Quality upgrading in the street food market: are better infrastructure and training sufficient?

We are providing (subsidizing) for free a set of infrastructure to vendors that are too costly for them to buy and that are not provided by the local authorities either. This is an in-kind transfer worth 400$ per vendor (roughly, 3-month net income in our context). In addition, we also plan to conduct a comprehensive series of discrete choice experiments on consumers to elicit their preferences and experiences with safe street food. The experiment is to test whether this subsidized investment is sufficient to sustain a “better equilibrium” in terms of adoption and practices (that we can observe and measure) concerning the production of safer food. The simple hypothesis underlying our intervention is that this subsidized investment could enable vendors to offer consumers a wider array of desirable and profitable products and services by removing the cost barrier. We intend to verify that this is appropriate in our context by investigating both that vendors have high perceived costs of providing hygienic services at baseline and that consumers have a positive willingness to pay for food that is perceived as more hygienic.

*** Intervention

We focus only on street food vendors selling heavy or medium-heavy food in a permanent built-up structure. Specifically, we sampled vendors satisfying the following criteria:

- The vendors must be selling cooked food (in the kiosk or other places);
- The vendors must be selling meal or lunch/dinner items;
- The vendors must be selling cooked snacks (at least 5 - 7 items) of the following items: Chowmein, Roll, momo, fried chicken items and dhosha, etc.

We are providing (subsidizing) for free a set of infrastructure to vendors that are too costly for them to buy and that are not provided by the local authorities either. The experiment is to test whether this subsidized investment is sufficient to sustain a “better equilibrium” in terms of adoption and practices (that we can observe and measure) concerning the production of safer food. The simple hypothesis underlying our intervention is that this subsidized investment could enable vendors to offer consumers a wider array of desirable and profitable products and services by removing the cost barrier. We intend to verify that this is appropriate in our context by investigating both that vendors have high perceived costs of providing hygienic services at baseline and that consumers have a positive willingness to pay for food that is perceived as more hygienic.

This study will run for 12 weeks. We will conduct a weekly monitoring survey to track whether vendors are using the infrastructure and what impacts this may be having on various business outcomes.

*** Details about the treatment groups

- C group: In the pure control group, vendors receive the personal banner and the menu card displaying prices.

- T1 group: In the first treatment group, vendors will receive the same set of items as given to C, plus the infrastructure to produce safer food. This infrastructure includes chlorine tablets to treat the main water source the vendor uses, storage facilities for both garbage and drinking water, soap for dish and handwashing, and disposable plates and gloves. This is a supply-side intervention

- T2 group: In the second treatment group, vendors will receive the same set of items as given to T1, plus capacity building. What we mean by this is that, each week, when we visit them, they will get a reminder about how to use better the infrastructure that we provided them. This is a stronger supply-side intervention.

2022-06-06

2022-08-19

Behavior of the street food vendors in terms of adopting safe and hygienic food handling methods. We are interested in whether or not they use the infrastructure provided; if so, what are the impacts on their business outcomes, if not, why and what can be done to improve the safety of street food given this. Some specific outcomes we are interested in are:

Is there a visible facility for the vendor to wash hands?
If yes, is there soap visible nearby?
How does the water tank to clean the dishes look at the moment?
Does the vendor use soap to clean the dishes?
Is there a towel/cloths for the vendor to wipe hands?
Is the vendor using an apron?
Is the vendor wearing a hair cover?
Does the stall have a garbage bin?
If yes, how does the garbage bin look like at this moment?
Are the ingredients, raw or half-cooked food separated from cooked food?
Is the cooked food covered?
Are the ingredients/raw food covered?
Are tongs/spatulas/other tools being used for cooking?
Is the food served with spoons, tongs or any other tools?
Does the vendor wash hands before cooking/handling food?
What does the counter where food is prepared look like?
Is the food being served on disposable plates?
If no, what is the food being served on?
Is there a garbage bin for consumers to dispose of their waste?
If yes, how does the garbage bin look like at this moment?
If there a facility for drinking water for customers?
If yes, how is drinking water distributed?
Is there a hand washing facility for customers?
If yes, how does this facility look at the moment?
If yes, is there soap available?

We will construct various summary measures of the key outcome variables in the form of weighted indexes.

Awareness of the street food vendors in terms of health hazards, as:

Are you familiar with the National Act for Urban Street Vendors 2014?
What do you think are possible contaminants of food; that is, things that can make food not safe to eat?
What do you use (or instruct employees to use) to have clean hands?
When do you think it is needed to wash hands?
Do you treat the water used for cooking or cleaning?
If yes, how do you treat the water used for cooking and cleaning?
Do you treat the drinking water provided to customers at your stall?
If yes, how do you treat the drinking water provided to customers at your stall?
How often do you replace the water in which dishes are washed?
How often do you empty your garbage bin?
How do you check if meat is cooked properly?
What do you do with leftover cooked food at the end of the day?
What do you do with leftover raw ingredients at the end of the day?

Over the past week, what days of the week was your kiosk closed?
Over the past week, on average, how many hours per day did you spend preparing the food to sell?
Over the past week, on average, how many hours per day did you sell street food?
Over the past week, on average, how many minutes per day did you spend cleaning at the end of the day?
Over the past week, on average, how many customers did you usually have each day?
Over the past week, on average, what were your daily sales?
Over the past week, on average, what were your daily total expenditures?
Over the past week, on average, what were your daily profits?
What do you expect your daily profits to be over the next 7 days?


3.2. Prices
Over the past week, did you display the personalized menu card that we gave you?
Over the past week, did you change any of the prices displayed on your menu card?
If yes, are the price changes publicly visible?
If yes, please note the new prices
If no, did you change the price of the items?

Our design is a stratified (blocked) random assignment with 2 levels. Hence, the sample is divided in clusters of street vendors which are then randomly assigned to different treatment groups, namely: treatment C, T1 and T2. Since the assignment is at the cluster level, all vendors in a particular cluster are assigned to the identical treatment group. The assignment of the treatments, coupled with regular monitoring of their food handling practices throughout the intervention, allows us to observe and identify changes in their behavior. The different treatment groups have different incentive schemes to reward the vendors for improving their practices, allowing us to identify the relative importance of factors that might be prompting their decisions. Finally, a measure of the spatial distance between different clusters allows us to estimate the spillover effects or externalities arising from the treatments.

*** Main characteristics of our design

● Areas:
o There are 3 areas: 1) Sector V, 2) Dalhousie and 3) Hazra.
o Sector V is the IT hub of Kolkata, where consumers have higher income and vendors adopt safer practices.
o Both Dalhousie and Hazra are quite large and diverse areas in KMC.
o Dalhousie is the central business district of Kolkata. It means that, practically, Dalhousie is the only office area, so major consumers are employees. This area has a greater number of vendors and a variety of foods available. There are only food vendors here.
o Hazra (+ Rashbihari) has offices, hospitals, religious temple (India famous Kali temple – every year hundreds of pilgrims from different parts of India visit the temple), markets, cinema hall and residential area. So, the consumer types are employees, tourists, pilgrims, shoppers, residential people, people visiting hospital etc. and that also implies the consumer’s financial capacity has a huge difference. This area has a smaller number of vendors than Dalhousie and also less variety of food. Also, there are not only food vendors; there are many types of vendors like garments, shoes, jewellery, etc.

● Clusters:
o In each area, we divide vendors in natural urban clusters of 1, 2 to 3, 4 to 8 vendors each.
o A cluster is defined in the following way:
▪ If a vendor does not have anyone nearby in a ray of 30 meters forms a cluster on his own. This is because he is a “monopolist” of food in his surroundings.
▪ If a vendor has 1 or 2 neighbours nearby in a ray of 30 meters forms a cluster of 2 or 3 vendors. They are “oligopolists” of food in their surroundings.
▪ In streets where there are many vendors one after the other, we take the distance from the leftmost vendor to the rightmost vendor and consider a cluster the collection of those vendors that are at most 30 meters apart from each other. This means that in some cases we have clusters of 4 vendors or 5 vendors. In rare cases, we have clusters of 6 vendors (1 case) or 8 vendors (3 cases).
▪ In our sample, there are 88 clusters in total. 31 in Sector V, 29 in Dalhousie, and 28 in Hazra.

● Vendors are allocated to 3 treatment groups:
o C: receive a personal banner and menu card.
o T1: receive the same as C, plus the infrastructure.
o T2: receive the same as T1, plus capacity building.

● Group sizes:
o C: 100 vendors, distributed in roughly 35 clusters.
o T1: 90 vendors, distributed in roughly 35 clusters.
o T2: 90 vendors, distributed in roughly 35 clusters.

Hence, our sample is made of 280 vendors, in roughly 90 clusters, distributed in 3 areas (Sector V, Dalhousie, Hazra).

Randomization done in office by a computer.

Randomization is done at the cluster level, i.e. small groups of vendors working close to each other in the same street.

Yes

105 natural urban clusters of vendors on the street.

280 street vendors.

100 street vendors in C, 90 street vendors in T1, 90 street vendors in T2.

Our design is a stratified (blocked) random assignment with treatment at level 2 and outcome variables at level 1. The calculations below show that our experiment is able to capture the following effects: For p_ind_total: 0.21 < effect size < 0.33 For p_ind_facilities: 0.21 < effect size < 0.37 For p_ind_handling: 0.21 < effect size < 0.34 For p_ind_costumer: 0.21 < effect size < 0.36 More in details, for power analysis, we can use the Stata command -pc_simulate- developed by Burlig, Fiona, Louis Preonas, and Matt Woerman (2020). "Panel Data and Experimental Design." Journal of Development Economics 144: 102548. The command provides power calculations for a setting like ours if you already have pilot data that you can use. We do have data from 2015-2016 on 4 indexes: 1) p_ind_facilities, 2) p_ind_handling, 3) p_ind_costumer and 4) p_ind_total. These data are for 1 wave before treatment and 5 waves after treatment. Since in this current experiment we have 12 waves after treatment, we can only do a power analysis assuming half of the sample size. This power analysis will be our upper bound. * cluster-level random assignment set seed 123 *** Index 1 pc_simulate p_ind_facilities, model(ANCOVA) mde(0.065 0.07 0.075) i(id) t(period) n(64) p(0.5) pre(1) post(5) alpha(0.05) /// vce(cluster) idcluster(block) sizecluster(3) bootstrap replace sum p_ind_facilities di 0.075/.2020061 * Effect size < 0.37127592 *** Index 2 pc_simulate p_ind_handling, model(ANCOVA) mde(0.065 0.07 0.075) i(id) t(period) n(64) p(0.5) pre(1) post(5) alpha(0.05) /// vce(cluster) idcluster(block) sizecluster(3) bootstrap replace sum p_ind_handling di 0.065/.1883959 * Effect size < 0.34501812 *** Index 3 pc_simulate p_ind_costumer, model(ANCOVA) mde(0.075 0.08 0.085) i(id) t(period) n(64) p(0.5) pre(1) post(5) alpha(0.05) /// vce(cluster) idcluster(block) sizecluster(3) bootstrap replace sum p_ind_costumer di 0.08/.223318 * Effect size < 0.35823355 *** Index 4 pc_simulate p_ind_total, model(ANCOVA) mde(0.035 0.04 0.045) i(id) t(period) n(64) p(0.5) pre(1) post(5) alpha(0.05) /// vce(cluster) idcluster(block) sizecluster(3) bootstrap replace sum p_ind_total di 0.045/.1363067 * Effect size < 0.33013784 To calculate a lower bound, we use the familiar -sampsi- command. This command does not require previous data, hence it can be used to perform power calculations assuming any number of repeated measures of the observations. The problem with -sampsi- is that it does not allow for cluster randomization. Hence, power calculation with -sampsi- will constitute a lower bound because, even though we can assume we have 12 waves, we need to assume that ICC = 0. More specifically, we need to make 3 simplifications: 1) Treatment is administered at level 1. ● In our experiment, the treatment is administered at level 2 for simplification and logistical reasons. ● However, vendors run their business individually, and the treatments are delivered at the individual level. Hence, given the context and the specific implementation, this is not a strong simplification. ● In practice this means that ICC is very small. 2) Stratification does not help explaining the outcome variables. • We use two variables on which we stratify on. • Based on the data from 2015, the variable “area” explains only 5% of the variation. • We don’t have data that helps understand how much cluster size explains. • So, in practice, this is a conservative simplifying assumption. 3) Spillover effects are negligible. • We don’t know whether ex ante spillover effects are negligible. • Based on the data from 2015, it is unlikely to observe spillover effects. Also because vendors are credit constrained. • In practice this means that, if there is an effect, this effect is only “direct”. The indirect effect = 0. • However, given how vendors are distributed on the street, and given the fact that we know exactly each vendor’s GPS location, we will be able to measure spillover effects. So, we will be able to relax this assumption. In our context, we have 90 vendors per treatment group. We observe each vendor 1 time before treatment, and 12 times after treatment. To calculate key parameters, I use the data from 2015. I type the following codes: use 1.data_vendors.dta,clear keep if food_meal==1 | food_heavy==1 tsset id period Then, for each index of behaviour, I calculate the correlation between follow-up measurements: reg p_ind_total L.p_ind_total i.area reg p_ind_costumer L.p_ind_costumer i.area reg p_ind_handling L.p_ind_handling i.area reg p_ind_facilities L.p_ind_facilities i.area Each of the following calculations guarantee a power > 0.80 * Power C vs T1 sampsi 0 0.21, sd(1) method(ancova) pre(1) post(12) r1(.3) n1(100) n2(90) sampsi 0 0.20, sd(1) method(ancova) pre(1) post(12) r1(.2) n1(100) n2(90) * Power C vs T1 + T2 sampsi 0 0.18, sd(1) method(ancova) pre(1) post(12) r1(.3) n1(100) n2(180) sampsi 0 0.17, sd(1) method(ancova) pre(1) post(12) r1(.2) n1(100) n2(180) * Power T1 vs T2 sampsi 0 0.22, sd(1) method(ancova) pre(1) post(12) r1(.3) n1(90) n2(90) sampsi 0 0.20, sd(1) method(ancova) pre(1) post(12) r1(.2) n1(90) n2(90) Hence, under the above simplifications, and using ANCOVA method, we should be able to estimate a MDES > 0.21.