transcribe

DoorDash Data Scientist Mock Interview (Commentary)

DataInterview · 14m · transcribed Jul 2026
More from DataInterview Business
𝕏 Share ▶ YouTube 📥 PDF 🤖 .md

Section Insights

# 0:00

Introduction to Mock Interview

What can candidates expect from this mock interview?

Candidates will go through product data science questions, focusing on statistics and case studies, with commentary on their performance.

  • The mock interview will last about 10 minutes.
  • Candidates should prepare for statistics and case study questions.
  • Using a scratch sheet for brainstorming is encouraged.
  • Virtual interviews require clear communication of thought processes.
# 2:54

Understanding Statistical Power

What does 90% power mean in statistics?

90% power refers to the probability of rejecting the null hypothesis when the alternative hypothesis is true, which differs from the candidate's incorrect definition related to type 2 error.

  • Candidates must grasp fundamental statistical concepts for interviews.
  • Incorrect definitions can lead to misunderstandings in interviews.
  • Using index cards for definitions can aid in preparation.
  • Practical significance in experiments must be understood alongside statistical significance.
# 5:48

Practical Significance in Experiments

How do you determine practical significance in an experiment?

Practical significance depends on the cost-benefit analysis and whether the results make business sense, not just statistical significance.

  • Statistical significance does not guarantee practical significance.
  • Business stakeholders should be involved in defining meaningful results.
  • Multiple metrics should be considered when evaluating experimental outcomes.
  • Understanding the implications of results is crucial for decision-making.
# 8:43

Type 1 and Type 2 Errors

How do you calculate type 1 and type 2 errors in statistical tests?

Type 1 error is the probability of rejecting a true null hypothesis (false positive), while type 2 error is failing to reject a false null hypothesis (false negative).

  • Type 1 error rate is typically set at alpha (e.g., 0.05).
  • Type 2 error rate is the complement of power.
  • Understanding these concepts is essential for accurate statistical analysis.
  • Candidates should be prepared to explain these concepts clearly.
# 11:37

Calculating Error Rates

How do you calculate type 1 and type 2 error rates given alpha and power?

Type 1 error rate is equal to alpha, while type 2 error rate is the complement of power, which can be calculated based on the definitions of these errors.

  • Candidates should be able to apply definitions to calculate error rates.
  • Struggling with calculations can be mitigated with hints from the interviewer.
  • Clear understanding of alpha and power is critical for statistical testing.
  • Practice with real scenarios can enhance confidence in explaining these concepts.

Transcript

0:01 hello candidates my name is dan lee i'm a data scientist interview coach at topdies.org in the next couple minutes we're going to go over a product data scientist mock interview at doordash the video itself is going to be about 10 minutes and i'm going to drop in some commentaries here and there to provide some suggestions in terms of what the candidate did well and what are some areas of improvement i really hope that you find this video

0:26 helpful for your upcoming mock interview hello hi so this is the doordash mock interview thank you for attending this call what we're gonna do is we're gonna spend about 30-40 minutes going through some product data science questions and the first five to ten minutes are going to be focused on statistics questions and then we'll spend the remaining 40 30 40 minutes just going through some case study questions covering a b testing analytics does that sound good

1:00 sounds quick okay awesome all right so what i'm going to do is i'm going to paste questions one at a time on the quarter pad and feel free to use it as your as a scratch sheet you know jot down your ideas and so forth and along the way if you have any questions feel free to ask okay okay cool here's a quick helpful tip for candidates who have upcoming interviews keep in mind that a lot of the interviews during

1:28 coven are virtual this means that you'll gain access to quarter pad or word document and when you're in a situation where you have to solve a case problem it's recommended that you actually flesh out the ideas by writing down your points rather than just overly explaining it this is going to be helpful in two fronts one it's going to be helpful for you to brainstorm ideas and two it helps the interviewer see your thought process

1:59 so when you have an upcoming interview please use your use decoder pad and word document as a place to actually drop down your ideas all right so let's start with the first status questions what does 90 statistical power mean so so it's 90 statistical power so statistical power is basically linked to the type one type of two errors which means what is the probability of finding a right observation when it

2:32 when it doesn't exist or finding an observation or or saying that we found enough we did not found an observation when it exists so 90 statistical power means the the probability of finding an observation which of not finding an observation which exists is 90 notice how the interview started the interviewer asked a fundamental question in statistics this is actually pretty typical in phone

3:04 screening and on-site rounds an interviewer will kind of pepper various questions on the fundamentals of statistics and machine learning theory in order to assess whether the candidate grasps a basic knowledge in data science and machine learning engineering so let's think about the question itself the question is what is what does 90 power mean now ninety percent power is the

3:35 probability of rejecting the null hypothesis given that the alternative is true this is actually not the same as what the candidate said so what the candidate said was it is the probability of not finding an observation when it exists in slightly off way i think what the candidate was conveying was that it's a type 2 error which is actually not the same as power so in this case the candidate's response

4:05 was incorrect now as as you're preparing for your interview you want to make sure that you brush up on these fundamental concepts because there's no room for interpretation when you're asked questions like these it's either you're right or wrong you're wrong so what i would recommend is that you create an index card and you type you basically put down the term and the definition in the back and you go through these index cards

4:36 as a way to practice explaining these basic concepts in statistics and machine learning theory okay all right next question how do you determine the practical significance in a in an experiment so when it depends on the type of experiment we run and the kind of metric we are looking at so matrix can be like mean like mean is a metric or

5:07 a rate can be a vector and metric can also be proportions right so it can be percentage of users that click on a particular link or things like that so in these both of these cases we do some kind of statistical tests like a two sided t test or a test of proportions and practical significant and the significance of those is determined by p value which in general it's if it's below 0.05 we say it's

5:37 significant now that is just the statistical part if you wanna look at the practical part we would also look have to look at the business metrics for example the kind of lift that we have seen and that will depend on the cost benefit analysis so let's say a five percent increase in ltv of a consumer on a particular app would be profitable for us and that would be the break-even point and anything above that would be useful whereas anything below

6:07 that would not be useful so if so if we look at that basically what we are seeing what we are saying is even if the results are significant statistically only if they make business sense then it would be practically significant and then there is one more aspect which is movement of multiple metrics so let's say we do an experimentation on one particular metric but and that comes out to be significant but the other metrics

6:37 move along with it in the experiment and that can be like detrimental to our experiment so we will have to use that as well in our stating the practical significance of the experiment let's break down the question number two in regards to the practical significance in an experimentation what the candidate mentioned are two things first the candidate mentioned the cost-benefit analysis of launching a product and secondly the candidate mentioned on multiple metrics

7:07 so looking at let's just say the primary metric and guard rail nature now in this case the candidate has a right understanding about the practical significance of an experimentation in an experimentation typically there's a socialization that happens upfront with the business stakeholder to determine what is a meaningful experimental result so just because an experimentation achieved successful significance so p-value is

7:38 less than alpha it doesn't mean that the treatment itself should replace the controller because it might not provide a necessary return on investment to actually launch that new future for instance the cost of actually building out the future for the wider audience to use might be more costly than the benefit or the or the additional revenue that you would gain by launching that in future and so the candidate here

8:09 had the right concept about the practical significance of an experimentation okay so how would you handle it's such a case then if one of the metric is practically significant but the other one isn't there can be a couple of ways to handle it one can be changing the design of experiment so making the experiment you know a bit narrowed down in terms of the users that are going into the

8:39 experiment that can make sure that we that we only move one metric and not the other metric and then we can also look at correlation or causation of the true matrix so if one metric is correlated to the other or one metric is a result of a causal inference from the other we can we would we should assume that the metric would move but if that is not the case then maybe we can you know put some controls in place and

9:08 then run the experiment again to get a better significance okay got it all right i'll ask the next question now how do you calculate type 1 and type your errors of your statistical tests okay so one and type two errors can be calculated by use english can calculate it by

9:42 i mean by by using like you know i'm not really sure about this so probably by let's see do you want to start by explaining what type one type of error error rates are sure a time for an error is basically nothing but a false positive so so it happens when the experiment incorrectly rejects

10:13 the true null hypothesis and you know in which actually it should not so then a type one error occurs a type two error on the other hand it's a false negative it happens when an experiment does not reject the null hypothesis but it's but you know but the result is actually false so it should reject the hypothesis okay got it so in this case what would be your type 1 error rate so type 1 error rate

10:46 you know in this case would be you know so basically let's say we have a null hypothesis that you know there is no significant change in the click-through rate of two types of users so there will be a type one error where you know this hypothesis will be rejected and and we find the results that there is a significant difference between the two you

11:17 know the the two sides and one way of controlling that i can think of is using the power so power is nothing but statistical power is nothing but beta and that's the probability of making like say a type two error so power of of the set of three tests is one minus beta so you know we can adjust for that value okay so i'm going to interject there because i think i want to kind of give

11:46 you some hint here so let's just suppose that alpha is 0.05 and your beta is let's just let's just suppose that your power is 0.8 then in that case how would you actually calculate your type 1 error type 1 and type tap to error rates sure so the problem so you mentioned b beta is 20 power power is point eight power and your alpha is point zero five

12:18 okay so if power is point eight it means a power is one minus beta so your beta would be 0.2 so the probability of making a type 2 error would be 0.2 and alpha you mentioned is 0.05 right so alpha is 0.05 and i think that's the probability of committing a type one error so it's like a five percent probability of committing a type one okay okay awesome notice how the candidate struggled a little bit in the

12:49 beginning eventually the candidate recovered because the interviewer gave the candidate a couple hits now in order to calculate the type 1 error rate type 2 error rate let's just think about what the definitions of those two are type 1 error rate is a probability of rejecting the null hypothesis when it is true type of errate is the rejecting of not rejecting the null hypothesis when the alternative is true

13:19 how do you actually calculate tempo and error rate and type 2 error rate so type 1 error rate is essentially your alpha so whatever your alpha you set so typically in successful testing it still be 0.05 sorry i meant to say at least in product experimentation it's typically 0.05 and so that is actually going to be the type on error rate now for tattoo air rate it is actually the complement of power so suppose that power is eighty percent

13:51 then in that case peptide rate is going to be the complement of that which is going to be 20 i hope you enjoyed a sample of this mock interview with some commentaries this is just a sample of what's more to come in the near future in the near future i'll publish more videos like this that will cover rounds focused on statistics sql statistical coding machine learning engineering and product sense if you like this video please subscribe

14:23 like and share with others and that's going to encourage me to continue to provide more videos like this thank you

Summary

The video features a mock interview for a product data scientist position at DoorDash, led by Dan Lee, a data scientist interview coach. Throughout the session, Lee provides commentary on the candidate's responses, highlighting both strengths and areas for improvement, particularly in fundamental statistics and practical significance in experiments.

- The interview consists of statistics questions followed by case study questions on A/B testing and analytics.
- Candidates are encouraged to write down their thought processes during virtual interviews to aid brainstorming and clarity.
- A key concept discussed is the meaning of 90% statistical power, which the candidate initially misinterprets.
- Practical significance in experiments is determined not just by statistical significance but also by business metrics and cost-benefit analyses.
- The candidate demonstrates an understanding of type 1 and type 2 errors but struggles with calculating their rates initially.
- Lee emphasizes the importance of mastering fundamental concepts in statistics and machine learning for successful interviews.
- The session aims to prepare candidates for real interview scenarios by providing insights into common questions and expected responses.

Questions Answered

What can candidates expect from this mock interview?

Candidates will go through product data science questions, focusing on statistics and case studies, with commentary on their performance.

What does 90% power mean in statistics?

90% power refers to the probability of rejecting the null hypothesis when the alternative hypothesis is true, which differs from the candidate's incorrect definition related to type 2 error.

How do you determine practical significance in an experiment?

Practical significance depends on the cost-benefit analysis and whether the results make business sense, not just statistical significance.

How do you calculate type 1 and type 2 errors in statistical tests?

Type 1 error is the probability of rejecting a true null hypothesis (false positive), while type 2 error is failing to reject a false null hypothesis (false negative).

How do you calculate type 1 and type 2 error rates given alpha and power?

Type 1 error rate is equal to alpha, while type 2 error rate is the complement of power, which can be calculated based on the definitions of these errors.

© transcribe · For agents Built with care and craft by Gokul Rajaram