)	n.s. (n.s.)	n.s. (n.s.)
NV on the general economy							t − 1−**, t − 2+ † (4.712*)
Constant	−0.008	0.045	0.012	−0.037	0.031	0.010	0.010
Lag structure	t − 1, t − 2	t − 1, t − 2	t − 1, t − 2, t − 3	t − 1, t − 2	t − 1, t − 2	t − 1, t − 2	t − 1, t − 2
N	111	111	110	111	111	111	111
Adjusted R2	.546	.262	.715	.425	.298	.108	.172
F	10.266***	3.672***	13.793***	6.674***	4.227***	1.859*	2.341**
RMSE	0.684	0.872	0.543	0.749	0.834	0.952	0.917
Durbin–Watson d statistic	2.081	2.079	2.088	1.863	2.006	1.962	2.018
LM test (chi2)	95.509*** (t − 2)	91.896*** (t − 2)	69.180* (t − 3)	51.982 (t − 2)	52.891 (t − 2)	66.458* (t − 2)	95.675* (t − 2)

Note. Δ: first differences, otherwise levels. Cells contain the lags and signs of significant VAR coefficients with Granger causality F values in parentheses. For Granger causality, the null hypothesis that the coefficients on all the lags of an endogenous variable are jointly zero is tested. EI for general economy is the production index; EI number for employment is the unemployment rate. The PE question for the general economy is, “How do you expect the general economic situation in this country to develop over the next twelve months?” The PE question for employment is, “How do you expect the number of people unemployed in this country to change over the next twelve months?” n.s. = not significant; NT = news tone; NV = news volume; NC = news tone consonance; EI = economic indicator; PE = public expectation; RMSE = root mean square error.

†

Significant on the 10% level. *Significant on the 5% level. **Significant on the 1% level. ***Significant on the 0.1% level.

Impact of News Coverage and EIs on PEs

Table 2 shows results for PEs being the dependent variable. Variables and interaction terms are added to the models step by step. As indicated by Granger causality and the level of significance of the coefficients, the production index (EI) can better forecast PE for the general economy (H1) and even better for unemployment (H2) than can NT, NC, or NV for the total observation period (Models 1a and 1b′). When the production index increases, PE for the general economy will become more optimistic (t − 3+*), and a decrease in unemployment is expected (t − 1−**, t − 3−*). Yet changes in real-world unemployment numbers cannot forecast public unemployment expectations (1b and 2b). Therefore, H1 is accepted in part and H2 in full.

OPEN IN VIEWER

Table 2.

Impact of EI and News Coverage on the PE for the General Economy and National Unemployment.

Dependent variable	PE on the general economy				PE on national unemployment
Model	1a	2a	3a	4a	1b	1b′	2b	3b	3b″	4b	4b″
ΔPast PE	t − 1+ †	t − 3+ †	t − 1+ †	t − 1+*	t−3+ †	n.s.	n.s.	t − 3+*	t − 2− †	n.s.	n.s.
ΔEI 2002-2011	t − 3+* (2.942 † )	t − 3+* (2.514 † )			n.s. (n.s.)	t − 1−**, t − 3−* (4.400**)	n.s. (n.s.)
ΔEI 2003/2004			n.s. (n.s.)	n.s. (n.s.)				t − 1+ † (3.142*)	t − 1−*, t − 3− † (2.169 † )	t − 2+ † (3.414*)	n.s. (n.s.)
ΔEI 2008/2009			n.s. (n.s.)	n.s. (n.s.)				n.s. (n.s.)	t − 1−* (n.s.)	t − 2+ † (n.s.)	n.s. (n.s.)
ΔEI no recession			t − 3+ † (n.s.)	t − 3+* (2.170 † )				t − 1+ † (3.338*)	t − 1−* t − 3– † (2.708*)	t − 2+ † (3.424*)	t − 1− † (n.s.)
NT 2002-2011	t − 1+ † , t − 3− † (n.s.)	n.s. (n.s.)			n.s. (n.s.)	n.s. (n.s.)	n.s. (n.s.)
NT 2003/2004			n.s. (n.s.)	n.s. (n.s.)				t − 2+* (n.s.)	t − 3+* (n.s.)	n.s. (n.s.)	n.s. (n.s.)
NT 2008/2009			t − 1+ † (n.s.)	n.s. (n.s.)				t − 3+* (n.s.)	t − 1−*, t − 3+ † (n.s.)	n.s. (n.s.)	n.s. (n.s.)
NT no recession			n.s. (n.s.)	n.s. (n.s.)				t − 2+* (2.097 † )	n.s. (n.s.)	n.s. (n.s.)	n.s. (n.s.)
NC 2002-2011	t − 3− † (n.s.)	n.s. (n.s.)		n.s. (n.s.)	n.s. (n.s.)	n.s. (n.s.)	n.s. (n.s.)			n.s. (n.s.)	t − 2− † (n.s.)
NV 2002-2011	t − 3+* (n.s.)	n.s. (n.s.)		n.s. (n.s.)	t−1−* (2.680 † )	t − 1−* (2.169 † )	t − 1− † (n.s.)			n.s. (n.s.)	n.s. (n.s.)
NT × NC 2002-2011		t − 3− † (n.s.)		n.s. (n.s.)			n.s. (n.s.)			n.s. (n.s.)	n.s. (n.s.)
NT × NV 2002-2011		t − 1+* (2.828 † )		t − 1+**, t − 2− † (5.172**)			n.s. (n.s.)			n.s. (n.s.)	n.s. (n.s.)
NT × EI 2002-2011		n.s. (n.s.)		n.s. (n.s.)			t − 2+ † (n.s.)			n.s. (n.s.)	n.s. (n.s.)
Constant	0.001	0.004	−0.001	−0.038	0.005	0.002	0.008	0.017	−0.008	0.013	−0.001
Lag structure	t − 1, t − 3	t − 1, t − 3	t − 1, t − 2, t − 3	t − 1, t − 2, t − 3	t − 1, t − 2, t − 3	t − 1, t − 2, t − 3	t − 1, t − 2, t − 3	t − 1, t − 2, t − 3	t − 1, t − 2, t − 3	t − 1, t − 2, t − 3	t − 1, t − 2, t − 3
n	110	110	110	110	110	110	110	110	110	110	110
Adjusted R2	.194	.312	.098	.238	.132	.194	.176	.153	.170	.139	.092
F	3.480***	3.815***	1.500 †	1.891*	2.017*	2.650**	1.871*	1.873*	1.993*	1.440 †	1.263
RMSE	0.249	0.230	0.264	0.243	0.251	0.242	0.245	0.248	0.245	0.250	0.257
Durbin–Watson d statistic	2.032	2.020	2.051	2.117	2.072	2.012	2.137	2.128	1.963	2.098	1.995
LM test (chi2)	27.153 (t − 3)	60.812 (t − 3)	63.628 † (t − 3)	194.642** (t − 3)	23.295 (t − 3)	26.890 (t − 3)	61.610 (t − 3)	47.705 (t − 3)	65.002 † (t − 3)	140.839 (t − 3)	174.078* (t − 3)

Note. ′: EI = real-world economic indicator production index. ″: EI = real-world economic indicator production index and NT, NC, NV = news on the general economy. Δ: first differences, otherwise levels. Cells contain the lags and signs of significant VAR coefficients with Granger causality F values in parentheses. For Granger causality, the null hypothesis that the coefficients on all the lags of an endogenous variable are jointly zero is tested. The PE question for the general economy is, “How do you expect the general economic situation in this country to develop over the next twelve months?” The PE question for employment is, “How do you expect the number of people unemployed in this country to change over the next twelve months?” PE = public expectation; EI = economic indicator; NT = news tone; NC = news tone consonance; NV = news volume; RMSE = root mean square error; n.s. = not significant.

†

Significant on the 10% level. *Significant on the 5% level. **Significant on the 1% level. ***Significant on the 0.1% level.

The interaction effect between NT and NC shows a negative sign of the coefficient for PE for the general economy (2a). Hence, the less consonant it is, the more NT tends to affect PE for the general economy. This interaction effect is not significant for unemployment expectations (2b). Therefore, H3 is dismissed.

There is also no support for H4 stating that the higher the unemployment numbers, the smaller the effect of news on PE. But NT effects tend to increase with higher unemployment numbers (2b, row NT × EI, t − 2+^†). This effect differs not significantly from zero in other models.

Model 3a relates to H5 and reveals that EI of three months ago tends to forecast PE on the general economy in non-recession times whereas NT of the previous month tends to forecast PE in the 2008/2009 recession. This NT effect increases with higher NV (4a, t − 1+**). Unemployment numbers positively forecast and Granger cause public unemployment expectations in non-recession times and the weaker 2003/2004 recession (3b). Employment NT surprisingly positively forecasts PE on national unemployment in all periods, i.e., the more negative the news, the less unemployment the public expects (3b). Also NT on the general economy (3b″) cannot forecast the public unemployment expectations in the expected direction in the 2003/2004 recession (t − 3+*) but can do so to some degree in the 2008/2009 recession (t − 1−*). To sum up, PE was not set by EI during the severe 2008/2009 recession. Only NT for the general economy of the previous month can forecast both PE for the general economy and for unemployment during the 2008/2009 recession in the expected direction. Therefore, H5 is accepted in part.

The role of NV is tested directly in Models 1a und 1b for the general economy and unemployment, revealing that more news leads to more optimistic PE for the economy in three months (t − 3+*) and the expectation of a decrease in unemployment numbers in the next month (t − 1−*). The interaction effect of NT and NV reveals that the NT effect on the PE for the general economy tends to increase with higher volume (2a and 4a, row NT × NV, t − 1+*, t − 1+**, respectively).

Concerning the quality of the estimations, residuals of Models 3a, 3b’’, 4a and 4b’’ contain further information according to the LM test. For PE for the general economy, Model 2a is the most efficient estimation with an explained variance of 31% (adjusted R²). However, it is not sufficient to explain national unemployment expectations with unemployment numbers or the production index and unemployment news or news on the general economy as adjusted R² allows for more explaining variables.

Impact of News Coverage and PEs on EIs

The results in Table 3, Model 1a, reveal that PE on the general economy positively forecasts the production index. More optimistic PE of the previous month Granger causes an increase in the production index (t − 1+** (3.842*)). Therefore, H6 is accepted. In contrast, unemployment expectations are not found to have the capacity to Granger cause the unemployment rate (1b, 2b).

OPEN IN VIEWER

Table 3.

Impact of News Coverage and PE for EI.

	EI
Dependent variable	Production index		Unemployment rate
Model	1a	2a	1b	2b
ΔPast EI	t − 3+ †	n.s.	t − 1+***, t − 2+**	t − 1+***, t − 2+*
ΔPE 2001-2011	t − 1+** (3.842*)	n.s. (n.s.)	t − 1+ † (n.s.)	t − 1+* (n.s.)
NT 2002-2011	n.s. (n.s.)	n.s. (n.s.)	n.s. (n.s.)	n.s. (n.s.)
NC 2002-2011	n.s. (n.s.)	t − 3− † (n.s.)	t − 2− † (n.s.)	n.s. (n.s.)
NV 2002-2011	n.s. (n.s.)	n.s. (n.s.)	n.s. (n.s.)	n.s. (n.s.)
PE × NT 2002-2011		n.s. (n.s.)		t − 3+ † (n.s.)
PE × NC 2002-2011		t − 3− † (n.s.)		n.s. (n.s.)
PE × NV 2002-2011		t − 2− *(n.s.)		n.s. (n.s.)
Constant	0.012	0.022	−0.004	−0.003
Lag structure	t − 1, t − 2, t − 3	t − 1, t − 2, t − 3	t − 1, t − 2, t − 3	t − 1, t − 2, t − 3
n	110	110	110	110
Adjusted R2	.136	.203	.651	.651
F	2.060*	2.090**	14.340***	9.325***
RMSE	0.210	0.202	0.044	0.044
Durbin–Watson d statistic	1.959	1.881	2.011	2.084
LM test (chi2)	46.347** (t − 3)	72.435 (t − 3)	23.295 (t − 3)	87.287* (t − 3)

Note. Cells contain the lags and signs of significant VAR coefficients with Granger causality F values in parentheses. For Granger causality, the null hypothesis that the coefficients on all the lags of an endogenous variable are jointly zero is tested. EI for general economy is the production index; EI number for employment is the unemployment rate. The PE question for general economy is, “How do you expect the general economic situation in this country to develop over the next twelve months?” The PE question for employment is, “How do you expect the number of people unemployed in this country to change over the next twelve months?” EI = economic indicator; PE = public expectation; NT = news tone; NC = news tone consonance; NV = news volume; RMSE = root mean square error; n.s. = not significant.

†

Significant on the 10% level. *Significant on the 5% level. **Significant on the 1% level. ***Significant on the 0.1% level.

News coverage has no direct unconditional effect on the production index in Model 1. Adding interaction terms between PE and news coverage improves the model quality for the production index (2a). Interaction terms reveal that the forecasting ability of PE for the production index tends to decrease with less consonant NT (t − 3−) and less NV (t − 2−). Hence, news coverage is not positively related to the production index. Therefore, H7 is rejected.