Think Like a Data Analyst: Business Logic to Insights

Start with Why to Define Business Objectives

Begin with a specific business requirement (e.g., increase user rides in a location), then ask why you need that outcome. This creates a chain of reasoning that connects the surface requirement to deeper business metrics and helps you identify the right KPIs to track.

Uber Case Study: Multiple Paths to One Goal

Increasing rides can be achieved through multiple interconnected metrics: decreasing empty miles (unpaid driver travel), increasing driver acceptance rates, reducing driver cancellation rates, and lowering driver churn. Each path reinforces the others and creates a holistic business strategy.

Four interconnected metrics to increase user rides

Acceptance Rate Complexity

Driver acceptance rate is difficult to optimize because it depends on subjective factors like driver preferences (city-only vs. long-distance trips, airport routes) and rider ratings. Unlike other metrics, it cannot be directly controlled by the platform.

Step 1: Define Business Goals and KPIs

Start by clarifying what you want to affect (user experience, loyalty, churn reduction) and translate this into realistic, actionable business questions. Set a baseline with important variables and analysis types you will use, maintaining flexibility to notice hidden patterns.

Step 2: Select Data, Build Pipelines, Clean and Wrangle

Identify which data tables address your business goals, merge tables appropriately, construct SQL pipelines and ETL processes, then handle duplicates, nulls, missing values, outliers, and formatting issues. This step consumes significant time but is crucial for analysis quality.

Data preparation workflow

Step 3: Exploratory Data Analysis (EDA) and Select Analysis Type

Perform EDA to understand data distributions, correlations, and statistical properties (mean, median, mode, skewness). Then select the appropriate analysis type (descriptive, exploratory, predictive, inferential, or causal) based on your business goals.

EDA components before model building

Step 4: Set Baseline and Validate Results

Establish a business baseline (e.g., historical churn rates from past quarters) to ensure analysis outputs are realistic. Compare model outputs against this baseline; if a model predicts 30% churn when historical range is 10-14%, investigate why. Validate results with colleagues, historical data, or statistical methods like SHAP.

Baseline validation prevents false conclusions

Step 5: Articulate Business Decisions

Communicate findings through storytelling, dashboards (Tableau), recommendations, or presentations (Google Slides). Build a narrative that highlights main features of your analysis and connects insights to actionable business decisions.

Descriptive Analysis: Summarize What Happened

Descriptive analysis summarizes data without interpretation. If you only report summarized data without deeper analysis, it remains descriptive. This is the foundation but does not answer why or predict future outcomes.

Exploratory Analysis: Discover Patterns

Exploratory analysis identifies patterns and discoveries in data but does not quantify whether findings will hold in new samples. It answers what patterns exist but lacks statistical rigor to generalize beyond the current dataset.

Predictive Analysis: Forecast Individual Outcomes

Predictive analysis uses variables (education, age, experience, gender, location) to forecast a target response (e.g., data science salary). It shows correlation and prediction capability but does not explain causation or why the prediction works.

Inferential Analysis: Generalize from Sample to Population

Inferential analysis draws conclusions about a population from a sample using statistical techniques like t-tests, ANOVA, Bayesian inference, and hypothesis testing. It answers what the data tells us about the broader population.

Causal Analysis: Understand Cause and Effect

Causal analysis identifies underlying cause-and-effect relationships between variables. However, using causal analysis alone is intuitive without solid statistical methods. Confounding variables (e.g., summer causing both ice cream sales and shark bites) can create false causation.

Causal Inference: The Gold Standard

Causal inference combines causal and inferential techniques using statistical methods to rigorously identify cause-and-effect relationships. Prominent methods include randomized controlled trials (RCTs) and difference-in-differences analysis. It addresses confounding bias and counterfactual scenarios.

Causal inference methodology

Confounding Variables: The Hidden Bias

A confounding variable is a third variable that influences both the independent and dependent variables, creating false causation. Example: smoking (confounder) causes both lighter carrying and lung cancer; omitting smoking suggests lighters cause cancer. Confounding tricks us into seeing causation where only correlation exists.

Confounding variable example: lighter, smoking, lung cancer

Counterfactual: The What-If Scenario

A counterfactual asks what would have happened in an alternative reality. Example: if John took medicine and recovered, the counterfactual is what would have happened if he had not taken it. Since we cannot measure two realities simultaneously, counterfactuals are challenging to estimate but essential for causal claims.

Correlation vs. Causation

Correlation does not imply causation. Two variables can move together without one causing the other. This cliche phrase describes a common analytical error where confounding variables create the illusion of causation when only correlation exists.