Exploratory Data Analysis and Data Quality Questions

Understanding and preparing an unfamiliar dataset before analysis or modeling. Covers systematic profiling through summary statistics, distribution and outlier inspection, and relationships between variables to form initial hypotheses, alongside turning raw data into a trustworthy base: handling missing values, deduplication, outlier treatment, type and consistency checks, and validation. Includes critical thinking about sampling and measurement bias and what a dataset can and cannot support.

HardTechnical

74 practiced

For a heavy-tailed metric (think financial transaction sizes), what robust descriptive statistics would you reach for beyond mean/variance -- trimmed mean, winsorized mean, median absolute deviation -- and what does each protect you against that the standard versions don't?

Sample Answer

HardTechnical

74 practiced

During EDA you find a strong correlation between a protected attribute (or a close proxy like ZIP code) and your target variable. What would you do next: how do you document the finding, what fairness checks would you run before anyone builds on this data, and who do you loop in?

Sample Answer

MediumTechnical

112 practiced

A KPI like 'average order value' can mislead when computed as a naive average across a skewed or unevenly-weighted population. How would you demonstrate this during EDA, and how would you explain to executives why the current number is misleading?

Sample Answer

HardTechnical

66 practiced

In practice, how do you actually tell MCAR from MAR rather than just defining them? Walk through a concrete diagnostic (for example, Little's MCAR test, or regressing a missingness indicator on the observed covariates), and be honest about its limitations on a real operational dataset.

Sample Answer

HardTechnical

68 practiced

A customer-lifetime-value analysis is skewed because older purchase histories are truncated (you only see what's left in the system). How would you quantify the bias this truncation introduces during EDA, and what would you flag as an assumption before anyone treats the resulting estimate as reliable?

Sample Answer

Direct answer Quantify truncation bias by comparing what you can see (the system's stored, truncated history) against any evidence of what's actually missing, and flag the resulting estimate as a likely UNDERSTATEMENT before anyone treats it as reliable. Older customers, whose early purchase history has fallen outside whatever retention window the system stores, will show a systematically lower observed lifetime value than their true value, purely because more of their history has aged out. Why this specifically biases the estimate If the system only retains, say, the last two years of purchase history, a customer with ten years of tenure has 80% of their true history invisible to any analysis running on the stored data, while a customer with one year of tenure has essentially all of their history visible. Averaging "observed lifetime value" across customers with very different tenures conflates two very different amounts of truncation, systematically understating true value more severely for your longest-tenured, and often most valuable, customers. Worked example Simulating this directly: assume a true, fully-observed average customer lifetime value of about $102, but a stored system that only retains roughly the last two years of history relative to each customer's total tenure. The resulting observed average, computed only from what's actually stored, comes out to about $51, roughly a 50% understatement of the true figure, driven almost entirely by the longest-tenured customers whose visible history represents the smallest fraction of their true lifetime. Trade-offs and pitfalls Before treating any truncation-corrected estimate as reliable, be explicit about the assumption behind the correction (in this case, an assumed relationship between tenure and how much history is missing) and flag it as an assumption, not a fact, since a wrong assumption about the truncation pattern can make a "corrected" number just as misleading as the uncorrected one, only with more false confidence attached.

Unlock Full Question Bank

Get access to all Exploratory Data Analysis and Data Quality interview questions and detailed answers.

Join thousands of developers preparing for their dream job.

Statistic	What it replaces	What happens to extreme values
Trimmed mean	the plain mean	dropped entirely before averaging
Winsorized mean	the plain mean	kept, but capped at a threshold
MAD	standard deviation / variance	barely influence it at all

Exploratory Data Analysis and Data Quality Questions

Sample Answer

Direct answer

What each protects against, specifically

Worked example

Trade-offs and pitfalls

Sample Answer

Direct answer

Why this needs more than a technical fix

Worked example

Trade-offs and pitfalls

Sample Answer

Direct answer

Constructing the weighted alternative

Worked example

Trade-offs and pitfalls

Sample Answer

Direct answer

Honest limitations

Worked example

Trade-offs and pitfalls

Sample Answer

Direct answer

Why this specifically biases the estimate

Worked example

Trade-offs and pitfalls

Unlock Full Question Bank