During this application season I also interviewed for the Amazon economist summer intern role, so I'm writing up my experience here to share with anyone who needs it.
Timeline:
- Mid-January 2025: had a senior refer me
- 03/04/2025: HR reached out
- 03/26/2025: first round
- 03/28/2025: got the email that I passed the first round
- 04/16/2025: second round
- 04/16/2025: two hours later that same day, heard I passed the second round and moved into team matching
Amazon has always done two rounds. In past years round one was case + BQ and round two was coding, but this year they changed it so both rounds were case + BQ.
Round 1:
60 minutes.
First we introduced ourselves to each other.
Then came the BQ (leadership questions) — the interviewer's two questions focused heavily on communication and ownership. The first was "tell me about a time when you failed a commitment," and the second was "tell me about a time you took on someone else's responsibility." Each question came with a lot of follow-up detail — for the first one, things like how the commitment was formed in the first place, whether it was a strong commitment, why it failed, and what the final outcome was; for the second, how you balance your own work against other people's, and why you made the decision you did.
Then the case:
Context: we want to add live streaming of certain sporting events to Prime Video.
Business sense: what are the potential benefits? I answered: 1. Prime Video subscriptions; 2. ad revenue during the livestream; 3. revenue from driving traffic to other Amazon apps during the livestream.
Model: we want to evaluate the impact of adding the livestream on subscriptions.
I listed feasible models — matching, IV, DID — and said I'd start with matching, first addressing bias from observables.
The evaluation happens at the individual level: define people who watched the livestream as the treatment group, and people who didn't as the control group.
I then talked through choosing the outcome, and which characteristics and user behaviors to match on.
In the back-and-forth I answered questions about the matching assumptions and how to actually execute it — for example, if you use KNN for matching, how exactly do you do that, are there other matching methods, what model do you use to estimate the propensity score, and what do you do if it overfits.
At the end of the interview I mentioned that if you wanted to address bias from unobservables, you could use IV — even though everyone technically has access to the livestream, you could introduce some random variation in how the livestream promotion is displayed, e.g. some people get a very prominent promotion and others a less prominent one, and use that as an instrument for whether they watched the game or not. The interviewer had a short follow-up asking how exactly I'd implement that, and I answered with the variable definition and 2SLS.
Round 2:
60 minutes.
This time the interviewer skipped the introductions and went straight into explaining the process.
First was BQ — two questions: "tell me about a situation where you dug deep to find a root cause" and "tell me about a situation where you were unsatisfied with certain aspects of your team," again with very detailed follow-ups.
Then the case:
Context: when a customer's item gets lost in shipping or arrives damaged, we normally issue a full refund or a replacement. In some regions we've started a "concession" experiment where we send affected customers a gift card instead, and we want to evaluate the impact of this program.
I started by asking a lot of clarifying questions: 1. When we handle these returns and refunds, and when we run these programs, are we trying to reduce negative reviews, or something else — this determines what our main outcome should be; 2. Roughly how many regions have this experiment running, and do we have data from before the experiment started — this determines whether we should use synthetic control, or whether matching or DID would work; 3. Do we have individual-level data for these cities — this determines what level the study should be conducted at; 4. What data do we actually have.
Once those questions were answered, I said we could start with DID, and we went into detail on the regression equation setup, what assumptions it relies on, and how to test them.
Then the interviewer asked what to do if the parallel trends assumption doesn't hold. I gave a few approaches, one of which was PSM-DID, and from there we moved into a discussion of PSM — the setup, the implementation, what assumptions it makes, and how to test them.
Finally, they asked: if you could design the experiment yourself, how would you do it — at the individual level or the city level?
Overall, the whole interview process was pretty...
Discussion
Loading comments…