Trust in official . An econometric search for determinants. The case of

Dr Serge Allegrezza. Director General. Institute of Statistics and Economic Studies (STATEC)

Paper prepared for the 16th Conference of IAOS OECD Headquarters, Paris, France, 19-21 September 2018 Session 2.C., Day 1, 19/09, 14:30: Communication & statistical literacy, strategic issues

Dr Serge Allegrezza [email protected] STATEC

Trust in official statistics. An econometric search for determinants. The case of Luxembourg

Prepared for the 16th Conference of the International Association of Official Statistics (IAOS) OECD Headquarters, Paris, France, 19-21 September 2018

2

ABSTRACT In a complex society, flooded with a deluge of from the Internet, trust in official statistics and in the national institutes of statistics play a key role in preserving public numbers as a reference and benchmark. This paper takes the view of the demand-side of statistics, investigating the perception of the trustworthiness of public numbers and the functioning of the national institutes of statistics. Trust is a matter of perception and opinion, which is as important as the supply-side view, i.e. compliance with peer reviewed processes and quality standards. The difficulty to assert statistical facts as reality, in particular in a context of “post-truth” social media and populist discourse, has increased.

Representative surveys on the trust of statistics are quite rare and available data is not systematically analysed. Comparative cross country studies at macro and micro level are missing. This study uses two surveys dating back to 2015 and 2017, carried out by an independent pollster (TNS-ILRES) which interviewed over 2500 residents in each wave in Luxembourg. The contains a bunch of questions on trust in institutions (parliament, government, central bank, NSI etc.) and on different types of media channels. The questionnaire asks interviewees about perceived political independence of the NSI, the use of official data in business, research and public debate and, finally, participation in official surveys. There are some questions on political preferences and confidence in fellow citizens. Furthermore, the questionnaire collects data on , income, gender, professional activity and age. Ideally, the econometric analysis (logit regressions) should confirm the independence, neutrality, non-discrimination and high trust of official statistics and the NSI. All the coefficients of variables linked to political independence or political preferences, gender, age and class (education and income) revealing some discrimination in the perception of official statistics, should not be significant. There should be, ideally, a positive effect of confidence resulting from the protection of personal data, the participation in official (compulsory) time consuming surveys and from the utilization of statistics. The analysis of data related to 2015 and 2017 strongly confirms the political independence of statistics and of the NSI (STATEC), the importance of data protection, the use of statistics and the importance of confidence in media as main distribution channel.

The conclusion advocates systematic and in-depth analysis of micro data on various dimensions of trust in statistics and their determinants, over time and across countries. The success of “fact checking” activities, training and tailor made communications (narratives for precise targets) might also be assessed by systematic and representative surveys on trust and numeracy. To gain more insights and share recommendations, I advocate that those surveys should be based on a common questionnaire (in the framework of the OECD) and exploited according to an agreed methodology.

Keywords: Trust in official statistics, cognitive science, 3

Introduction

“Damned lies and statistics”. Funny jokes about statistics are popular and stories about the way unabashed politicians or ruthless advertisers misuse statistics are quite familiar. In this paper, we will examine trust in official statistics and its determinants. Trust in official statistics, both in the official numbers and in the organizations producing the numbers, depends on trustworthiness, the standards by which the production process works and the availability of data to the public. Trust in statistics is defined and measured by the guidelines set up by the OECD. An early version of the paper1 was presented at the Marrakech ISI conference in July 2017 analysing the data from the 2015 trust survey. This paper has been extended to incorporate the most recent survey on trust in statistics of 2017 in the case of Luxembourg and deepens former research on this topic (Allegrezza, 2014). In the first part, the paper is arguing why should focus on perception, underscoring the subjective view of “official numbers” and of the functioning of NSIs. Perception, with all its cognitive flaws and biases, as described in the emerging cognitive science literature, is an essential element to assess the trustworthiness of official statistics, both numbers and organizations producing the latter. The second part of the paper describes the data at hand and the relevant variables used in a bunch of logit regressions. The results of the analysis of the two data sets are compared. All in all, the findings confirm that, all else equal, personal data protection and political independence have a positive and significant impact on the probability of trust in official statistics and the NSI. Unfortunately, there is knowledge segregation: more educated people have more trust in numbers than the less educated.

Trust and cognitive barriers to official statistics

Trust is a complex phenomenon and the communication between official statistical institution and the public is not as straightforward as the basic “sender-receiver” scheme might suggest. There are few empirical studies available on trust in statistics and their determinants. The receivers of statistical information are very heterogeneous, endowed with different knowledge competencies in statistics and social sciences, have different expectations or wishes and are limited by time and money constraints. We propose to dig deeper in the relationship between the perception of trustworthiness, and the actual use of statistics by focusing on the subjective evaluation of the trustworthiness of official statistics produced by the national statistics institute in Luxembourg (STATEC). The communication of data and information has changed dramatically since the Internet and the social media are regarded as the main channels to access information. There seems to be some anxiety about a growing defiance of science and expertise in general, despite a growing level of education in our countries. It is difficult to assess quantitatively to what extent the limitations in literacy, numeracy and the resistance towards science are increasing2. In a

1 “Statistics in a post-truth society. Determinants of confidence, independence and usage of official statistics” 2 Eurobarometer survey (2010) suggests that reluctance towards science is significant (in 2010, 38% think we depend too much on science and not enough on faith) but this proportion is decreasing.

4 more general context, trust vis à vis public institutions (government, parliament, bureaucracy) and trust in other fellow citizens might be linked to trust in official statistics, interesting enough to be explored empirically.

The suspicion of statistics has multiple reasons

The first reason is certainly rooted in the awkward representation of society as a whole, structured around Quetelet’s “average man” wiping out uniqueness of individual characters, peculiar situations, contexts and biographies inherent to a single person. Olivier Rey, a French philosopher and mathematician, who has studied the emergence of numbers as convenient way of depicting the world, puts forcefully forward the denial of the singularity and distinctiveness of an individual by official numbers. He claims that this is a major cause explaining frustration with statistics. William Davies on the same lines argues that “both statisticians and politicians have fallen into the trap of talking like a state, giving the impression of having lost touch with single citizens3. The second reason of suspicion: statistics is a branch of mathematics. Probability for example entails a way of thinking which is distinct from everyday thinking. Our brains are wired in such a way that we need a conscious effort to dismiss the default of reasoning: using Bayes theorem does not come to our mind spontaneously and many solutions are counter intuitive. Thinking statistically is hard, as Nobel Prize winner Daniel Kahneman has demonstrated by manifold in his famous book “Thinking, Fast and Slow”.

It has been forgotten that trust in statistics is the result of a long and painful history. Theodor Porter (1996) in his book “Trust in numbers” makes it clear that official statistics can’t be understood properly unless examined through the lense of science history. The authority of official statistics is linked to the progressive emergence of quantification, the standardization of measurements and the process of validation of social numbers. Objectivity, taken as a synonym of realism, has been cultivated by promoting rules of fairness, impartiality, impersonality. Since a genuine, ontological “absolute objectivity” is not possible, scientists must cultivate proxies like “disciplinary objectivity”, guaranteed by specialists or “mechanical objectivity” which is obtained by following rigorous rules that reduce personal biases or preferences. As Porter shows through examples, including the engineering of official statistics, “mechanical objectivity” is difficult to realize fully, because tacit knowledge, experience, wisdom, intuition, skills and craft play an important role in scientific activity. “In public affairs, reliance on nothing more that seasoned judgment seems undemocratic… Ideally, expertise should be mechanized and objectivised. It should be grounded in specific techniques sanctioned by a body of specialists” (p7). The faith in objectivity tends to be associated with political democracy or at least with systems in which bureaucratic actors are highly vulnerable to outsiders”. He recognizes that the “exaltation of objectivity of science is often confused with elitism” (p.75), quantification is a “technology of distance”.

3 "How statistics lost their power (https://www.theguardian.com/politics/2017/jan/19/crisis-of-statistics-big-data-democracy )

5

The struggle to establish bureaus of statistics in Western Europe and to acquire the legitimization by the general public and decision makers went through a long process, well described by Alain Dérosières (2010). The institution of statistics is a political, social and cultural process which rests on a broad consensus and the convenience of the language of quantification. This consensus founding official statistics might fray if the broader social fabric, on which it rests, were to unravel.

New insights on the way that we are using information and transforming it, eventually, to decisions come from cognitive sciences4 which have uncovered many dimensions of the way our brain works. A bunch of complex mechanisms, riddled with various biases, impede “pure reason” to work perfectly. Effective users or the potential users of official statistics are seizing numbers and figures and their interpretations through those distorting cognitive mechanisms.

Mercier and Sperber dismiss the dogma which takes “for granted that the job of reasoning is to help individuals achieve greater knowledge and make better decisions”, but reason is “hopelessly biased and lazy”. In their interpretation, reason’s bias and laziness is a tool for social , quite imperfect, but developed to help people to find “reasons that support their point of view because this is how they can justify their actions and convince others to share their beliefs”. This is a huge barrier to cross to get statistical facts through. In the same vein, Jonathan Haidt (2013) claims that the “mind is divided, like a rider on an elephant, and the rider’s job is to serve the elephant. The rider is our conscious reasoning— the stream of words and images of which we are fully aware. The elephant is the other 99 percent of mental processes— the ones that occur outside of awareness but that actually govern most of our behaviour”. Sentiments are dominating reason and associated numbers.

Behavioural economics offer some further explanations on the way information is processed and decisions are taken, at odds with the canons of the “homo economicus”. Tirole and Bénabou (2017) propose three concepts to cope with irrational thinking: strategic ignorance, self-signalling and reality denial to explain some awkward decision patterns. In psychology, the concept of self-deception is connected to the way our self is playing with data and information. Self-deception has a long history in psychology and philosophy and continues to inspire new perspectives and explanations of the paradox that it entails. The idea that the mind can conceal information from itself is puzzling and disturbing, producing different conceptions of self-deception and different views on the consequences of it (Bachkirova, 2016).

As a consequence of the less than rational behaviour of human beings, the usual linear scheme of communication of NSIs, which confronts, on one side, the supply of accurate, timely and reliable numbers and, on the other side, the demand for objective information, in order to allow rational decisions, is naive and therefore useless. In fact, the communication between producer and user of official statistics seems

4 “Cognitive science is a network of interrelated scientific disciplines engaged in researching human cognition and its brain mechanisms. Cognitive science is made up of experimental psychology cognition, philosophy consciousness, neuroscience, cognitive anthropology, linguistics, computer science and artificial intelligence” (Marina Bogdanova, 2017). We would add behavioral economics. 6 much more complex in an information society where citizens, consumers and business are drowned in a deluge of data of all sorts which they get mostly for free in traditional and social media. The issue of trustworthiness in numbers and their concrete use in various situations has been around for quite some time, but has not been, to our knowledge, studied extensively. Lewandowsky Ecker and Cook (2017) made some important contributions on the way contentious (mis)information is used and corrected: “The post-truth world emerged as a result of societal mega- trends such as a decline in social capital, growing economic inequality, increased polarization, declining trust in science, and an increasingly fractionated media landscape. Misinformation in the post-truth era can no longer be considered solely an isolated failure of individual cognition that can be corrected with appropriate communication tools. Rather, it should also consider the influence of alternative epistemologies that defy conventional standards of evidence”.

Ignorance of basic numbers (gdp, inflation, etc.) has probably always been a problem, as result of poor education, especially in mathematics and statistics. The level of literacy and training determine the level of analytical capacity. Ignorance or a lack of competence may nurture defiance of official statistics. But this link between literacy and confidence in numbers has still to be explored. An OECD (2016) study on literacy and numeracy gives some hindsight on skills and proficiency in information processing. The survey realised from 2013 until 2015 in 33 countries showed that a sizable proportion of adults has poor reading and poor numeracy skills (22.7% on average). One in four adults has no or limited experience with computers or lack confidence in their ability to use computers. Proficiency in literacy and numeracy peak at around age 25, older adults score less than younger adults.

Is illiteracy a serious problem? Steven Sloman and Philip Fernbach (2017) underscore the myth of individual thought and the power of collective wisdom (pp. 4-5). They state: “individual knowledge is remarkably shallow, only scratching the surface of the true complexity of the world, and yet we often don’t realize how little we understand. The result is that we are often overconfident, sure we are right about things we know little about.” So Ignorance is the normal state of human society which suffers from a “knowledge illusion”. In this context, post truth is not as a dangerous threat, as generally assumed. Post-truth, which was announced as the international word of the year 2016 in the Oxford Dictionary, is defined as “relating to or denoting circumstances in which objective facts are less influential in shaping public opinion that appeals to emotion and personal belief”. There is another form of post truth or alternative facts blend which has been coined as “bullshit”. Harry Frankfurt (Dieguez, 2018) describes this special form of misinformation as a special blend of lie. “Liars know the truth and try to hide it; bullshitters don’t know or care about the truth and try to hide their lack of commitment to it. Thus bullshitting is more like bluffing or faking. Frankfurt thinks bullshit is more dangerous than lies because it erodes the possibility of the truth existing and being found.

Grasping a scientific reasoning is quite difficult for laymen. As Gorman puts it “We will assert many times that the problem is not simply a lack of information, although that can be a factor. Irrational behaviour occurs even when we know and understand all the facts. Given what we now understand about brain function, it is probably not even appropriate to label science denial as, strictly speaking, “irrational.”

7

Rather, it is for the most part a product of the way our minds work. This that simple education is not going to be sufficient to reverse science denial. Grigoreff et al. (2016) study whether providing information about immigrants affects people’s attitude towards immigration. In a large representative cross-country they show that, when people are told the share of immigrants in their country, they become less likely to state that there are too many of them. In two online experiments in the U.S., half of the participants were fed with statistics about immigration, before evaluating their attitude towards immigrants with self-reported and behavioural measures. This more comprehensive intervention improves people’s attitude towards existing immigrants, although it does not change people’s policy preferences regarding immigration.

Nyhan and Reifler (2010) showed in lab experiments, that citizens bulked at evidence contradicting their partisan opinions and ideological attachments. Rather than ignoring factual information actually they were even more convinced by their prejudices. This is what the authors called the “backfire effect” reflecting that facts actually may compound ignorance. Later research (Porter and Wood, 2017) in the same vein found that the backfire effect may be tenuous. It suggests that citizens heed factual information, even when such information challenges their ideological commitments. Drew Westen in “Political Brain” analyses partisan beliefs.

Ideology and political preferences, partisan views of right against left, progressive versus conservative could play a role as a grid distorting the nature and trustworthiness of figures and numbers and of institutions producing the latter. Some other authors could be of relevance for the issues of trust in science and objective facts. The philosophical context of “post-modernism” has played some role too. In this philosophical current, science, is perceived as dangerous, dominated by business and essentially flawed (see Lyotard, Derrida, Foucault, Rorty…). Those popular philosophers paved the way to post-truth and alternative facts approaches. The German philosopher F. Nietzsche brings the relativist view to the point: "There are no facts, only interpretations”. This is of direct concern for official statistics which asserts to convey objective facts, as a standard to assess “reality”. Steven Pinker in his book “Enlightenment now” pinpoints the change of mood against reason, science, humanism, or progress and the missions of all the institutions of modernity—schools, hospitals, charities, news agencies, democratic governments, international organizations. Since the 1960s, trust in the institutions of modernity has sunk, and the second decade of the 21st century saw the rise of populist movements that blatantly repudiate the ideals of the Enlightenment. Steven Pinker reminds us that counter-Enlightenment has pushed back the ideals of progress and freedom from the very beginning, as for example the Romantic Movement in the 19th century. Populist movements are the heirs of those ancient ideas.

8

Measuring trust in official statistics

The paper is about measuring trust in official statistics and in the credibility of organizations producing them as to support various decision makers and citizens in a democratic society. In this paper, the framework proposed by the OECD to apprehend trust in official statistics has been adopted as a guideline to draft the questionnaire and organize the data analysis. Ignorance of statistical numbers is widely spread; people have a very fuzzy idea of basic macroeconomic indicators, despite being omnipresent and largely commented in the news. In the same context, the OECD has shown in a recent report about adult’s skills in literacy and numeracy that 18.5% of adults, on average, have poor reading skills and poor numeracy skills. Around one in four adults has no or only limited experience with computers or lacks confidence in their ability to use computers (OECD, 2016). A recent report on financial literacy shows that in G20 countries, only 52% of adults on average reached a minimum target of basic skills like establishing a budget (OECD 2017). What level of numeracy is required from citizens in order to understand and use official statistics which are deemed essential for running the , managing government or intervening in public debate?

A Eurobarometer poll (2015) has revealed some striking facts. For instance: 6% of Europeans estimated their national growth rates correctly, 31% don’t know, most overestimate growth rates. None could give the correct answer on inflation (31% did not know the number, which was highly exaggerated compared to the true number). 23% correctly estimated the unemployment rate in their country, 20% don’t know the figure, and all were overestimating the unemployment rate.5 To what extent does this stounding lack of numeracy impact trust in ? The Eurobarometer survey shows that only half of the interviewees trust economic statistics (10% don’t know). There are off course significant differences among countries: it ranks highest in Denmark (83%) and lowest in Cyprus (36%), Luxembourg is 8th (63% expressing trust). I found few analyses, to my best knowledge, of the relations between trust in official statistics and in NSI’s on the one hand and macro- economic factors (economic and social situation, trust in institutions, cultural context…) on the other hand. The same lack of studies, at micro level, exploring relations between trust in numbers and individual characteristics (income, level of education, trust in institutions…) is surprising, given the challenges national statistics are facing.

5 The true number could be questioned, at least in Luxembourg: for example unemployment mixes up the national rate (ADEM) and the harmonized unemployment rate (), the same holds for inflation (national price index versus harmonized European). GDP is blurred by the quarterly accounts and successive annual revisions of the same reference year. 9

Trust in government versus confidence in statistics 80

60

40

20 y = 0.518x + 33.895 R² = 0.4719 0 0 10 20 30 40 50 60

Trust gov Linear (Trust gov)

Trust must be seen in a wider context. For example, using the Eurobarometer data, it is straightforward to depict a linear relation between confidence in official statistics across Europe and confidence in government. This means that when interpreting the trust in official numbers or in statistical offices, the more general context has to be taken into account. Some national statistics institutes collect data on confidence, based on the OECD framework, in official statistics. See for example the studies by New Zeeland and United Kingdom. For our study, we will focus on a survey ordered by INSEE and realized by CEVIPOF. Chiche and Chauvrie (2016) made an econometric analysis of the defiance by the French public in various statistical fields like growth, unemployment, immigration, public deficit etc. explained by variables on age, gender, education level, income, and political preferences. The results show that trust in official numbers is correlated with confidence in political institutions (government, parliament, president). The results show also differences between gender, age, level of education, income and political preferences. Public statistics are not uniformly appreciated by a heterogonous public. The richer and more educated, the more left leaning, the more interviewed persons tend to trust official statistics. There are also different explanatory patterns depending on the statistical fields.

The Luxembourg surveys

The data

Measurement of trust or confidence (the two terms are used indifferently in our study) in statistics as an institution has been conducted in several representative surveys. The surveys were realized to assess the reputation of the Institute, to foster better products and to improve targeted communication. Besides the “peer review”, based on the European code of conduct of official statistics, which aims at assessing the conformity with agreed quality standards, the surveys presented in this study look for the perception of statistics by a representative sample of persons.

10

As official statistics and public numbers should be as objective as possible, the professional reputation of the national institutes is of utmost importance. The questions aim at measuring the degree of confidence or trustworthiness of statistical numbers and the confidence in the national statistical institute (STATEC). The explanatory variables are of two sorts. First, there are variables on confidence. The most contentious is the perceived political independence of the Institute (Independence). The confidence in the handling of personal data (Protection) is very important, given the sensitivity of the public on this issue. The confidence in Media (Media) is important since it represents still the most important channel to spread statistical data, press statements and studies. There are two behavioural variables. The first derives from a question on whether respondents use or refer to statistics published by the Institute (Usage), which refers to their behaviour, validating the trust in official numbers. The second derives from a question asking whether the interviewee participated in a survey realized by the national institute. This variable shows to what extent contributors to an official survey (LFS, EUSILC …), who give some of their precious time to respond to questions, trust the numbers and the Institute (Participant). Second, the surveys collected data on socio-demographic characteristics. There are three variables: income categories (Income), education levels (Education), gender (sec) and age categories (age). The survey of 2015 includes also some questions about economic opinions, used as proxy for political preferences. Questions appreciating public debt level and the public deficit were meant to capture typical attitudes of fiscal conservatives versus fiscal progressives. Some remarks on the expected results: as statistics should be a public good, confidence should be non- discriminatory: no gender and “class differences” by income and education, independence from government and from political interferences. So the best econometric result would consist in no significant effects from those variables on confidentiality. In sharp contrast, participation in surveys and usage of official statistics should have positive, strong and significant effects on trust in statistics. Those are the ideal results we are aiming at. The survey of 2017 includes two other questions about trust in other fellow citizens, family and relatives, neighbours and a question about political preferences (left- right scale).

Results of the analysis

A bunch of various specifications of logit regressions have been run taking as dependent variables: trust in public statistics, political independence, use of statistics by citizens and institutions and a positive opinion on the performance of the NSI (STATEC). The dependent variables of the four equations have been used as determinants (exogenous variables) for the other regressions.

The 2015 survey

Confidence in public statistics is influenced positively by the perception of political independence as well as guaranteeing personal data protection, which is highly and positively significant. Those are the expected effects and reassuring results. The confidence in media, used as main distribution mechanism of statistics, is positively linked to trust in statistics. This is important since news reporting is often biased or contaminated by comments or

11 criticism (online comments on social media). Participating in an almost compelling survey organized by STATEC has no significant impact on trust in official statistics. The level of education has a positive impact on trust in statistics. While gender (man) and income (dummies) have no significant effect. Trust in statistics diminishes with age. Those comments hold for the results of the equation on trust in the NSI (STATEC). Trust in statistics has a significant positive impact on the perception of political independence. Interestingly, the 2015 survey shows a negative impact of nationality (Luxembourger) on political independence. Nationals are less convinced by the political independence of official numbers but show no effect on trust in the NSI. Another interesting result is the use of statistics which is positively impacted by participating in an official survey. This suggests that interviewees are more prone to use statistics when they are becoming aware of them by giving data to the NSI.

Table 1. Determinants of Trust in statistics, Political independence, Usage of statistics and trust in NSI (Survey of 2015, Basic Model) Trust in Political Usage of Trust in NSI (STATEC) Survey 2015 Public Independence statistics statistics Trust x +++ ++ +++ Independence +++ x NS +++ Political Usage of statistics ++ + ++ +++ Protection of data ++ +++ ++ +++ Participant NS + ++ NS Official survey

Media trust ++ ++ - ++ Sex (male) NS NS ++ NS Income (categories) NS NS NS NS Education +++ NS +++ NS (levels) Age --- +++ --- NS (categories) National NS - NS NS McFadden R2 0,17 0,162 0,105 0,3 N

Significance p <0.001 : +++ or --- p<0,01 ++ or –- p<0,05 : + or - NS = not significant X= excluded

12

The extended model included two additional sets of variables on knowledge of basic statistics (estimate of the public debt, percentage of GDP, unemployment rate, …) and two opinion questions on whether the government is entitled to run a public deficit during recessions and whether public debt is always excessive. Data shows that ignorance of public data is widely spread (less than half of respondents gave the right answer, except for unemployment where more than half gave the right answer) and confidence works as a substitute for some sort of minimum knowledge. The variables did not turn out as significant and no additional explanation emerged. The only variable with a significant (negative) impact on trust in official statistics is the variable on fiscal conservatism (55% of the sample hold the opinion that public debt is always excessive, also it is less than 25% of gdp, the second lowest in the euro area). But no effect on other dependent variables has been detected. It suggests that trust can be influenced by political opinions. Some other techniques have been used to take care of endogeneity. Some of the dependent variables could be exogenous in the other equations of interest in a simultaneous equation framework. To take care of this effect, simultaneous equations models for dichotomous dependent variables, bivariate and trivariate probits6 have been used. For example trust in statistics and in the NSI might have contemporaneous mutual influence asking for a more sophisticated model. But the results (reported in a companion document, available on request with the Stata code) did not show any significant difference to the results reported in table 1.

The 2017 survey

The same regressions have been run on the new data gathered from a survey realized in 2017 with the same questionnaire. Table 2 summarizes the results for the four equations: trust in official statistics, political independence, usage of statistics and trust in the NSI (STATEC). Trust in official statistics is determined by trust in the institute, the political independence; protection of personal data and the usage of statistics as well as media confidence have all a positive and significant effect. Education and age have a significant negative impact. To appreciate the magnitude of the effects, it is interesting to have a look at marginal effects, i.e. the probability of trusting given the individual determinants taken one by one. The most important effect in size is trust in the NSI (STATEC) followed by the level of education, then come political independence and data protection (cf. table in the appendix).

6 SEM and bi and trivariate probit in STATA 13

Table 2. Logit regressions on trust, political independence, usage and trust in NSI (Survey of 2017)

Trust in Survey 2017 Public Political Independence Usage of statistics Trust in NSI (STATEC) Statistics

Trust x +++ ns +++ STATEC +++ +++ +++ x

Independence Political +++ x ns +++

Usage of statistics ++ ns x +++

Protection of data ++ +++ +++ +++

Participant ns + +++ ++ Official survey

Media trust +++ +++ ns +++

Sex (male) ns + +++ ns

Income (categories) ns ns --- ns

Education +++ + + ns (levels)

Age --- + --- + (categories)

National ns ns + ns

McFadden R2 0.229 0.213 0.097 0.384 N=2589

Significance p <0.001 : +++ or --- p<0,01 ++ or –- p<0,05 : + or - NS = not significant X= excluded 14

The additional variables on political attitudes (left to right) and confidence in other fellow citizens have been added to the basic regressions. (In the appendix, the full regressions results are exhibited for trust in statistics). It turns out that political preferences (dummy variable coded 1 for left) have no significant effect while trust in fellow citizens (variable altrui) has a significant and positive effect in all the specifications. Confidence in fellow citizens, a dimension of social capital, increases the probability to trust statistical figures from the NSI.

The results of the surveys, as illustrated in table 1 and 2, carried out in 2017 and 2015, show very similar results.

Conclusions

The two data sets on truth in statistics used in this study are quite rich and provided some useful information, confirming the “felt” political independence of official statistics.

The rate of trust in official statistics reaches circa 69% in 2017, which is quite similar to what its value in 2015 and before (at par with trust in police). We showed that political independence and political attitudes have no bearing on confidence in statistics. But there are still more than 30% of respondents which have only limited confidence in official statistics! What is the optimal level of trust in a given point in time and in a particular social and economic context?

Some methodological considerations are unavoidable .The analysis has to be carried forward to extract all the relevant information. Multiple questions on trust in institutions and fellow citizens should be explored furthermore using . Logit regressions with dichotomous variables should also be completed to take into account a greater variability of the responses (full trust, partial trust, low trust and no trust at all). Ordinal models should be tested as a consequence. The samples which are weighted to increase representativeness should be questioned in order to correct for additional biases arising with online surveys. Finally, the two samples of 2015 and 2017 could be pooled and used for further analysis.

15

References

- Allegrezza, Serge. « La confiance dans les statistiques publiques », 2014, Economie et Statistiques, April 2014, http://www.statistiques.public.lu/catalogue-publications/economie-statistiques/2014/74-2014.pdf. - Bachkirova, Tatiana. « A New Perspective on Self-Deception for Applied Purposes ». New Ideas in Psychology 43 (décembre 2016): 1-9. https://doi.org/10.1016/j.newideapsych.2016.02.004. - Bénabou, Roland, et Jean Tirole. « Mindful Economics: The Production, Consumption, and Value of Beliefs ». Journal of Economic Perspectives 30, no 3 (août 2016): 141-64. https://doi.org/10.1257/jep.30.3.141. - Chiche, Jean, et Flora Chanvril-Ligneel. « Confiance dans les statistiques publiques : une relation contrariée », 2016, 25. - Sara E. Gorman, Jack M. Gorman « Denying to the Grave: Why We Ignore the Facts That Will Save Us » - European Commission. Eurobarometer. Science and Technology. no 340. July 2010, http://ec.europa.eu/commfrontoffice/publicopinion/archives/ebs/ebs_340_en.pdf. - European Commission. Europeans and Economic statistics. Report. Standard Eurobarometer 83, spring 2015. http://ec.europa.eu/commfrontoffice/publicopinion/archives/eb/eb83/eb83_stat_en.pdf. - Desrosières, Alain. La politique des grands nombres : Histoire de la raison statistique. Editions La Découverte, 2010. - Dieguez, Sebastian. Total bullshit ! : Aux sources de la post-vérité. Paris: Presses Universitaires de France - PUF, 2018. - Grigorieff, Alexis, Christopher Roth, et Diego Ubfal. « Does Information Change Attitudes Towards Immigrants? Representative Evidence from Survey Experiments ». SSRN Electronic Journal, 2018. https://doi.org/10.2139/ssrn.2768187. - Gorman, Sara E., et Jack M. Gorman. Denying to the Grave: Why We Ignore the Facts that Will Save Us. Oxford ; New York: OUP USA, 2016. - Haidt, Jonathan (2013) The Righteous Mind: Why Good People Are Divided by Politics and Religion (Vintage) Paperback, - Lewandowsky, Stephan, Ullrich K.H. Ecker, et John Cook. « Beyond Misinformation: Understanding and Coping with the “Post-Truth” Era ». Journal of Applied Research in Memory and Cognition 6, no 4 (décembre 2017): 353-69. https://doi.org/10.1016/j.jarmac.2017.07.008. - Nyhan, Brendan, et Jason Reifler. « When Corrections Fail: The Persistence of Political Misperceptions ». Political Behavior 32, no 2 (1 juin 2010): 303-30. https://doi.org/10.1007/s11109-010-9112-2.

16

- OECD. « G20/OECD INFE report on adult financial literacy in G20 countries - OECD », 2017. http://www.oecd.org/finance/g20-oecd-infe-report-adult-financial-literacy-in-g20- countries.htm. - OECD Skills Matter: Further Results from the Survey of Adult Skills. OECD Skills Studies. Paris: OECD, 2016. - Pinker, Steven. Enlightenment Now: The Case for Reason, Science, Humanism, and Progress. New York, New York: Viking, 2018. - Porter, Theodore M. Trust in Numbers – The Pursuit of Objectivity in Science and Public Life. Reprint. Princeton, N.J: Princeton University Press, 1996. - « Rational ignorance versus rational irrationality ». Consulté le 3 février 2017. http://econfaculty.gmu.edu/bcaplan/pdfs/rationalignorancevs.pdf. - Rey, Olivier. Quand le monde s’est fait nombre. Paris: Stock, 2016. - Sloman, Steven, et Philip Fernbach. The Knowledge Illusion: The Myth of Individual Thought and the Power of Collective Wisdom. Macmillan, 2017. - Dan Sperber, Hugo Mercier: « The Enigma of Reason: A New Theory of Human Understanding eBook: Amazon.de: Kindle-Shop ». Consulté le 16 août 2018. - Drew Westen « The Political Brain: The Role Emotion in Deciding the Fate of the Nation: How We Make Up Our Minds Without Using Our Heads. - « The Righteous Mind: Why Good People are Divided by Politics and Religion: Amazon.de: Jonathan Haidt« Trust ». Our World in Data. Consulté le 7 août 2018. - Wood, Thomas, et Ethan Porter. « The Elusive Backfire Effect: Mass Attitudes’ Steadfast Factual Adherence ». SSRN Scholarly Paper. Rochester, NY: Social Science Research Network, 31 décembre 2017. https://papers.ssrn.com/abstract=2819073.

Appendix

The appendix contains output of a Logit regression with trust in statistics (conf) as a dependent variable. Three specifications have been tested. The first with all the exogenous variables : participation in surveys (particip), data protection (protection) trust in STATEC (statec), gender-man (sex), level of educational attainment (EDU2-EDU8), income of the household (REV2-REV13), age (2,Agen-5.Agen). The second includes all the former exogenous variables and adds polit (1=left) and trust in fellow citizens (altrui). The third retains only significant variables. The second table shows the marginal effects of the last Logit regression in order to capture the magnitude of the coefficients on probability of Trust. The last table shows all the variables used and their .

17

(1) (2) (3) conf conf conf conf util 0.217 0.312** (1.91) (2.82) particip -0.115 -0.0358 (-0.98) (-0.31) protection 0.543*** 0.891*** 0.543*** (3.33) (6.08) (3.43) statec 1.551*** 1.556*** (9.99) (9.80) media 0.627*** 0.667*** 0.504*** (5.37) (5.79) (4.35) indep 1.024*** 1.315*** 0.978*** (8.44) (11.55) (8.01) sex 0.0138 -0.0604 (0.12) (-0.54) lux -0.219 -0.206 (-1.72) (-1.68)

EDU2 0.103 0.154 (0.31) (0.50)

EDU3 0.609 0.580 0.525*** (1.85) (1.89) (3.50)

EDU4 0.897** 0.982** 0.856*** (2.62) (3.07) (5.13)

EDU5 1.111** 1.109** 1.064*** (3.06) (3.25) (5.09)

EDU6 1.425*** 1.470*** 1.415*** (3.96) (4.32) (7.11)

EDU8 0.384 0.293 (0.84) (0.74)

REV2 -0.931 -0.840 (-1.32) (-1.17)

REV3 -0.338 -0.237 (-0.44) (-0.30)

REV4 0.200 0.136 (0.28) (0.19)

REV5 0.320 0.403 (0.56) (0.65)

REV6 0.305 0.383 (0.55) (0.63)

REV7 0.314 0.283 (0.49) (0.42)

REV8 0.131 0.0945 (0.24) (0.16)

REV9 0.303 0.376 (0.58) (0.65)

REV10 0.221 0.227 (0.43) (0.39)

REV11 0.556 0.661 (1.06) (1.14)

REV12 0.714 0.737 (1.35) (1.27)

REV13 0.275 0.257 (0.54) (0.45)

1bn.Agen . . . .

2.Agen -0.379 -0.343 (-1.57) (-1.54)

3.Agen -0.850*** -0.769*** (-3.66) (-3.63)

4.Agen -0.875*** -0.807*** (-3.71) (-3.74)

5.Agen -0.632* -0.527* (-2.46) (-2.18) polit -0.119 (-1.03) altrui 0.572*** 0.553*** (4.43) (4.29)

_IAgen_3 -0.552*** (-3.56)

_IAgen_4 -0.615*** (-3.81)

_IAgen_5 -0.448* (-2.29)

_cons -2.134*** -1.596** -2.067*** (-3.79) (-2.60) (-9.24)

N 2589 2589 2589 pseudo R-sq 0.229 0.197 0.226 t statistics in parentheses * p<0.05, ** p<0.01, *** p<0.001 18

Table A2 : Marginal effects of last logit equation specification Marginal effects after logistic y = Pr(conf) (predict) = .72635701

variable dy/dx Std. Err. z P>|z| [ 95% C.I. ] X

statec* .3526753 .0372 9.48 0.000 .279774 .425577 .819163 protec~n* .1160421 .03594 3.23 0.001 .045608 .186476 .841523 media* .098771 .02215 4.46 0.000 .055365 .142177 .444774 indep* .203415 .02569 7.92 0.000 .153069 .253761 .63413 altrui* .1044843 .023 4.54 0.000 .05941 .149558 .315594 EDU3* .0988821 .02665 3.71 0.000 .046652 .151112 .289247 EDU4* .1513625 .0258 5.87 0.000 .100792 .201933 .228878 EDU5* .1699831 .02544 6.68 0.000 .120117 .219849 .109624 EDU6* .2142368 .02176 9.85 0.000 .171589 .256885 .13963 _IAgen_3* -.1152123 .03312 -3.48 0.001 -.180124 -.050301 .276014 _IAgen_4* -.1304062 .0354 -3.68 0.000 -.199792 -.06102 .231111 _IAgen_5* -.0946041 .04323 -2.19 0.029 -.17933 -.009878 .165815

(*) dy/dx is for discrete change of dummy variable from 0 to 1

Table A3 . List of variables

Variable Obs Mean Std. Dev. Min Max

particip 2,589 .4028582 .4905675 0 1 protection 2,589 .8455002 .361497 0 1 media 2,589 .4499807 .4975879 0 1 conf 2,589 .6740054 .4688358 0 1 polit 2,589 .6635767 .4725769 0 1

indep 2,589 .634608 .4816329 0 1 statec 2,589 .8215527 .3829628 0 1 sex 2,589 .4963306 .5000831 0 1 EDU1 2,589 .0239475 .1529151 0 1 EDU2 2,589 .1958285 .3969138 0 1

EDU3 2,589 .2900734 .4538837 0 1 EDU4 2,589 .2220935 .415734 0 1 EDU5 2,589 .1096949 .3125694 0 1 EDU6 2,589 .1367323 .3436309 0 1 EDU8 2,589 .02163 .1455002 0 1

lux 2,589 .6828891 .4654409 0 1 REV1 2,589 .0057937 .0759105 0 1 REV2 2,589 .0084975 .0918071 0 1 REV3 2,589 .00618 .0783848 0 1 REV4 2,589 .0150637 .12183 0 1

REV5 2,589 .03399 .1812383 0 1 REV6 2,589 .0336037 .1802416 0 1 REV7 2,589 .0417149 .1999756 0 1 REV8 2,589 .0652762 .2470602 0 1 REV9 2,589 .1154886 .3196724 0 1

REV10 2,589 .1096949 .3125694 0 1 REV11 2,589 .1490923 .3562483 0 1 REV12 2,589 .1819235 .385856 0 1 altrui 2,589 .3151796 .4646771 0 1 .

19