Changes in voter behavior after an information signal: An experimental approach for Senegal

Henning, Christian H. C. A.; Petri, Svetlana; Diaz, Daniel Working Paper Changes in voter behavior after an information signal: An experimental approach for Senegal Working Papers of Agricultural Policy, No. WP2020-11 Provided in Cooperation with: Chair of Agricultural Policy, Department of Agricultural Economics, University of Kiel Suggested Citation: Henning, Christian H. C. A.; Petri, Svetlana; Diaz, Daniel (2020) : Changes in voter behavior after an information signal: An experimental approach for Senegal, Working Papers of Agricultural Policy, No. WP2020-11, Kiel University, Department of Agricultural Economics, Chair of Agricultural Policy, Kiel This Version is available at: https://hdl.handle.net/10419/235902 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by-nc/4.0/ The Agricultural Working Paper Series is published by the Chair of Agricultural Policy at the University of Kiel. The authors take the full responsibility for the content. Christian Henning Svetlana Petri Daniel Diaz Department of Agricultural Economics University of Kiel Changes in Voter Behavior after an Information Signal: An Experimental Approach for Senegal AGRICULTURAL POLICY WORKING PAPER SERIES WP2020-11 ISSN: 2366-7109 WORKING PAPERS OF AGRICULTURAL POLICY Christian Henning Svetlana Petri Daniel Diaz Changes in Voter Behavior after an Information Signal: An Experimental Approach for Senegal Department of Agricultural Economics University of Kiel Kiel, December2020 WP2020-11 http://www.agrarpol.uni-kiel.de/de/publikationen/working-papers-of-agricultural-policy About the authors: Prof. Christian Henning has the professorship of the Chair of Agricultural Policy at the Institute of Agricultural Economics in the University of Kiel (Germany). Before he joined the faculty in Kiel he did research at the University Mannheim, Stanford University and at the "Mannheimer Zentrum für Europäische Sozialforschung". He also operated as a consultant for the FAO, the European Union and further international and national organizations. Dr. agr. Svetlana Petri is a post doctoral researcher at the Department of Agricultural Policy at the Agricultural Economics Institute in the University of Kiel (Germany). Among her research interests are: Election Models of local and central Governments, Decentralisation of Infrastructure Delivery in Developing Countries, Determination of Capture and Accountibility Factors in Developing Countries, and Influence of Elite-Networks on the political decision Process in Developing Countries Daniel Diaz is doctoral student in the Department of Agricultural Policy at the Agricultural Economics Institute in the University of Kiel (Germany). He studied business administration and he acquired an MSc in international business economics. His main research interests are probabilistic voter models and modelling capture and accountability for central Governments. Corresponding author: c[email protected], sp[email protected], [email protected] 3 Abstract Electoral competition is considered a control mechanism to guarantee a good performance of the government. However, in real life it often leads to a distorted policy implementation due to Government Capture and low Government Accountability. Therefore, the analysis of voter behavior is a key factor to understand government performance. More specifically, if voters choose more policy and retrospectively oriented, the government has greater incentives to implement efficient policies. In this sense, if voters have more information on politics, they are more likely to base their decision on policy issues. To assess changes in voter behavior, we carried out a political experiment, where information about the performance of the Senegalese government was delivered to a randomly selected group of voters. Then, based on election surveys data collected before and after the information signal, a probabilistic voter model with latent class using a panel data set was developed. Additionally, to evaluate changes in the relative importance of the three voting motives (policy, non-policy and retrospective), marginal effects and relative marginal effects were estimated. As expected, after the information signal, the relative importance of the policy and the retrospective components increased significantly. Furthermore, to see the impact on government performance, indicators for Capture and Accountability were calculated. Even though, the accountability index in Senegal is low, after the delivery of the political information, we observed an increment. Likewise, some capture indices changed from the first to the second round. Finally, we measured the changes in the optimal policy position of the incumbent party. Since there were changes in the policy positions of voters, if the incumbent wants to maximize its probability of winning the elections, it should make changes on its policy platform correspondingly. Keywords: probabilistic voter model, capture, accountability, agricultural policy, Senegal, Africa JEL classification:Q18, C33, C35, C38 4 1 Introduction In political theory, electoral competition is considered a control mechanism, as voters have the power to either punish the bad performance of the government or reward the good one through their vote. However, in real life, electoral processes often lead to a distorted policy implementation due to Government Capture and low Government Accountability. Therefore, the analysis of voter behavior is a key factor to understand government performance. One of the most influential authors concerning public choice theory [Downs, 1957], states that voters evaluate candidates based on their policy platfoms, as well as, on an estimation of what such candidates would do were they in power (i.e. policy oriented). On the other hand, Grossman and Helpman [1996] affirm that voters have both, policy oriented and non-policy oriented voting motives. From their perspective, the relative importance of these voting motives depends on the level of information that voters have about politics. To elaborate further, if voters have limited information on politics by the time of casting a vote, they are likely to base their decision on non-policy issues, such as charisma or religion, giving less consideration to government policy positions. This behavior in turn, reduces the incentives for the government to implement efficient policies. Other authors also highlight the importance of information when it comes to make electoral choices. For instance, DellaVigna and Kaplan [2006] studied the impact of media bias upon voting, finding a significant effect of exposure to news on voting decision. Such exposure induced a substantial percentage of the viewers to change their decision. Pande [2011] in turn, explained that limited information is an explanation for low-quality politicians in low-income democracies. Therefore, information about the political process and politician performance improves electoral accountability. According to Coate [2004], voters update their beliefs rationally given the information they have received from advertising campaigns. Furthermore, Banerjee et al. [2011] found evidence that voters change quite substantially their electoral choice when they are given information about government performance and qualifications of the incumbent. They pointed out the fact that voters demonstrated sophistication using the information to judge performance and qualifications, as oppose to the fear that information would simply confuse them. According to Khemani [2001], a large number of voters are motivated by party affiliation and other non-policy variables, while others that are indifferent between candidates or parties on ideological grounds, vote based on economic information (macro economic variables such as economic growth, inflation, poverty and income inequality). He interpreted voter responses to economic performance in Indian elections and argues that the evidence is consistent with greater voter vigilance and government accountability at local level. As pointed by Caplan [2007], in practice, democracies frequently adopt policies 5 that are damaging. This is partly due to the fact that voters embrace a long list of misconceptions that lead them to act irrationally and vote accordingly. Bardhan and Mookherjee [2000] distinguish between informed and uninformed voters. Informed voters are politically aware and choose based on the utility they expect to obtain. On the contrary, uninformed voters are swayed by campaign spending. In this sense, policy biases (Government Capture) emerge due to the existence of uninformed voters. Additionally, parties choose their policy platforms in order to maximize their probability of winning the elections. In this regard, we conducted a political experiment where an information signal is delivered to different groups of voters. Then, based on voter survey data collected before and after this information signal, changes in voter behavior were measured by means of a probabilistic voter model. We were also interested in measuring the changes in the relative importance of the three voting motives (policy, retrospective and non-policy), as well as, in the government performance indicators. Finally, based on the first and the second order conditions, we measured the changes in the optimal policy position of the incumbent party BBY for three policy issues. 2 Experimental Study The experimental study took place in Senegal few weeks before the presidential election of February 24th, 2019. Planning the implementation of the experiment and surveys according to upcoming elections is crucial for voter behavior analysis. In this sense, we assumed that the Senegalese electorate had made up their mind for the election, which provides reliable data regarding the actual voting decision. Additionally, all political parties had chosen their policy platforms and candidates. The experiment was carried out in five regions of Senegal. It was a random experiment as individuals were randomly assigned to different groups. It consisted of a first round of 1000 interviews, conducted face-to-face in the corresponding dialect/language (Serere, Wolof, Pular and French). The next step, was the delivery of an information signal. To this end, the total sample of interviewees was divided into three groups: group 1 received a positive treatment, group 2 received a negative treatment and group 3 received a placebo treatment. After receiving the signal, a second round of interviews was conducted containing just some of the questions from the first round. 2.1 Information Signal The tool implemented as information signal was a series of videos comprised of two parts. The first part, contained information about the role of the government and the power held by voters either to reward or punish their performance. The second part, showed 6 the performance of the government regarding the implementation of agricultural policies in the framework of the Malabo Declaration. The aforementioned declaration is a re-commitment to the goals of the Comprehensive Africa Agriculture Development Programme (CAADP) agreed by the Heads of State and Government of the African Union to provide effective leadership for the attainment of specific goals by the year 2025. The goals include ending hunger, tripling intra-African trade in agricultural goods and services, enhancing resilience of livelihoods and production systems, and ensuring that agriculture contributes significantly to poverty reduction. One of the Malabo strategic objectives is the Commitment to Mutual Accountability to Actions and Results, which includes a Biennial Agricultural Review Process that involves tracking, monitoring and reporting the progress. These results are then presented in individual scorecards for each country, where the performance indicators are shown. These scorecards were obtained from the Regional Strategic Analysis and Knowledge Support System (ReSAKSS) [African Union, 2018] and an example of one of them can be found in the appendix (figure 5). With the information contained in these scorecards, we decided to design a traffic light rating system. This system has the advantage of being universally recognized, and using the three colors of the real traffic lights (green, yellow and red), good and poor performance can easily be identified. The table showing the scores and thresholds from which the performance is defined to be either good or bad is also available in the appendix (figure 6). Based on these results, we proceeded to select three good and three bad indicators of the Senegalese performance taking into account the scores of the neighbor countries. The indicators used to deliver the signal were: Positive indicators: •Public expenditures in agriculture •Strengthening social protection •Tripling Intra-African Trade for agriculture commodities and services Negative indicators: •Ensuring resilience to climate related risks •Establishing CAADP based cooperation, partnership and alliance •Establishing Intra-African policies and institutional conditions The information signal was exhibited as a map displaying the selected indicators of the Senegalese performance compared to the same indicators for most of the neighbor 7 ECOWAP countries. Examples of the maps are presented in the appendix (figures 7 and 8). As regards the placebo video, it was not related to the agricultural policy implementation, but instead it was a short documentary about the process of desertification in Senegal. In the video, rural communities receive training on planting patterns to enrich the soil and stop the desertification process. Unlike the other videos, the idea was not to change the opinion of the audience about the performance of the government in the agricultural sector [Elsen, 2016]. 3 Methodology 3.1 Voter Behavior To analyze voter behavior we estimated a probabilistic voter model that makes possible the inclusion, in the utility function, of a stochastic term containing all unknown factors. These models are usually estimated with Discrete Choice models, as they can explain choices between two or more alternatives. In this study, the alternative Abstention was also included in the choice set. To derive the Discrete Choice model, it is common to apply a Random Utility Maximization (RUM) Model. Here, if voter idecides to participate in the election, he chooses party konly if this party provides him the highest utility Vik. Similarly, if the voter chooses not to participate, the greater utility comes from the alternative Abstention. Also, we assume that the stochastic term is independently, identically extreme value distributed (iid) and thus a logit model was derived. This model was extended to a multi-alternative estimation based on McFadden [1974], meaning that voters can choose an alternative k from a set of alternatives K. Pik(K) = eVik K P k=1 eVik (1) Given the nature of the experiment, we created two datasets. Both of them contain all variables that did not change from round 1 to round 2, such as, the socio-demographic characteristics. Additionally, each dataset includes those variables built from questions asked in both rounds, like, distances, choice, satisfaction with president and satisfaction with policy. Then we combined the datasets by rows to create the wide panel data. Finally, in order to perform the estimations, the latter was transformed into a long panel data. An example is displayed in Table 1. 8 4 Data We designed two rounds of voter surveys including questions on socio-demographic characteristics, voting behavior, policy positions and network characteristics. The first and second round of interviews were applied in Senegal on the same day in January 2019 by the Senegalese Agricultural Research Institute. These were conducted face-to-face in the respective dialect or language of the interviewees. The sample contains 1000 individuals from five different regions across the country. After data cleaning, 844 complete observations remained for the analysis of voters’ behavior. The observations with missing values in the variables choice-pre and choice-post were eliminated. 4.1 Dependent Variable In a probabilistic voter model the dependent variable is usually the actual or intended vote choice. Given the approach of this paper, the interviewees had to answer, in both rounds, to the following question: If a presidential election were held tomorrow, which party would you vote for? To include the alternative Abstention, we followed the approach of Thurner and Eymann [2000]. They explain that the number of people who decide not to participate in an election is usually underestimated in surveys due to effects of social (un)desirability. Therefore, we have considered the interviewees who revealed their intention of abstaining, as well as, the potential non-voters. In other words, we have taken into account those respondents who answered “Will not vote" and “Don’t know" as part of the Abstention alternative. Table 2 shows the results of both surveys, as well as, the official presidential election outcome. Even though none of the surveys’ results are close to the actual election outcome, the party in power BBY is a clear winner. For the analysis in the empirical section we consider all parties and Abstention. Then, the whole set of alternatives is: K = {BBY, Rewmi, Pastef, PUR, Niang and Abstention}. Table 2: Senegalese presidential election results BBY Rewmi Pastef PUR Niang Abstention Presidential election 2019 38.48% 13.55% 10.35% 2.69% 0.98% 33.95% Own survey 2019 (First round) 70.46% 3.72% 5.30% 1.13% 0.34% 19.05% Own survey 2019 (Second round) 73.53% 3.96% 5.32% 1.02% 0.34% 15.84% Source: [Constitutional Council of Senegal, 2019], own survey 15 4.2 Independent Variables To explain the dependent variable, only independent variables with less than 10% of missing values were considered. Furthermore, the variables with missing values were imputed with the mean value, except for the policy positions that were imputed via linear regressions. For this study, they were divided into policy, retrospective and non-policy. Policy Variables: Respondents were asked about their policy positions and their perceived policy positions of the parties on nine different issues. The positions were asked, based on a five-point scale, on the following issues: 1. Social 2. Ideology 3. Investment in: Public services vs. Economic growth (PSvsEG) 4. Investment in: Education and health services vs. Insecurity and violence reduction (EHvsIV) 5. Development of: Agricultural sector vs. Industrial sector (AGRvsIND) 6. Increase productivity of: Food crops vs. Cash crops (FoodvsCash) 7. Benefit the agricultural sector through: Technological progress vs. Access to markets (TPvsAM) 8. Agricultural sector should be: Taxed vs. Protected (TaxvsProtect) 9. Accountability Then, distances for the parties were calculated as the difference between the voters’ own policy position and the perceived policy position of the parties. In the case of the alternative Abstention, the distance was set to 0. Therefore, the utility of non-voting is greater than the utility of voting and hence the voting paradox is fulfilled. Retrospective Variables: Questions of satisfaction with government performance were asked in both rounds. More specifically, there were questions addressing the level of satisfaction with the performance of the current president, as well as, the implementation of agricultural policies by the government. Non-policy Variables: A whole set of sociodemographic variables such as gender, rurality, marital status and education was included, as well as, other variables measuring the level of trust of voters on different types of institutions. Moreover, regions and ethnic 16 groups were coded as dummy variables. For the specific characteristics of the candidates, we performed a factor analysis to reduce the number of variables resulting in a two-factors solution that we called factor “Image" and factor “Origin". Furthermore, to measure party loyalty, the variables Party ID were created as alternative specific dummies, where “1" indicates party affiliation for that specific party and “0" otherwise. In the case of the alternative Abstention, the variable was set to “0" since there is no such thing as party identification for Abstention. Based on Mattes [2008], a Lived Poverty Index (LPI) was estimated, where the level of poverty is high if it is closer to 5 and low if it is closer to 0. Additionally, we created the dummy variables “treated”, “positive”, “negative” and “placebo” regarding the type of treatment. Finally, a political Knowledge Index (PKI) was designed based on the answers of the voters to a number of exogenous questions about political knowledge. Then, we created a dummy variable that defines if the voter is informed or uninformed. 5 Empirical Application and Results 5.1 Latent Class Model With the data described in the former section, we estimated a probabilistic voter model with latent class using a panel data set to determine which factors influence and change voting behavior in Senegal. This latent class model (LCM) approach takes into account the heterogeneity of the data, which is relevant because voting motives differ across voters. The estimated LCM consists of two sub-models, the model for choices that determines which alternative is chosen and the model for classes that defines class membership. In the former, all variables that changed between rounds and the variable Party Identification were included, whereas in the latter, all variables that did not change between rounds were considered. As it was mentioned in the methodology section, the classes were fixed to see the changes between rounds, that is, classes 1 and 2 correspond to the first round and classes 3 and 4 correspond to the second round. Different model specifications were estimated and the goodness of fit was measured with the Akaike Information Criterion (AIC). The models included only the significant independent variables chosen via the z-score test. In this paper we show the results of the two best models which are displayed in tables 3 and 4. However, for simplicity, present the entire analysis only for Model 1. 17 Table 3: Model 1 - Latent Class Model with Panel Data AIC = 4274.7902 Class 1 (0.4423) Class 2 (0.0577) Class 3 (0.4566) Class 4 (0.0434) VARIABLES Coeff. z-value Coeff. z-value Coeff. z-value Coeff. z-value Model for Choices Attributes Abstention:(intercept) 3.3521 4.2739 *** 0.3105 0.2166 2.2568 4.3481 *** -3.2386 -0.1268 BBY:(intercept) 1.7450 2.3241 * -8.7911 -1.6717 . 0.8109 1.6174 0.5391 0.0210 Niang:(intercept) -1.5231 -0.6657 -1.6990 -0.4296 -1.3392 -0.8646 -11.0344 -0.0867 Pastef:(intercept) -1.0477 -0.8765 5.7851 2.4529 * 0.1731 0.2170 3.7602 0.1471 PUR:(intercept) -0.1366 -0.1128 0.3934 0.1443 -1.8141 -1.6230 -5.2329 -0.1974 Rewmi:(intercept) -2.3897 -2.0041 * 4.0011 1.9985 * -0.0875 -0.0916 15.2065 0.5781 PSvsEG -0.0265 -0.8357 -0.3246 -1.7913 . 0.0377 1.2372 -1.6306 -1.8658 . FoodvsCash -0.0561 -1.5531 -0.5460 -2.1038 * -0.0633 -2.0617 * -1.6810 -1.8981 . Ideology 0.0282 0.6458 -0.4729 -2.3786 * -0.0013 -0.0352 -0.9580 -1.7646 . Party_ID 6.6453 6.9021 *** 7.8855 2.4386 * 6.2240 10.0347 *** -3.7839 -1.3712 Predictors Abstention:Satisfaction_president -0.3949 -1.2144 0.6396 1.4274 -0.1153 -0.5486 1.3419 0.1617 BBY:Satisfaction_president 0.6818 2.2169 * 1.3344 1.1531 0.8856 4.3668 *** 1.2000 0.1440 Niang:Satisfaction_president -0.3822 -0.2880 0.6318 0.5114 -0.3665 -0.5196 1.6750 0.0405 Pastef:Satisfaction_president 0.0348 0.0729 -1.5185 -1.9046 . -0.2835 -0.8159 -0.8992 -0.1079 PUR:Satisfaction_president -0.1543 -0.3252 -0.2750 -0.2526 0.2982 0.7393 1.6735 0.1949 Rewmi:Satisfaction_president 0.2148 0.4552 -0.8123 -1.1693 -0.4185 -0.9913 -4.9913 -0.5801 Model for Classes Covariates classes:intercept 0.6739 1.7509 . 0.2227 0.3613 0.9482 2.4698 * -1.8448 -2.3120 * classes:Trust_president 0.3207 4.3349 *** -0.3851 -3.0394 * 0.2738 4.0480 *** -0.2095 -1.7308 . classes:LPI -0.1791 -1.8370 . -0.0204 -0.1275 -0.2119 -2.2062 * 0.4114 2.0849 * classes:factor_char_image -0.3110 -3.9941 *** 0.2189 1.6799 . -0.3114 -3.8461 *** 0.4035 2.5727 * ***p<0.001, **p<0.01, *p<0.05, . p<0.10 Source: Own estimation 18 Table 4: Model 2 - Latent Class Model with Panel Data AIC = 4281.6220 Class 1 (0.4389) Class 2 (0.0611) Class 3 (0.4518) Class 4 (0.0482) VARIABLES Coeff. z-value Coeff. z-value Coeff. z-value Coeff. z-value Model for Choices Attributes Abstention:(intercept) 3.1506 4.4931 *** -0.8215 -0.2234 1.9732 3.8387 *** -1.7868 -0.0664 BBY:(intercept) 1.5194 2.4380 * -9.5455 -1.6580 . 0.5900 1.2100 2.3040 0.0853 Niang:(intercept) -1.9535 -1.1498 -1.5656 -0.0916 -1.5425 -0.9868 -9.6910 -0.0721 Pastef:(intercept) -0.7120 -0.6644 6.4115 1.5359 0.4050 0.5272 4.6385 0.1721 PUR:(intercept) -0.2448 -0.2228 1.0923 0.2359 -1.9998 -1.7592 . -3.4059 -0.1249 Rewmi:(intercept) -1.7597 -1.5702 4.4288 1.1490 0.5741 0.7060 7.9412 0.2935 PSvsEG -0.0282 -0.8586 -0.4604 -2.6375 ** 0.0401 1.2800 -2.7491 -2.2451 * FoodvsCash -0.0616 -1.6388 -0.5841 -2.2111 * -0.0593 -1.8338 . -0.8361 -2.1035 * TPvsAM 0.0127 0.4139 -0.5145 -2.4933 * -0.0035 -0.1211 -0.2740 -1.6791 . Ideology 0.0290 0.6468 -0.5119 -2.5895 ** 0.0012 0.0301 -0.7011 -1.6273 Party_ID 6.3380 8.2442 *** 9.3519 3.4598 *** 6.3038 9.2443 *** -4.5043 -1.5453 Predictors Abstention:Satisfaction_president -0.4281 -1.7649 . 1.2401 0.5723 -0.0225 -0.1058 0.6408 0.0730 BBY:Satisfaction_president 0.6897 3.3489 *** 1.5614 0.6594 0.9728 4.8071 *** 0.6727 0.0764 Niang:Satisfaction_president -0.0108 -0.0169 -1.0038 -0.0949 -0.2883 -0.4050 1.0083 0.0230 Pastef:Satisfaction_president -0.0716 -0.1856 -1.2438 -0.5518 -0.3491 -1.0353 -1.4409 -0.1637 PUR:Satisfaction_president -0.1797 -0.4459 -0.0604 -0.0245 0.3440 0.8258 1.2630 0.1422 Rewmi:Satisfaction_president 0.0005 0.0013 -0.4935 -0.2245 -0.6570 -1.6206 -2.1440 -0.2426 Model for Classes Covariates classes:intercept 0.1325 0.6347 0.1089 0.3500 0.3228 1.7660 . -0.5641 -1.7990 . classes:Trust_president 0.2993 4.5215 *** -0.3560 -3.1316 ** 0.2503 4.1650 *** -0.1936 -1.8043 . ***p<0.001, **p<0.01, *p<0.05, . p<0.10 Source: Own estimation 19 Among the attributes in the model 1, are the alternative specific constants, that absorb all information not explicitly incorporated in the model. Also included, are the policy issues PSvsEG, FoodvsCash and Ideology, which are significant with negative coefficients for at least one class. This means that the greater the distance between the voter’s position and the perceived position of the party, the less is the utility and the probability to choose that party. In the case of the alternative Abstention, the utility of non-voting is always greater than the utility of voting, which is consistent with the voting paradox. The last attribute was Party Identification, that turned out to be significant for three classes with positive coefficients. This means that if a voter feels close to a political party, the probability that he will chooses the corresponding candidate increases. On the other hand, those voters that are not close to any political party, do not increase their utility by casting a vote for any candidate, hence, they rather abstain. As regards the predictors, the variable Satisfaction with President is significant for classes 1 and 3 with positive coefficients for the alternative BBY which is the incumbent party. This is consistent with the theory, as the greater the satisfaction with the president, the greater the probability to support the incumbent in the elections. Likewise, for class 2 the variable is significant for Pastef with negative coefficient, meaning that the greater the satisfaction with the president, the less the probability to choose this alternative. Concerning the covariates, the intercepts for classes 1 and 3 are significant with positive coefficients which reflects the existence of bias towards belonging to class 1 in the first round and class 3 for the second round. Similarly, the probability to belong to class 1 in the first round and class 3 in the second round increases when voters trust the president, when their LPI is not high and when the image of the candidate is not relevant to them. The size of the class memberships is approximately as follows: Class 1: 44.23 (%) , Class 2: 5.77 (%) , Class 3: 45.66 (%) and Class 4: 4.34 (%). This evidences a weak heterogeneity. With respect to the attributes and predictors in Model 2, the results are very similar. Nevertheless, it is worth noting that the policy issue TPvsAM also resulted significant. Furthermore, the variable Satisfaction with President turned out to be significant with negative coefficient for the alternative Abstention for class 1, which indicates that the greater the satisfaction with the president, the less likely is that the voter will abstain. Regarding the covariates, the variable Trust President was the only significant variable, which suggests a higher probability to belong to class 1 for the first round and class 3 for the second round when voters trust the president. Finally, for model 1 we estimated the utilities and probabilities. Table 5 shows the mean probability for each alternative and round. As expected, there was a change from the first round to the second round. Notwithstanding, it is clear that in both rounds the incumbent party BBY had the highest probability of winning the elections. 20 Table 5: Mean Probabilities Alternatives Round 1 Round 2 Abstention 18.07% 16.44% BBY 71.72% 73.71% Niang 0.36% 0.34% Pastef 5.22% 4.28% PUR 1.05% 0.97% Rewmi 3.58% 4.26% Source: Own estimation 5.2 Government Performance Indicators As mentioned in the methodology section, the probabilistic voter model is a logistic regression model. Therefore, its coefficients only allow to measure the direction of the impact, but to evaluate the magnitude of such impact, marginal effects (ME) had to be calculated. In the case of the LCM, these can be calculated only for the variables included in the model for choices. In addition to ME, the relative marginal effects (RME) for each voting motive and each round were calculated. As displayed in figure 1 the RME of the non-policy component is the highest in both rounds. On the contrary, the policy voting motive is the less relevant. However, it is important to notice that after the information signal, the RME of the nonpolicy component decreased from 84% in round 1 to 76,84% in round 2. On the other hand, the RME of the policy voting motive increased from 4,33% to 10,74%. Finally, the RME of the retrospective voting motive had a slight increment as it went from 11,67% to 12,42%. As expected, the information signal changed the voting behavior as voters chose more policy and retrospectively oriented in the second round. Even when we look at the treatments separately, the importance of the non-policy component decreased, whereas, the importance of the policy and retrospective components increased significantly. The previous analysis applies even to the recipients of the placebo signal, which lead us to think that the information contained in that video did not fulfill its mission to keep unchanged the intended vote choice of the interviewees. 21 Figure 1: Relative Maginal Effects (a) Round 1 (b) Round 2 Source: Own estimation Additionally, we analyzed the impact of the information signal on informed and uninformed voters, where similar results were observed. Table 6 shows that the importance of the non-policy voting motive is higher for the uninformed group. Also, the information signal had a higher impact on the informed group, than it had on the uninformed group. 22 Table 6: RME Informed vs. Uninformed Round 1 Round 2 p-value Non-Policy Informed 83.76% 75.28% <2.2e-16 Uninformed 84.37% 79.19% 0.0000 Policy Informed 4.91% 12.55% 0.0000 Uninformed 3.45% 8.01% 0.0000 Retro Informed 11.33% 12.17% 0.0017 Uninformed 12.18% 12.79% 0.0140 Source: Own estimation 5.2.1 Government Accountability Governments act accountable when they implement policies serving the needs and desires of voters rather than favoring special interest of lobbying groups or intrinsic policy preferences of politicians. Based on the estimated model, the accountability indices were calculated for each round and also for each type of treatment as can be seen on table 7. The low index of accountability for all cases suggests that the function of elections of holding accountable the government is not fulfilled. However, it is important to highlight that the index increased for the entire sample, as well as, for each treatment group after the delivery of the information signal. Also relevant, is the fact that the biggest change was the one experienced by the group that received the positive treatment. Table 7: Accountability indices Round 1 Round 2 Whole sample 16.00% 23.16% Positive 16.17% 23.69% Negative 15.94% 22.75% Placebo 15.89% 23.06% Source: Own estimation 5.2.2 Government Capture Electoral competition can be biased in favor of special interests. To measure the political weight of certain groups of voters, government capture indices were calculated. Thus, we 23 could identify which groups were being favored at the expense of others. In table 8 is evident that for round 1, the rural population captures the urban, men capture women, young people capture the old, uneducated people capture the educated, married people capture people with other marital status, the muslim religion captures other religions, farmers capture non-farmers, residents of the region of Saint Louis capture residents of other regions and people belonging to the ethnic group Pulaar Toucouleur captures other ethnic groups. However, for round 2 we can see that in two cases the direction of the capture index changed. More specifically, the old people capture now the young, and people belonging to other ethnic groups capture those belonging to the Pulaar Toucouleur ethnic group. Table 8: Capture indices Round 1 Round 2 Rural vs. Urban 1.0901 1.0601 Women vs. Men 0.8889 0.8133 Old vs. Young 0.8738 1.1901 Educated vs. Uneducated 0.8115 0.8656 Married vs. Other 1.1363 1.2728 Muslim vs. Other 1.3247 1.0170 Farmer vs. Non Farmer 1.1403 1.3025 Saint Louis vs. Other 1.2390 1.7352 Pulaar Toucouleur vs. Other 1.0820 0.7265 Source: Own estimation Government capture indices were also estimated per type of treatment and we could see slightly different results between them. Tables 9, 10 and 11 display such results in the appendix. 6 Nash Equilibrium Based on the FOC, we identified the optimal policy positions of the incumbent party BBY for the issues PSvsEG, FoodvsCash and Ideology. The results are shown in figures 3. The curves in the Kernel distribution reflect the own policy positions of the voters, whereas the dots reflect the optimal policy positions of BBY. 24 Table 9: Capture indices for Positive Treatment Positive Treatment Round 1 Round 2 Rural vs. Urban 1.3755 1.1970 Women vs. Men 0.8741 0.7779 Old vs. Young 0.9627 1.4624 Educated vs. Uneducated 0.7472 1.2853 Married vs. Other 1.2356 1.2560 Muslim vs. Other 1.2886 0.8142 Farmer vs. Non Farmer 1.1562 1.6415 Saint Louis vs. Other 1.5568 1.4505 Pulaar Toucouleur vs. Other 1.1622 0.7819 Source: Own estimation Table 10: Capture indices for Negative Treatment Negative Treatment Round 1 Round 2 Rural vs. Urban 0.8530 1.1370 Women vs. Men 0.9876 0.9259 Old vs. Young 0.8926 0.9098 Educated vs. Uneducated 0.8291 0.6811 Married vs. Other 1.4227 1.7249 Muslim vs. Other 1.1856 3.0503 Farmer vs. Non Farmer 1.0491 1.0846 Saint Louis vs. Other 0.9502 1.5081 Pulaar Toucouleur vs. Other 0.9513 0.7331 Source: Own estimation 31 Table 11: Capture indices for Placebo Treatment Placebo Treatment Round 1 Round 2 Rural vs. Urban 1.1050 0.9026 Women vs. Men 0.8171 0.7621 Old vs. Young 0.7650 1.1330 Educated vs. Uneducated 0.8635 0.5787 Married vs. Other 0.8676 1.0192 Muslim vs. Other 1.8623 0.8225 Farmer vs. Non Farmer 1.2122 1.1868 Saint Louis vs. Other 1.2382 2.3785 Pulaar Toucouleur vs. Other 1.1395 0.6717 Source: Own estimation 32 Table 12: Model 1 - Latent Class Model Pre AIC = 991.6471 Class 1 (0.7114) Class 2 (0.2886) VARIABLES Coeff. z-value Coeff. z-value Model for Choices Attributes Abstention:(intercept) 4.2939 1.7413 . 0.5885 0.7624 BBY:(intercept) 1.6162 0.7055 0.1380 0.1623 Niang:(intercept) -5.2956 -0.5013 -3.1480 -1.2148 Pastef:(intercept) -0.2549 -0.0841 1.6658 2.0877 * PUR:(intercept) -1.3721 -0.4552 -0.3579 -0.2580 Rewmi:(intercept) 1.0125 0.2066 1.1136 1.4065 PSvsEG 0.0574 0.8937 -0.1728 -2.6754 ** FoodvsCash -0.0268 -0.4393 -0.1764 -1.9599 . Ideology 0.0854 0.8696 -0.0790 -1.0183 Party_ID 14.1557 1.8741 . 2.5418 4.0494 *** Predictors Abstention:Satisfaction_president -0.3469 -0.3242 0.3909 1.4573 BBY:Satisfaction_president 1.2987 1.2900 0.5141 1.5737 Niang:Satisfaction_president 0.6045 0.1412 0.4695 0.5538 Pastef:Satisfaction_president 0.2919 0.2431 -0.6343 -1.7181 . PUR:Satisfaction_president 0.6432 0.5553 -0.3602 -0.4520 Rewmi:Satisfaction_president -2.4913 -0.9271 -0.3801 -1.1145 Model for Classes Covariates classes:intercept -0.0853 -0.2081 0.0853 0.2081 classes:Trust_president 0.2901 3.0131 ** -0.2901 -3.0131 ** classes:LPI -0.1321 -1.3731 0.1321 1.3731 classes:factor_char_image -0.1861 -2.0168 * 0.1861 2.0168 * ***p<0.001, **p<0.01, *p<0.05, . p<0.10 Source: Own estimation 33 Table 13: Model 1 - Latent Class Model Post AIC = 942.2289 Class 1 (0.8937) Class 2 (0.1063) VARIABLES Coeff. z-value Coeff. z-value Model for Choices Attributes Abstention:(intercept) 2.7840 4.7969 *** -7.8903 -0.4925 BBY:(intercept) 1.3270 2.4180 * -8.0017 -0.4949 Niang:(intercept) -0.9484 -0.5983 -14.5631 -0.1874 Pastef:(intercept) -1.4893 -1.3836 18.4580 1.0726 PUR:(intercept) -1.8763 -1.5180 -6.8607 -0.4228 Rewmi:(intercept) 0.2031 0.2230 18.8578 1.0840 PSvsEG 0.0388 1.2450 -0.9397 -2.4714 * FoodvsCash -0.0687 -2.2419 * -1.3138 -2.2070 * Ideology -0.0160 -0.4187 -0.1604 -0.6880 Party_ID 6.4254 9.1402 *** -1.0964 -0.8223 Predictors Abstention:Satisfaction_president -0.3176 -1.4590 4.2837 0.8051 BBY:Satisfaction_president 0.7197 3.5303 *** 4.8304 0.8945 Niang:Satisfaction_president -0.4969 -0.6916 4.0698 0.1629 Pastef:Satisfaction_president 0.2618 0.7095 -8.2485 -1.3230 PUR:Satisfaction_president 0.2811 0.6524 3.4436 0.6300 Rewmi:Satisfaction_president -0.4482 -1.1557 -8.3792 -1.3255 Model for Classes Covariates classes:intercept 1.3636 2.8698 ** -1.3636 -2.8698 ** classes:Trust_president 0.2209 2.6619 ** -0.2209 -2.6619 ** classes:LPI -0.3099 -2.5293 * 0.3099 2.5293 * classes:factor_char_image -0.4391 -3.8899 *** 0.4391 3.8899 *** ***p<0.001, **p<0.01, *p<0.05, . p<0.10 Source: Own estimation 34 Table 14: Model 2 - Latent Class Model Pre AIC = 975.7887 Class 1 (0.8847) Class 2 (0.1153) VARIABLES Coeff. z-value Coeff. z-value Model for Choices Attributes Abstention:(intercept) 3.2017 3.3647 *** -0.0285 -0.0017 BBY:(intercept) 1.7758 1.9382 . -2.3945 -0.1387 Niang:(intercept) -2.4863 -1.1634 -7.5223 -0.0951 Pastef:(intercept) 1.1781 0.9903 6.1496 0.3599 PUR:(intercept) 0.1619 0.1292 -3.1533 -0.0886 Rewmi:(intercept) -3.8313 -1.3540 6.9489 0.4060 PSvsEG -0.0146 -0.4628 -0.5086 -2.3477 * FoodvsCash -0.0490 -1.3910 -0.5934 -1.8754 . TPvsAM 0.0117 0.4063 -0.7326 -2.8707 ** Ideology 0.0273 0.6288 -0.6641 -2.4915 * Party_ID 8.3664 3.6665 *** -2.8236 -1.6208 Predictors Abstention:Satisfaction_president -0.3421 -1.1634 0.4012 0.0732 BBY:Satisfaction_president 0.5817 2.0271 * 2.4152 0.4350 Niang:Satisfaction_president 0.1203 0.1637 0.6562 0.0257 Pastef:Satisfaction_president -0.6170 -1.4337 -1.4968 -0.2712 PUR:Satisfaction_president -0.3053 -0.6867 -0.2693 -0.0242 Rewmi:Satisfaction_president 0.5624 0.6496 -1.7065 -0.3091 Model for Classes Covariates classes:intercept 0.3536 1.6903 . -0.3536 -1.6903 . classes:Trust_president 0.2177 2.9032 ** -0.2177 -2.9032 ** ***p<0.001, **p<0.01, *p<0.05, . p<0.10 Source: Own estimation 35 Table 15: Model 2 - Latent Class Model Post AIC = 955.7233 Class 1 (0.8944) Class 2 (0.1056) VARIABLES Coeff. z-value Coeff. z-value Model for Choices Attributes Abstention:(intercept) 2.3633 4.3623 *** -6.1072 -0.4201 BBY:(intercept) 0.9728 1.9372 . -4.4858 -0.3084 Niang:(intercept) -1.2332 -0.8080 -11.0318 -0.1549 Pastef:(intercept) -0.2156 -0.2652 7.0703 0.4851 PUR:(intercept) -2.0355 -1.6679 . -3.4915 -0.2335 Rewmi:(intercept) 0.1482 0.1707 18.0459 1.1102 PSvsEG 0.0360 1.1228 -1.6544 -2.4979 * FoodvsCash -0.0719 -2.2005 * -0.4605 -1.6006 TPvsAM 0.0023 0.0770 -1.5804 -2.2027 * Ideology -0.0204 -0.5097 -0.0340 -0.1438 Party_ID 6.0629 10.3910 *** -1.5263 -0.9699 Predictors Abstention:Satisfaction_president -0.2357 -1.0940 3.0295 0.6420 BBY:Satisfaction_president 0.8270 4.2357 *** 3.1500 0.6652 Niang:Satisfaction_president -0.3883 -0.5606 2.7018 0.1190 Pastef:Satisfaction_president -0.0985 -0.3111 -2.9431 -0.6082 PUR:Satisfaction_president 0.3088 0.7186 2.1778 0.4489 Rewmi:Satisfaction_president -0.4134 -1.1097 -8.1160 -1.3233 Model for Classes Covariates classes:intercept 0.4224 1.9358 . -0.4224 -1.9358 . classes:Trust_president 0.2113 2.9697 ** -0.2113 -2.9697 ** ***p<0.001, **p<0.01, *p<0.05, . p<0.10 Source: Own estimation Table 16: Mean Probabilities for Separate Models - Model 1 Alternatives Pre Post Abstention 18.11% 16.32% BBY 71.89% 73.55% Niang 0.30% 0.34% Pastef 4.99% 4.67% PUR 0.99% 0.89% Rewmi 3.72% 4.23% Source: Own estimation 36 Figure 9: Relative Maginal Effects for Separate Models - Model 1 (a) Pre (b) Post Source: Own estimation 37 Table 17: Accountability indices for Separate Models - Model 1 Pre Post Whole sample 15.54% 29.69% Positive 15.09% 30.89% Negative 15.98% 30.36% Placebo 15.54% 27.87% Source: Own estimation Table 18: Capture for Separate Models - Model 1 Pre Post Rural vs. Urban 0.9829 1.0075 Women vs. Men 1.0079 0.8316 Old vs. Young 1.0695 1.1305 Educated vs. Uneducated 0.9199 0.8265 Married vs. Other 1.0737 1.3266 Muslim vs. Other 0.9771 0.9345 Farmer vs. Non Farmer 1.0703 1.1122 Saint Louis vs. Other 1.1568 1.7804 Pulaar Toucouleur vs. Other 0.9092 0.7085 Source: Own estimation 38 Figure 11: Optimal Policy Positions of Incumbent Party BBY for Separate Models - Model 1 (a) Public Services vs. Economic Growth (b) Food Crops vs. Cash Crops (c) Socialism vs. Capitalism Source: Own estimation References James Adams. Party Competition and Responsible Party Government. Ann Arbor, MI: University of Michigan Press, 2001. 39 James Adams, Samuel Merrill, and Bernard Grofman. A Unified Theory of Party Competition. Cambridge Univ. Press, Cambridge, 2005. African Union. African Transformation Scorecard progress report 2017. Published on the Thirtieth Ordinary Session of the Assembly of the Union, January 2018. Abhijit Banerjee, Rohini Pande Selvan Kumar, and Felix Su. Do informed voters make better choices? experimental evidence from urban india. Unpublished manuscritp, 2011. Pranab Bardhan and Dilip Mookherjee. Capture and governance at local and national levels. Political Economy, Governance, and Development, 90(2):135–139, 2000. Bryan Caplan. The Myth of the Rational Voter - Why Democracies Choose Bad Politics. Princeton University Press, Princeton, 2007. Stephen Coate. Political competition with campaign contributions and informative advertising. Journal of the European Economic Association, 2(5):772–804, September 2004. Constitutional Council of Senegal. Senegalese Presidential Election Results 2019. Published by the Constitutional Council of the Republic of Senegal (website), March 2019. URL: https://conseilconstitutionnel.sn/decision-n-4-e-2019-affaire-n-25-e-19/. Otto Davis, Hinrich J. Melvin, and Peter C. Ordeshook. An expository development of a mathematical model of the electoral process. American Political Science Review, 64: 426–48, 1970. Stefano DellaVigna and Ethan Kaplan. The fox news effect: Media bias and voting. Working Paper 12169, 2006. URL http://www.nber.org/papers/w12169. Anthony Downs. An Economic Theory of Democracy. Harper and Row, New York, 1957. Wouter Elsen. Fighting the Desert with Gardens in Senegal. Video published by Storyhunter on Youtube, November 2016. URL: https://www.youtube.com/watch?v=gt78gZmPiU4. James M. Enelow and Melvin J. Hinich. The Spatial Theory of Voting: An Introduction. Cambrigde University Press, New York, 1984. Robert Erikson and David Romero. Candidate equilibrium and the behavioral model of the vote. American Political Science Review, 84:1103–26, 1990. Morris Fiorina. Retrospective Voting in American National Elections. Yale University Press, New Haven, 1981. 40