Gender price gaps and competition: Evidence from a correspondence study

2018-07-02

2018-11-30

total price estimates

The primary variable of interest is the price estimate received by each gender and customer type.
Prices will be defined using the following rules:
1) Use the total price provided in the estimate
2) Use the average price if price ranges are provided,
3) Use the price after discounts, whenever discounts are offered,
4) Use the first price provided by a shop, whenever the same shop provides more than one price
in separate emails
5) Drop prices when either the labor or radiator price are explicitly excluded in the estimate,
6) Drop extremely low or high prices, below $100 and above $2,000, and estimates where the
price range provided is so wide that the maximum price is more than twice as large as the minimum
and price composition is not given. These thresholds are arbitrary but conservative values. I will
add robustness exercises varying these thresholds.
7) Use price provided by the shop, whenever price matching is offered. In alternative specifications,
I will use the price picked from a random draw of lower nearby prices obtained by the same
user and customer type.

The proportion of emails replied, the proportion of estimates given, the proportion of estimates given on first email reply, price composition

I will include indicator variables for receiving a shop reply, replying to a shop once before receiving a price and receiving a price estimate. I will use these variables to compare response rates and effort it takes to obtain prices. Shops often provide detail on the total estimate composition. I will keep track of included and explicitly not included parts, along with their prices and quantities. I can potentially explore if
shops tend to recommend overtreatment and how price components vary with customer characteristics.

The experimental design uses a correspondence-study approach.

The experimental design uses a correspondence-study approach, in which I send emails to shops varying the perceived gender and customer information revealed to each shop.

I do a simple randomization at the individual shop level. For the first round of emails, I randomly assigned one of the eight treatment groups (based on the combinations of perceived gender and customer type) to each shop. I determined the distribution of customer types using the optimal shares obtained in the power calculations. I assigned 18.3% standard scripts, 15.7% low-search cost scripts, 8% high-education and 8% uninformed scripts to each gender.4 I also randomized which other characteristics are used to contact each shop. These characteristics are the script within each customer type, the email subject and the car year. The scripts and email subjects provide the same information but are reorganized and slightly differently worded, and the car years are 2009 and 2010, which belong to the same car generation.
The purpose of these additional variations is to avoid sending a second email that is identical in all but customer’s name and type.
For the second round of emails, I will keep all valid auto repair shop emails. I will change the gender and customer type with respect to that used for the first email, restricting the within shop comparison to always include a standard type-script.

The first experimental design is simple randomization.
The second experimental design is at the shop level, sending a second email to the same shop.

No

Approximately 60,000 shops

Approximately 60,000 shops.

Initial sample size is approximately 60,000 shops. Half will be contacted using an account with a female name, half with a male's account name. 18.3% of shops will be contacted by a female customer using a standard script, 15.7% by a low-search cost female customer, 8% by a high-educated female and 8% by an uninformed female. Equivalent shares are assigned to each male-type combination.

I should be able to detect a difference in prices between males and females within types of $14.7 to $21.9, given a statistical power of 80%, 1% significance level and sample size of 10,000 observed prices.