Diagnose sales correlations without claiming causality

You support a sales team and are asked to find "which outreach channels correlate with higher deal win rate" without building predictive models. You have deal-level data: deal_id, account_id, rep_id, created_at, closed_at, is_won, amount_usd, product_line, region; and touch-level data: account_id, rep_id, touch_date, channel (email/call/demo/webinar), is_primary_contact. Design a correlation-focused analysis that is decision-ready but avoids causal claims: (a) Define a defensible exposure window (e.g., touches within the first 14 days after created_at) and justify how you’ll handle right-censoring for open deals and late touches; (b) Specify stratifications and/or matching you’ll use (e.g., region, segment, deal size buckets, rep tenure) to control confounding without modeling; (c) Show exactly how you’d compute within-rep, within-segment correlations to avoid between-rep composition bias (outline de-meaning or fixed-effects style differencing before correlating); (d) List bias risks (reverse causality when hot deals drive more touches, missing-not-at-random touches on lost deals, seasonality) and propose sensitivity checks (pre-registration of windows, placebo windows before deal creation, leave-one-rep-out analysis, randomization inference) to see if correlations are robust; (e) Describe two plots you’d present that can reveal Simpson’s paradox across regions or segments and how you’d detect and communicate it; (f) Write the exact decision guardrails you’ll present to sales leadership to prevent causal overreach and how you’d phrase them.