❔How do you handle missing or corrupted data in a dataset?
✅ Answer:
I approach this by first investigating the nature of the missingness. I categorize it into Missing Completely at Random (MCAR), Missing at Random (MAR), or Not at Random (MNAR), as the reason why data is missing dictates the solution.
In the short term, I apply immediate cleaning techniques. If the missingness is negligible, I might use listwise deletion. If the data is critical, I use imputation methods, simple ones like mean/median for numerical data, or more advanced ones like K-Nearest Neighbors (KNN) or MICE for preserving the relationship between variables.
Long term, I focus on root cause analysis. I work with data engineers to identify if the corruption is happening at the source (e.g., a broken sensor or a UI bug in a form). My goal is to implement validation checks at the data entry level to ensure the pipeline remains clean and reliable for future analysis.
❔How do you distinguish between Correlation and Causation when analyzing a trend?
✅ Answer:
I start by acknowledging that correlation is only a mathematical relationship, while causation requires a functional mechanism. My first step is to use visualizations (like scatter plots) and coefficients (like Pearson’s) to confirm if a relationship actually exists.
Next, I look for confounding variables and "Spurious Correlations." I use techniques like partial correlation or stratified analysis to see if a third factor is influencing both variables. For example, ice cream sales and shark attacks both rise in summer, but the "cause" is the temperature, not the ice cream.
To truly prove causation, I would ideally advocate for a controlled experiment (A/B Testing). If an experiment isn't possible, I use quasi-experimental designs or "Causal Inference" models like Difference-in-Differences or Propensity Score Matching to estimate the causal impact using historical data.
❔Walk me through your process for Exploratory Data Analysis (EDA) on a new dataset.
✅ Answer:
I begin with a structural audit. I check the shape of the data, data types, and the presence of duplicates. I generate summary statistics to get a sense of the "center" and "spread" of the data, which helps me spot obvious anomalies right away.
Next, I move to Univariate and Bivariate analysis. I use histograms to check for skewness and box plots to identify outliers. I then use heatmaps and correlation matrices to see how variables interact with each other. This is where I start formulating hypotheses about which features might be the strongest drivers of our KPIs.
Finally, I tie everything back to the Business Context. EDA isn't just about math; it’s about finding a story. I look for segments or trends that contradict our current business assumptions and prepare a summary that translates these technical observations into actionable questions for the product or marketing teams.
❔A metric you are tracking suddenly drops by 20%. How do you investigate the cause?
✅ Answer:
I follow a "drilling down" methodology. First, I perform a Technical and Seasonal check. I verify with the engineering team if there were any deployment changes or tracking pixel failures. I also compare the drop against historical patterns is this a typical weekend dip or a holiday-related trend?
Next, I segment the data. I break the 20% drop down by dimensions like geography, device type, browser, and user acquisition channel. If the drop is only happening on Android, it’s likely a technical bug. If it’s happening across all segments, it’s more likely a broader market shift or a competitor’s move.
Finally, I look at the User Journey. I analyze the conversion funnel to see exactly where the drop-off is happening. Is it at the "Add to Cart" stage or the "Payment" stage? Once the bottleneck is identified, I present the data-backed reason for the decline along with a proposed fix or a further test to mitigate the loss.