Conserving wildlife by reducing demand: A restaurant experiment in Kinshasa

2023-10-26

2023-11-16

1) Meal choices: We will compare meal choices of treatment subjects (who watch bushmeat video) and of control subjects (who watched a non-bushmeat video)

2) Take-up: We will test whether treatment differentially affected coupon use.

3) "Extensive margin” choice of restaurant: We will test whether treatment affects the type of restaurant chosen.

4) Mechanisms: If subjects who saw a bushmeat video order less bushmeat, we will look into why they might have done so through a single index that captures different features of bushmeat (being sustainable, fresh, tasty, cool, legal, healthy, or connecting them to their place of origin)

5) Social desirability bias: We will compare a social desirability score (constructed following Dhar et al. (2022) between treatment and control units.

1) Specifically, estimate the following equation by ordinary least squares regression:

1{Bushmeat_ijt} = beta*T_i + alpha*X_i + kappa_j + zeta_t + epsilon_ijt (1)

where Bushmeat_ijt is an indicator that equals 1 if subject i ordered bushmeat and equals 0 if subject i did not order bushmeat, j denotes the table where the subject participated in the experiment, t denotes the date at which the subject participated in the experiment, beta is the coefficient of interest, T_i is an indicator that equals 1 if subject i was in the control group and equals 0 if subject i was in the treatment group, alpha is a vector of coefficients, X_i is a matrix of individual controls, kappa_j are table fixed effects, zeta_t are date fixed effects, and epsilon_ijt is the error term. In all regressions where the unit of observation is an individual subject (a person), we will cluster standard errors at the individual subject level since that is the level at which treatment will be assigned. We include individual-level control variables and fixed effects in order to increase the precision with which beta is estimated. We include the following 8 variables in X_i: age of subject in years, an indicator that equals 1 if the subject is male and equals 0 otherwise, an indicator that equals 1 if the subject has attained university (post-secondary) education and equals 0 otherwise, the number of years of education the subject has, an indicator that equals 1 if the subject has earned a salary in the past 7 days and equals 0 otherwise, an indicator that equals 1 if the subject has earned income as the proprietor of a business in the last 7 days and equals 0 otherwise, the number of times the subject has eaten bushmeat in the last 30 days, and an indicator that equals 1 if the subject eats bushmeat at formal OR informal restaurants and equals 0 otherwise. If an individual-level control variable is missing and the variable is not an indicator variable, we will impute the value with the mean among all non-missing values. For example, if 50 subjects have missing age, we will assume their age equals the mean age among all 550 subjects with non-missing age.

2) Take-up: We will test whether treatment differentially affected coupon use by estimating the following equation by ordinary least squares regression:

1{Use Coupon_ijt} = beta*T_i + alpha*X_i + kappa_j + zeta_t + epsilon_ijt (2)

Where 1{Use Coupon_i} equals 1 if subject i used the coupon (and ordered a dish at their restaurant of choice) and equals 0 otherwise, and the other variables are defined in Equation 1.

3) “Extensive margin” choice of restaurant: We will test whether treatment affects the type of restaurant chosen by estimating the following equation by ordinary least squares regression:

1{Bushmeat Restaurant_ijt} = beta*T_i + eta_1*Dist_ij + eta_2*Price_i + alpha*X_i + kappa_j + zeta_t + epsilon_ijt (3)

Where 1{Bushmeat Restaurant_ijt} equals 1 if subject i chose one of the two restaurants that feature more bushmeat on the sample menu we show to subjects and equals 0 if subject i chose one of the two restaurants that features less bushmeat on the sample menu, eta_1 and eta_2 are coefficients, Dist_ij is the distance in km between subject i’s table j and the restaurant they chose, Price_i is the average price in XOF over all dishes on the sample menu of the restaurant that i chose, and all other variables are defined in Equation 1. We define 1{Bushmeat Restaurant_ijt} and Price_i using the sample menus we show to subjects in the experiment. Specifically, the enumerator will read to the subject, for each restaurant, the different dishes available at that restaurant and the price of each dish in XOF. We will rank restaurants by the percentage of dishes made up by bushmeat. For example, if the sample menu we provide to subjects for Restaurant 1 contains 1 bushmeat dish and 1 non-bushmeat dish, the percentage for Restaurant 1 is 50%. The two restaurants that feature more bushmeat (1{Bushmeat Restaurant_i} = 1) will be the two with the highest percentage. If there is a tie, we will break the tie with the number of bushmeat dishes on the menu. For example, if Restaurant 2 has 2 bushmeat dishes and 2 non-bushmeat dishes on the menu, we would rank Restaurant 2 above Restaurant 1. If there is still a tie, we will calculate the average price of bushmeat dishes at each restaurant, and rank restaurants with less expensive bushmeat dishes above those with more expensive bushmeat dishes. If there is still a tie, we will randomly assign tied restaurants to be defined as featuring more or less bushmeat on the menu.

4) Mechanisms: If subjects who saw a bushmeat video order less bushmeat, why might they have done so?

After watching the video and choosing the restaurant where they want a coupon to, we will ask subjects eight questions related to their view of bushmeat. Each question asks subjects whether they completely disagree, partially disagree, neither disagree nor agree, partially agree, or completely agree with a statement. The first seven statements are of the form: bushmeat is…. And the seven individual attributes are: sustainable (durable), fresh (fraîche), tasty (savoureuse), cool (cool), legal (légale), healthy (saine), and connects them to their place of origin (relie à votre lieu d’origine). The eighth statement is that they are proud of the environment of Democratic Republic of Congo (fier de l'environnement de la République démocratique du Congo).

We aggregate responses to these eight questions into a single index. For each question, we code a value of 1 if the response indicates a negative view of bushmeat. For each of the first seven statements, if the subject partially OR completely disagrees with the given statement we code the response as 1, and if the subject neither disagrees nor agrees, partially agrees, or completely agrees we code the response as 0. For the eighth and final statement, if the subject partially agrees or completely agrees we code the response as 1, and code the response as 0 otherwise. Then we sum the responses over the eight questions for each subject. Thus each subject can have a value of between 0 and 8. Finally we standardize the summed score by subtracting the mean summed score across all subjects, and then dividing by the standard deviation of the summed score across all subjects.
We estimate the following equation by ordinary least squares regression:

StdScore_ijt = beta*T_i + alpha*X_i + kappa_j + zeta_t + epsilon_ijt (4)

where StdScore_ijt is subject i’s standardized response score and all other variables are defined in Equation 1.

To further understand any potential mechanisms, we will also estimate the effect of treatment assignment on each of the eight variables individually. The dependent variable in each regression will be 1{Unfavorable View_i}, an indicator variable that equals 1 if subject i partially OR completes disagrees with the given statement (in the case of the first seven statements). In the case of the eighth statement regarding pride in Congo’s environment the indicator variable will equal 1 if the subject partially agrees or completely agrees. We will therefore estimate eight equations of the following form by ordinary least squares regression:

1{Unfavorable View_i} = beta*T_i + alpha*X_i + kappa_j + zeta_t + epsilon_ijt, (5)

5) Social desirability bias: Several times during the primary experiment the enumerator will instruct the subject that they should order whatever dish they most want to eat at the restaurant, whether that dish be bushmeat or not. Nonetheless, it is still possible to feel concerned that treated subjects may order less bushmeat due to experimenter demand effects, rather than due to having been persuaded by the bushmeat video. We therefore implement the method of Dhar et al. (2022; American Economic Review) in order to assess the threat that experimenter demand effects pose to our interpretation of beta in Equations 1 and 2. Dhar et al. (2022) “use a 13-question short form of the Crowne and Marlowe (1960) module developed by Reynolds (1982).” Given each statement, subjects answer whether they completely agree, partially agree, neither agree nor disagree, partially disagree, or completely disagree. We code the response to each statement as 1 if the subject gives a socially desirable answer. For example, if a subject completely disagrees OR partially disagrees with the statement “I sometimes feel resentful when I don’t get my way”, we would code their response as 1 (and code the response as 0 if they neither disagree nor agree, partially agree, or completely agree). We will sum the coded responses over statements, so that subjects have a social desirability score of between 0 and 13. Then we will standardize the score by subtracting the mean score across all subjects, and then dividing by the standard deviation of the score across all subjects. We augment Equation 1 by adding two variables: the standardized social desirability score, and the interaction of the treatment indicator and the standardized social desirability score. If the coefficient on the interaction term is not different from 0, then it is less likely that social desirability bias/experimenter demand effects are differentially changing what treated subjects order at restaurants. We estimate the following equations by ordinary least squares regression:

1{Bushmeat_ijt} = beta*T_i + sigma_1*StdSoc_i + sigma_2*T_i*StdSoc_i + kappa_j + zeta_t + epsilon_ijt, (5)

where sigma_2 represents the coefficient of interest for assessing differential experimenter demand effects, StdSoc_i is the standardized social desirability score of subject i, and all other variables are defined in Equation 1.

6) Consumption bushmeat at restaurants where price of non-bushmeat is subsidized vs. at places where the price of non-bushmeat is not subsidized.

We will answer this question using daily restaurant sales data as well as the variation in Moambe Chicken prices induced by our secondary experiment. Specifically, we will use restaurant-day-level treatment assignment as an instrument for Moambe Chicken price in the first-stage ordinary least squares regression:
Log(P_{it}) = gamma_1*T_{it} + delta_1*Q_{i0} + delta_2*Price_{i0} + epsilon_{it} (7)
Where Log(P_{it}) is the log price of Moambe Chicken (a common chicken dish) in XOF at restaurant i on day t, gamma_1 is a coefficient, T_{it} equals 1 if restaurant i was treated on day t (paid by us to reduce the price of Moambe Chicken) and equals 0 otherwise, delta_1 and delta_2 are coefficients, Q_{i0} is the number of plates of bushmeat sold the week before the secondary experiment, Price_{i0} is the average price in XOF over all dishes on the sample menu of restaurant i that we showed to subjects, and epsilon_{it} is the error term. We cluster standard errors at the restaurant-day level since that is the level at which we will assign treatment.
We estimate the reduced form equation by ordinary least squares regression:
Log(Q_{it}) = gamma_2*T_{it} + delta_3*Q_{i0} + delta_4*Price_{i0} + epsilon_{it} (8)
Where Log(Q_{it}) is the log number of plates of bushmeat sold at restaurant i on day t, gamma_2, delta_3, and delta_4 are coefficients, and all other terms are defined in Equation 7.
Then our object of interest, the elasticity of bushmeat consumption with respect to Moambe Chicken price, is gamma_2 / gamma_1. In practice, we will estimate the instrumental variables regression in a single command in R using the feols() command in order to obtain the cross-price elasticity’s correct standard error.
As a validation of the secondary experiment, we will also estimate the own-price elasticity of Moambe Chicken consumption by re-estimating Equation 8 except replacing the dependent variable with Log(number of plates of Moambe Chicken sold by restaurant i on day t), and dividing the treatment coefficient by gamma_1. In practice, we will similarly also estimate this instrumental variables regression in a single command in R using the feols() command in order to obtain the own-price elasticity’s correct standard error.

Primary experiment:

1. Set up 4 tables in upscale parts of Kinshasa. Encourage passersby to approach the tables to possibly receive a coupon to their restaurant of choice. Only one subject may participate in the experiment at a time. The enumerator will not begin an experiment with the next subject until the previous subject has left the area (to avoid the next subject having a conversation with the previous subject, or seeing the video watched by the previous subject).
2. Ask subject for their consent and record their identifying information (to reduce repeat participants). Ask subject if they have consumed bushmeat in the last month. If they haven't, drop them from the experiment. If they have, proceed to the next step.
3. Subject watches a short video on a tablet with headphones. Subjects in the treatment group watch a 90-second clip produced by the Republic of Congo's Forestry Ministry for their 2019 “Let's eat less bushmeat in the city" campaign (https://www.facebook.com/MEFCG/videos/2393355434020632/). Subjects in the control group watch a 90-second clip from a Congolese soap opera that does not discuss bushmeat or the environment (https://www.youtube.com/watch?v=-qAE46cqHbY; clip begins at 38:15 and ends at 39:45). Randomization is at the individual-level.
4. Subject chooses which restaurant they would like a coupon to. The restaurants on the list will be “approved establishments" in the sense that the owners have agreed to accept coupons from us. The list will contain information about each restaurant (location, name, dishes offered by the restaurant, and the price of each dish).
5. Subject completes short survey, answering questions related to their personal views of bushmeat and the bushmeat consumption patterns of their social network (e.g., their friends, colleagues, and family), and whether they have seen the video before. Subjects also complete 13 questions related to social desirability bias so that we can assess experimenter demand effects (Dhar, Jain, & Jayachandran, 2022).
6. Subject receives a coupon for 3000 XOF. Each coupon contains a unique ID. Enumerator tells subject (for the second time) that they should order whatever they want at the restaurant.
7. Subject goes to the restaurant they chose and shows the worker their coupon. Subject may order whatever they want (bushmeat or non-bushmeat). The worker messages us the coupon ID and what the subject ordered. Our survey firm (Innovative Hub for Research in Africa) pays the restaurant owner for every coupon that is redeemed. Our sample size is 600 subjects, not all of whom will actually redeem their coupons. We will recruit 4 restaurants to participate in the experiment.

Secondary experiment:

Participating restaurants will also be asked to reduce the price of Moambe Chicken on certain days by 1000 XOF. In exchange, we will pay restaurants 1000 XOF for each plate of Moambe Chicken they sell on those days. Randomly reducing the price of Moambe Chicken will allow us to assess whether reducing the price of non-bushmeat affects consumption of bushmeat. We will also ask participating restaurants to share daily sales data with us, specifically the number of bushmeat, Moambe Chicken, and other non-bushmeat plates sold and revenue from each, so that we may assess whether sales of bushmeat and non-bushmeat change.

Randomization will be done by a number directly generated on the survey platform. Specifically, the number will be randomly generated to be between 0 and 1. If the number is less than 0.5, then the subject will be in the control group. If the number is more than 0.5, then the subject will be in the treatment group.

Randomization will be at the individual level.

Yes

600

600

300 individuals for both treatment and control groups