Biases in Fiscal Multiplier Estimates

The "true" size of fiscal multipliers is widely debated by economists and policy makers as large (small) multipliers provide arguments to expand (cut) public spending. Within a meta-analytical framework, we ask whether the large observed variance in multiplier estimates can be explained by the national imprint and various author incentives. For this purpose, we use data on economists' personal characteristics including results from a selfconducted author survey. Our evidence is consistent with the hypotheses that the national background of researchers and the interests of donors financing the research matter for the degree and direction of multiplier estimates.<br><br>These potential biases largely disappear for teams of international co-authors.


Introduction
This paper tests for the presence of biases in the literature on fiscal multipliers.
Fiscal multiplier estimates are an important input for policy design: they measure the impact of discretionary fiscal policy on output. Multipliers are typically defined as the ratio of a change in output at a particular horizon as a response to a change in fiscal policy (see, e.g., Batini et al. 2014).
These estimates range widely for objective reasons. For example, they may differ across policy instruments (e.g., spending or taxes), time-horizons, business cycles, monetary environments, geographic settings, etc. (Ramey 2019). A meta-analysis of 104 scholarly papers by Gechert (2015) reveals a wide distribution of multipliers which range from -0.19 to 2.27 at the bottom and top 5 percentiles of the distribution respectively. Figure 1 plots the distribution of multipliers separately for general spending, tax reliefs, public investment, and transfers. The means of these multiplier types vary with smaller values for transfers (0.39) and tax cuts (0.52), and larger ones for general spending (0.97) and investment expenditure (1.27).
In addition to such objective reasons that can explain the range of multiplier estimates, it may be the case that researchers' prior beliefs and personal incentives also impact the results (Paldam 2018). In particular, three types of biases might potentially play a role.
First, the economic policy orientation of a researcher may influence her empirical findings (see Section 2.2 below). This is a relevant concern as the size of a fiscal multiplier matters in policy debates on the appropriate level and timing of government spending and taxation, and the potential of anti-cyclical fiscal policy. This means that economists, being not different from laypeople, may have prior beliefs Our study is novel as it is the first that systematically studies the variance of multiplier findings with respect to author background and the three types of potential biases. The further innovation is that, for one important strand of the literature, we look at diverse biases jointly. We also add to the meta-analytical literature by documenting evidence of factors that amplify or mitigate the distortion. As a bias amplifier, we consider the author's active role in media communication. Media presence tends to identify researchers who have a "mission" and, hence, a possibly stronger tendency to present research insights that raise public attention. As a potentially mitigating factor, we analyze (international) co-authorship as fostering impartiality. The presumption is that authors exert mutual control for professional standards so that author-teams might better be able to contain biases compared to a single author.
We study the possible relevance of the above mentioned three biases through a variety of tests. Our database originates from a meta-study on fiscal multipliers by Gechert (2015) that we augment by various author-and funding-related variables.
On the ideological bias, we test whether different proxies for the authors' national imprint correlate with their estimated fiscal multipliers. As proxies we use the author's home country's government-spending-to-GDP ratio and the level of economic freedom. We also add an author-specific preference indicator derived from a self-conducted survey among the authors of primary studies. To measure a potential funding bias, we collect data on project financing as well as study the type of 5 a researcher's workplace. We search for a publication bias first through testing for asymmetries in the relation between the precision and the size of the estimate (in line with Gechert 2015). In addition, we search for systematic differences between journal articles and working papers as well as between estimates of non-tenured and tenured researchers. We measure an author's media involvement through her presence on the VoxEU blog. For the analysis of co-authorship, we distinguish between single authors, co-authors from the same country and co-authors from an other country with the expectation that the latter will provide the strongest checks against a biased research design.
Ultimately, we find evidence that is consistent with the ideology bias. We find a mitigating effect from co-author monitoring that is most pronounced for teams of international authors, and weak evidence for the media-involvement amplifier as well as the funding bias. We do not find any evidence for publication selection in the considered fiscal multiplier studies.
Two clarifying notes of caution are necessary. Even if evidence for biases can be detected, this does not necessarily point to conscious manipulations (Kirchgässner 2014). Instead, a researcher's ideological position or self-interests could impact her choice of modeling or testing approaches. Additionally, certain priors could unconsciously affect the author's trust in differing findings, thus creating a bias of judgment in the selection of results. Moreover, a financing and a publication bias can be largely outside authors' responsibility as donors and editors/reviewers may select researchers on the basis of their (past) results.
The second caveat refers to causality. Our data structure does not offer an opportunity to exploit a natural experiment. Endogeneity concerns differ across hypotheses: they are more pronounced for the funding source but are smaller for 6 the national background. The national background though is still associated with various other dimensions than just economic policy orientation, and we cannot exclude that omitted variables drive the results. Thus, we are able to show to which extent observable correlations are consistent with the existence of biases (and their antidotes) without claiming the identification of the causal channel. We continue to discuss the relevance and impact of a possible endogeneity for several of the key findings in the presentation of detailed results.
The paper is structured as follows: Section 2 summarizes the relevant strands of literature and Section 3 develops the hypotheses. Section 4 describes the data and estimation model. Section 5 presents the results and Section 6 concludes.
2 Relevant Literature

Selective publication of research results
The first related strand of literature comprises of contributions that question the neutrality of quantitative research in economics. Over recent years, metaanalytical approaches have shown that empirical researchers benefit from considerable discretion in selecting results and that they use this freedom. The direction of selection bias will correspond to author interests along various dimensions, such as successful publications, financial interests, and intrinsic or ideological motives (Ioannidis 2005

Impact of prior beliefs
The idea that economists use their discretion to produce research consistent with prior beliefs and a self-serving bias is not new. A first relevant finding is that the significant differences in perceptions, beliefs, and economic policy preferences between economists and non-economists can to some extent be explained by the typical socio-economic status of economists (Caplan 2002;Blendon et al. 1997): economists may have a more optimistic perspective on trade, liberalized labor 8 markets, small governments, and low taxes because they are typically a part of an affluent population segment with above-average incomes and low labor market risks.
Apart from the fact that economists constitute a specific sample of the population, the role of researchers' ideological biases has gained greater attention. For example, in the empirical literature on the deterrence effect of the death penalty, Kirchgässner (2014) finds evidence of an impact from prior political convictions on results, even if advanced statistical methods are applied.
The link between national tradition and economic policy preferences is another field with mounting evidence on the importance of economists' national imprint.
These aspects have received considerable attention in a macro-context since the outbreak of the euro area debt crisis. The discussion on how to appropriately manage the crisis has revealed systematically different views among both politicians and economists from Northern and Southern Europe. Alesina et al. (2017) show that the economic integration in Europe has not reduced the heterogeneity of general norms (including the value of hard work or obedience). Guiso et al. (2016) analyze the Greek crisis and identify a cultural clash as a fundamental cause. Dyson (1999)

Funding bias
Funding-induced biases have received considerable attention in the pharmacoeconomic discipline through comparisons between publicly-and industry-financed research results on new drugs. Some examples that find a positive correlation between private sponsorship and a favorable test outcome for the pharmaceutical company are given by Friedberg et al. (1999), Baker et al. (2003), and Bell et al. (2006). Bekelman et al. (2003) review that a conflict of interest in biomedical research through financial relationships between researchers and industry alter research results. This is especially surprising as most of these studies make use of randomized control trials, which are otherwise thought of as the "gold standard" of empirical research. For a systematic overview of research on pharmaceutical industry funding and its impact on study outcomes see Sismondo (2008).

10
Evidence for funding biases are also seen in media studies and health economics.
DellaVigna and Hermle (2014) analyze movie reviews by media outlets which are owned by a production company. Their results suggest the absence of any bias.
With a similar focus but arriving at the opposite conclusion, Dobrescu et al. (2013) test the independence of book reviews when the author is connected to a media outlet. Li (2017) scrutinizes the interdependency of experts' conflict of interest and the quality of their judgment in the context of peer review at the National Institute of Health. Her findings suggest the existence of a bias in favor of projects close to the evaluators' own research.
We add to this literature on funding bias by applying it to macro-economic research. In this context, a possible funding bias is unrelated to any specific private business interest. Instead, it could be the result of public research donors' incentives and their interest in demonstrating the usefulness of public spending.

Publication bias
The publication bias is the most extensively researched bias in empirical economic research (De Long and Lang 1992). It results from the competitive strive of authors for scarce space in reputed journals and can be present when referees, researchers, or editors have an ex-ante preference for statistically significant or other specific results (see, e.g., Frey (2003) for a discussion of editors' and referees' influence on research articles). Doucouliagos and Stanley (2013) find evidence for widespread and substantial publication bias in the majority of the 87 economic areas they study. Similarly, Ioannidis et al. (2017) cover a wide range of fields of economics including international economics, labor economics, growth and devel-11 opment, microeconomics, macroeconomics, finance, and public economics. This "meta-meta-analysis" suggests that publication bias is omnipresent and is closely related to a low power of research designs that "forces" researchers to search for results until effect sizes are so large that they reach significance (Stanley and Doucouliagos 2012).
In his meta-analysis, Gechert (2015) provides an initial analysis for a publication bias in the multiplier literature. Following meta-analytical conventions, he searches for asymmetries in the precision of estimates around the most precise estimate.
Any asymmetry would indicate that published estimates are not a representative sample from the total underlying population. Gechert finds only weak evidence for a publication bias in the multiplier literature. According to his results, if such a bias exists at all, it benefits smaller multiplier estimates. This could relate to the attraction of "surprising" results that challenge conventional wisdom. We take this as our starting point and augment Gechert's approach through tests for the impact of author-specific features that approximate different degrees of publication pressure (e.g., tenured versus non-tenured authors).

Hypotheses
Our hypotheses cover three biases that relate to (i) the impact of an author's own ideological imprint, (ii) donor interests, and (iii) the publication process. In addition, we take account of the bias-enforcing effect of a researcher's involvement in media debates and the bias-mitigating effect of co-authorship.
With the first hypothesis we follow the observation that a researcher's politi- In the context of the multiplier literature, the direction of bias is not obvious.
In this literature, the crucial controversy is on the size rather than on the significance of multipliers. As mentioned in Section 2.4, Gechert (2015) finds weak evidence that the publication bias in the multiplier literature is, if anything, negative. Given these features of the multiplier literature, the search for a publication bias should target two distinctive symptoms that are first, suspicious asymmetries in the precision of estimates around the most precise estimate (see Section 2.4) 14 and, second, a preference for surprisingly small multipliers. The latter can be detected by comparing results from different author types (e.g., tenured vs. nontenured researchers, as both groups differ with respect to publication incentives) or publication types (e.g., working paper vs. journal article, as a working paper represents an earlier stage of scientific production before the editor/reviewer selection sets in). Published articles or researchers with high publication pressure should provide smaller multipliers. Therefore, we construct the following hypothesis: H3: Multiplier estimates are subject to a publication bias that leads to asymmetries in the precision of estimates and, possibly, smaller estimates in published studies (compared to working papers) and from authors with high publication pressure.
As a bias amplifier, we take account of an author's involvement in media debates.
Any such activity can be taken as signal of a "mission" and, hence, a stronger policy interest. We expect that the amplifier can potentially be important for both the ideology bias (H1 ) and the funding bias (H2 ). Researchers with strong positions in public debates might also be more willing to oversell results with the help of like-minded external donors. Note that this bias does not define a genuine direction of bias. Instead, it reinforces an existing primary bias (that originates from national imprint or funding). Therefore, this amplifier is tested through an interaction of the media involvement indicator with the proxy for primary bias. We do not see any theoretical argument to expect an amplifying effect of an authors' media exposure on publication bias (H3 ). This leads us to the following hypothesis: H4: Active participation in the media debate on economic policy increases the effects of country imprint (H1) and financing source (H2) on multiplier estimates.
Our final hypothesis relates to a potential monitoring effect that originates from co-authorship. Several papers show that monitoring agents can decrease tax evasion or corruption. 1 Moreover, it is a robust finding of the tax morale literature that singles are more likely to evade taxes than people living in marriage (Alm and Torgler 2006). The explanation is that close social interactions have a monitoring function that tends to enforce both written and social norms. In this sense, a single agent is less constrained compared to an entity of individuals that has to agree on joint decisions. Likewise, interaction in researcher teams can be expected to activate professional norms and improve authors' respect for high scientific standards.
We expect that mutual monitoring in researcher-teams should mitigate all three biases (ideology, funding, publication). Again, this consideration does not suggest a primary bias but only an effect relative to existing primary biases. In contrast to the amplifier of media involvement, co-authorship should moderate the primary bias. As with H4, the detection strategy will make use of interaction terms. Hence, we evaluate the validity of the following hypothesis: H5: Mutual monitoring from co-authors reduces the biases related to the hypotheses H1 (national imprint), H2 (donor interests), and H3 (publication bias).

Data
We construct our database by combining the meta-analytical data from Gechert We obtain information on authors from hand-collected CVs and personal websites.
This allows us to identify the authors' country in two different definitions: the country where the author received the highest educational degree, and the country of work (at the time of publication). From the CVs, we also collect the institution of employment. From the published (working) papers, we collect information about project grants. Summary statistics for the employed variables are provided in Table A1 in the Appendix. Figure 2 shows the distribution of authors across the countries represented in our sample.
Our proxies to test the impact of national imprint on an author's ideological stance according to H1 are the size of government and the degree of economic freedom. To limit issues from omitted national variables, we also add an authorspecific measure. To obtain this individual score, we conducted a survey among all authors to learn more about their policy preferences in macro-policy debates.
From mid February to mid-March 2019, we contacted 159 of the authors and received 54 replies (34%). Figure A1 in the Appendix shows the questionnaire comprising seven statements on fiscal and monetary policy issues. Researchers could agree (= 9) or disagree (= 1) with the statements in incremental steps of 1. Based on their responses, we calculate a dummy that classifies an author as market-orientated. 2 We also employ four further data sources to obtain a proxy for the market orientation of authors that did not respond to our survey. First, we use data on petitions signed by economists and classified by Hedengren et al. (2010)

Estimation
We conduct a meta-analytical regression analysis to test the hypotheses developed above. Our dependent variable is the fiscal multiplier as it is derived in the underlying primary study. No further normalization is needed as this measure is already dimensionless and comparable across all studies (Gechert 2015). Our unit of observation is the author-estimate. Hence, one estimate from a n-author team provides n observations. In order to prevent studies with multiple authors to have a larger weight in the analysis, we weight each observation by the inverse of the number of authors. We specify the estimation model as follows: where mult ai is our dependent variable and captures the size of the fiscal multiplier estimated. The index a denotes the author and i is the particular estimate from this author. Individual researchers may be the author of several papers and many articles contain numerous estimates due to different specifications and robustness checks. The coefficient β 1 represents our coefficient of interest and measures the impact of the bias inducing source. Source ai can be author-dependent (e.g., national background) or study-dependent (e.g., project grant). 6 M odel X ai covers controls about the model employed in the study (e.g., Real Business Cycles (RBC), VAR, DSGE, etc.). 7 T ype X ai accounts for the type of multiplier (e.g., spending or tax multiplier) with its obvious relevance for size differentials. Country X ai includes a battery of dummy variables for the country coverage of the underlying study. Finally, X ai summarizes other controls such as, for example, the time horizon of the study. The error term is clustered at the study level.

National Imprint (H1 )
To study the impact of an author's national background, we need an indicator that rates countries according to how free-market oriented they are. For that purpose, we make use of two proxies: the government expenditure-to-GDP ratio and the Economic Freedom of the World indicator (EFW) provided by the Fraser Institute.
The EFW is an index that ranges from 0 to 10 where a higher value reflects more economic freedom. 8 We measure the country indicator in the year of the (working) paper publication.
For the expenditure ratio and the EFW, our national imprint hypothesis predicts a positive and negative coefficient, respectively. This represents the view that living in a more pro-market country with a smaller government reflects a government-skeptical position and causes a bias towards smaller fiscal multipliers.
The underlying assumption is one of revealed preferences: through the observable size of government and the extent of governmental interference with market processes, a country's population reveals its fundamental economic policy preferences.
Thus, we are able to test whether authors, in their research, are influenced by the overall policy orientation of their country of origin.
For internationally mobile researchers, "country of origin" is, of course, ambiguous. Therefore, we work with two different definitions: country of work and country of the highest educational degree.
Besides the two country indicators, we also use our author-specific indicator of policy orientation. This author-specific indicator provides a particularly important robustness check as a correlation between our two country indicators and the size of multipliers can be driven by omitted national variables. If results for the author indicator are similar to those for the country indicators, this signals that the empirical support for H1 is not merely an artifact driven by omitted national variables.  (1)-(4), (5)-(8), and (9)-(12) present various specifications for our three indicators of ideological orientation: expenditure ratio, economic freedom, and our author-specific indicator of ideological orientation, respectively.
For the expenditure ratio and economic freedom regressions, we provide two variants that relate those indicators to either the author's country of workplace or education. For both country definitions, an obvious endogeneity exists. First, economists (or students in economics) might migrate to those countries that offer a public sector in line with their preferences so that neither the country of education, nor of work is truly exogenous. However, this kind of endogeneity does not compromise our testing strategy. If such a Tiebout-migration does actually characterize economists' location choices, this would even strengthen the case that a country's governmental features are a useful proxy for author ideology.
For the author-specific dimension, we use the indicator that takes our survey results only, and the augmented one that adds data on campaign donations and other sources. All specifications are presented with and without country fixed effects. We always include the full set of control variables accounting for the type 22 Electronic copy available at: https://ssrn.com/abstract=3420838 of multiplier, model, country-coverage 9 , and other features of the underlying study.
Detailed results for all these controls are presented in Table A2 in the Appendix. 10 The results are consistent with our hypothesis in all but one specification. Authors from countries with larger governments or lower economic freedom come up with larger multipliers. The same holds for authors that are classified as having a pro-government orientation through our author-specific indicator. With one exception, all specifications that use country fixed effects are estimated with high statistical precision. Effect sizes are fairly large: A 10 percentage point increase in spending-to-GDP ratio increases the fiscal multiplier by 0.07 to 0.47 points on average (or 8-55% of the mean). A one point increase in the EFW indicator is associated with a decrease in the multiplier of up to 0.62 points. Lastly, the average difference between market-and government-oriented researchers amounts to a magnitude of between 0.1 and 0.21 points in the size of multipliers.
9 Country fixed effects refer to author country; country-coverage controls refer to the country groups that are included in the primary study. The inclusion of the latter is a safeguard against the risk that a correlation between country origin and country coverage drives the results.
10 The point estimates of the controls for the model and multiplier type presented in Table  A2 in the Appendix are in line with the findings of Gechert (2015). For example, relative to government consumption, investment multipliers are large and tax or transfer multipliers small. Among the models, RBC approaches tend to arrive at the smallest multipliers as expected.

Funding (H2 )
As formulated in H2, the funding bias should lead to higher multiplier estimates for government-financed studies. In the following, we test this hypothesis for direct funding through research grants and indirectly through institutional funding. Table 2 summarizes the various specifications. As in the preceding section, all regressions include the full set of control variables as previously described. Likewise, we again provide specifications with and without country fixed effects.
Columns (1) and (2)  In order to obtain a more detailed picture, we differentiate between the various sources for project grants in columns (3) and (4). These finer-grained results indicate that grants coming directly from national institutions (either the national government, the national science funding agencies, or the central banks) are associated with higher multiplier estimates. Interestingly, this effect is only precisely estimated for projects funded by national science agencies and partially for central banks, though not for those which received a research grant directly from the government. Grants from privately financed foundations or from research institutes are significantly associated with smaller multiplier estimates. This supports the 25 hypothesis as these donors should not share the same interest in proving the government to be efficient. 11 Columns (5) and (6) look at the impact of government financing when it is given in an indirect way to finance the researchers' workplace. The reference workplace is university employment. 12 The positive correlations for government and the negative for private institutions correspond to our expectations. However, all estimates lack statistical precision. The weaker link compared to project grants is not surprising since project grants offer a more direct channel for bureaucratic and political influence on research outcomes compared to institutional financing.
Overall, we find some evidence for a funding bias. As usual in the funding bias literature, we abstain from speculating on the direction of causality that drives the correlation. It is only one possible case that the source of funding has an impact on the conduct and results of a research project. It may also well be the case that a researcher's (prior) work changes her success rate in obtaining external finance.

Publication Bias (H3 )
We test for a publication bias by means of three approaches. First, we search for asymmetries in the relationship between an estimate and its precision. Second, we look for systematic differences between journal articles and working papers. Finally, we ask whether non-tenured researchers (due to higher publication pressure) come up with different estimates on average than tenured researchers. We base 11 The coding of identified grants into the 5 categories is depicted in Table A3 in the Appendix. 12 The coding of workplaces into the 4 categories is depicted in Table A4 in the Appendix.

27
Electronic copy available at: https://ssrn.com/abstract=3420838 our second and third approach on Gechert's (2015) finding that the publication bias in the multiplier literature, if it exists, favors smaller estimates (that may be perceived as an interesting surprise and a challenge to conventional wisdom).
To assess asymmetries in statistical precision, one would usually rely on the standard errors of the respective estimate (Doucouliagos and Stanley 2009). However, the standard errors of the employed studies are not readily comparable such that, similar to Gechert (2015) and suggested by Stanley and Doucouliagos (2012), we rely on the number of observations used to obtain the multiplier estimates as a second best proxy for precision. We start out with a graphical investigation of the relationship between the multiplier estimates and the underlying number of observations. Figure 3 depicts a funnel graph with the two variables. No asymmetries are visible which points to the irrelevance of the publication bias in that literature.
Since any such graphical investigation is prone to subjective (mis-)interpretation, we additionally rely on an econometric analysis and estimate the following simple model similar to Doucouliagos and Stanley (2009): where f (N i ) are various functions of the number of observations N i which where used to estimate the fiscal multiplier mult i in paper i. The error term continues to be clustered at the study level. The results are presented in Table 3. For columns (1)-(4) and (7)-(10), a publication bias would imply statistically significant coefficients for f (N ), whereas for columns (5) and (6) one would observe a statistically significant intercept. Since none of the specifications suggests the presence of such a bias, we do not find any evidence for its relevance in the underlying fiscal multiplier studies. Columns (1) and (2) of Table 4 answer the direct question of whether the size of multipliers reported in journal articles with their entry barriers differs systematically from those in working papers or other non-journal publications. Depending on the inclusion of country fixed effects, the sign switches and results are far from being statistically significant.
Columns (3) to (6) present regression results that look into the findings of researchers that, due to safe academic positions, are under lower publication pressure compared to non-tenured colleagues. Non-tenured researchers tend to produce larger multipliers, which runs against the expectation that the publication bias would favor smaller multipliers. None of the coefficients is significant. As an alternative proxy for publication pressure, we look at researchers with a full 29   (5) and (6) we follow Doucouliagos and Stanley (2009) and weight the fiscal multiplier estimates as the dependent variable by log(N ) and √ N , respectively. professorship versus those without. These results again do not support the hypothesis. Overall, there is no support for a significant publication bias in the multiplier literature.

Media Involvement (H4 )
Hypothesis H4 claims that researchers who actively participate in media debates push their opinions in their research as well. We would then expect that both the ideological bias and the funding bias are reinforced. We therefore focus on the interaction of the H1 -and H2 -related variables with our indicator for media involvement. We proxy media involvement through an author's presence on the  With statistical significance, this is the case for the educational definition of author origin. For the interaction with the economic freedom indicator in models (5) to (8), we again find the expected (negative) sign for the country of education but without statistical significance. 14 Results are therefore mixed and provide only indicative evidence for an amplifying effect of media involvement. Table 6 illustrates the interaction between media involvement and the funding bias. The findings do not provide for a strong confirmation of the hypothesis.
The key interaction is insignificant for grants in general (columns (1) and (2)).
Looking at specific funding sources (columns (3) and (4)), only the interaction with private donors yields a significant estimate with the expected sign. Concerning the workplace perspective (columns (5) and (6)), the interactions rather point 14 Figures A2 and A3 plot the overall marginal effect of media involvement on the multiplier estimates for different levels of government size and economic freedom. We do not have an ex ante expectation about this overall effect and the graphs also do not show any robust systematic relationships.
to a bias-moderating role: authors from government institutions who publish on VoxEU show a diminished tendency to come up with large multipliers.

Co-authorship (H5 )
In this final section, we summarize the evidence on the bias-mitigating effects of co-authorship. We limit the analysis to the ideological bias H1 and the funding bias H2. Since we could not detect any hint of the existence of a publication bias, the search for a counterbalancing effect from author teams is senseless.
We start with the ideological bias. We expect that monitoring from international author teams should be particularly effective in mitigating an ideological bias compared to purely national collaborations. Table 7 makes use of a dummy for authors coming from different countries. The effect of co-author monitoring is indeed highly visible and significant through all specifications that assign author preferences to their country of education. In all cases, the interaction counterbalances the original direction of the bias to a large extent and with a high statistical significance. 15 As collaboration of (international) authors appears to be helpful to contain biases from one's national background, mutual monitoring and exchange may equally serve to reduce funding biases. We test for this mitigating effect through interacting our monitoring dummy for the presence of co-authorships with funding indicators (Table 8). For aggregated (columns (1) and (2)) and disaggregated project grants (columns (3) and (4)), the monitoring dummy identifies multiple authors from different countries. For the workplace definition, the monitoring variable is a dummy for multiple authors (columns (5) and (6)) or for teams with at least one co-author working at a non-government institution (columns (7) and (8)).
15 Table A5 in the Appendix presents results for a multiple author dummy that includes purely national teams. Interactions tend to have the right sign but often fail to be significant. For aggregate grants, no significant impact of co-authorship emerges, though there is a clear pattern in line with our expectations for the disaggregation. International co-authorship out-balances the impact of both government grants and private grants on the size of multiplier estimates. More specifically, the signs of the interaction coefficients are reversed compared to the plain effects and the magnitude of the counter-balancing effect is sufficient to neutralize the bias. Similar clear effects cannot be detected for the workplace analysis. However, the absence of a moderating effect is of less relevance given that there was no strong evidence for the existence of a workplace effect at all. Our findings on multiple authors and biases might not necessarily reflect a causal impact of co-authorship. An alternative explanation is that more biased authors may self-select into single-authorship. No matter which of these mechanisms drive the results, the essential finding is that multiplier estimates of (international) author teams tend to show less symptoms of an ideological and funding bias.

Conclusion
It is well known that fiscal multiplier estimates vary largely because of different country contexts, multiplier types, or econometric models. Our contribution is that we shed light on more subjective reasons behind the observed variance of estimates related to authors' ideology, incentive effects of external research funding, and rules of the academic publication process. We find that the variance of multiplier estimates can indeed be better explained if we take account of author-specific characteristics.
We show that a researcher's economic policy orientation, as proxied either by national background or an author-specific classification, correlates with her estimated multipliers. In addition, our evidence is consistent with the hypothesis that government-financed projects are associated with larger multipliers. Our analysis does not detect evidence for a publication bias in this strand of literature. The analysis delivers some evidence that researchers with an active involvement in media debates are particularly prone to the production of multipliers that support their prior (national) beliefs on the role of government.
This work underlines the need for great caution and scientific neutrality when designing research projects. Moreover, it emphasizes that it is important for policy 37 Electronic copy available at: https://ssrn.com/abstract=3420838 makers to carefully compare various sources when seeking guidance from empirical research and to take into account the conditions and schools of thought under which research projects were conducted.
Our results also suggest that co-authorship in general, and international teams in particular, are an antidote to the distorting effects of national or funding biases.
If this particular insight from our results can also be applied to other strands of literature, this would carry great significance for economics in general. Of most significance in this case is the conclusion that biases, which originate from the narrowness of national debates, might be counterbalanced through more international collaboration and mutual surveillance of research teams.

% of observations
Notes: The graphs show the marginal effects of the VoxEU variable on the multiplier estimates. The regressions include country fixed effects and correspond to specifications (6) and (8) of Table 5, respectively.