Source Data
Sourcing data that can bear the question
Once the question is set, you identify sources and judge fitness for purpose. Know the vocabulary: primary vs secondary data, internal vs external, structured vs unstructured, and the acquisition route (extract, API, survey, purchase). A data dictionary and lineage record - where each field came from and how it was transformed - are deliverables, not paperwork.
Fitness is judged on quality dimensions: accuracy, completeness, consistency, timeliness, validity and uniqueness. Data profiling (row counts, null rates, duplicate rates, value ranges, date coverage) is how you test those dimensions before analysis, not after.
Traps. Assuming a system of record is correct because it is official. Using only data that is easy to get, which silently biases the sample (survivorship and selection bias). Confusing volume with quality. Ignoring privacy, consent and regulatory constraints until late. Forgetting that sampling method - random, stratified, convenience - determines what you may later generalize.
Rule of thumb - profile before you promise. Spend the first pass on a profiling report: expected vs actual row counts, null and duplicate percentages, min/max dates, and a list of who is missing from the data. If a group is absent, note it as a limitation now.
Worked example. A CRM extract returns 40,000 customers but finance reports 52,000 active accounts. Rather than proceeding, you reconcile: 9,000 are duplicates from a legacy merge and 3,000 have no email address. You document the gap, define the analytic population as the 40,000 de-duplicated records, and flag that email-based findings cannot speak for the 3,000 without addresses.