Imagine a report about the Yankees and Red Sox. Inventing match results is fabrication. Changing recorded losses into wins is alteration. Quietly deleting valid losses while presenting the remainder as the full record misrepresents the evidence even if every retained row is genuine.

Data cleaning can legitimately remove duplicates or repair errors. The difference is a defensible rule, a traceable record, and disclosure—not whether the cleaned result looks favorable. The American Statistical Association calls for transparent methods and resistance to pressure to predetermine or selectively interpret results.[2]

Ask for the chain from original records to reported totals.A polished dashboard cannot replace the underlying evidence.

At a 40% baseline, the visible lengths represent 13.1 and 2.4 percentage points. The first is about 5.46 times the second. At zero, the lengths represent 53.1 and 42.4, a ratio of about 1.25. Both panels retain correct labels; only one preserves the ratio of the reported rates.

That is why the UK Government Analysis Function recommends a zero baseline for bars. If the purpose is to inspect a small difference, use an appropriate alternative and make its scale clear.[4]

The Yankees’ displayed rate is 10.7 percentage points higher. Relative to the Red Sox rate, it is about 25.2% higher. These are compatible descriptions. Calling the difference “25.2 percentage points” is incorrect; calling it “25.2% more wins” would require comparable game counts that the image never supplies.

Reverse the comparison and the reference changes: the Red Sox rate is about 20.2% lower than the Yankees rate, because you now divide the gap by 53.1. “Higher” and “lower” percentages need not be symmetric. Always name the reference.

04 · Cherry-picking the period

Suppose you publish only a team’s strongest stretch and describe it as typical performance. Those wins may be real; the selection hides the stretches that would change the impression. Ask for the full relevant period and whether the start and end dates were chosen before the results were examined.

05 · Sampling and survivorship bias

A fan poll about which team is better does not measure a win rate. If you recruit only from one team’s fan page, it may not even represent baseball fans generally. A large response count does not by itself establish representative recruitment. Survey transparency requires information about the population, sampling, recruitment, questions, and weighting.[6]

Survivorship bias is a related selection problem: reviewing only clubs that reached the playoffs omits those that did not. You cannot explain why a strategy succeeds across all clubs by studying only its survivors. This is a possible failure mode, not a finding about the supplied image.

06 · Moving the denominator or definition

A win rate needs both wins and eligible games. Excluding unfavorable games, mixing home-only results with all-game results, or quietly redefining an “eligible game” changes the measure. In administrative reporting, the same problem appears when a success rate improves because difficult cases disappear from eligibility. Ask for numerator, denominator, exclusions, and definition changes together.

“53.1% of what, over which dates, with which exclusions?”If the report cannot answer, its precision is ahead of its documentation.

If you extend the team report to runs scored per game, a few very high-scoring games can pull up the mean. A median answers a different question about the middle of the distribution. Neither is universally the honest choice; disclose the measure and show enough spread to support the interpretation.[7]

Pooling win rates creates another trap. The combined rate is total wins divided by total eligible games. Taking a simple average of subgroup percentages works only when their denominators are equal. Home and away rates must be weighted by the corresponding game counts.

With different subgroup weights, an overall comparison can even reverse the comparisons within every subgroup: Simpson’s paradox. You need the subgroup counts to check for it; the two image values cannot establish that it happened here. And disaggregating is not automatically the correct causal analysis—the relevant grouping depends on the question and how the data arose.[8]

You cannot calculate a justified confidence interval or significance test from these two percentages alone. You would need the game counts, the underlying records, and a defensible model. Independence cannot simply be assumed: teams may play each other and face different schedules.

If the records cover every game in a defined season, the rate describes that season without sampling uncertainty about the recorded total. Predicting future performance introduces a different source of uncertainty. Measurement mistakes, missing records, and model assumptions can matter in either case.

Even a valid small p-value would not tell you the probability that the favored team is “truly better,” or whether its advantage matters in practice. Statistical significance does not measure effect size or importance.[9]

“The new coach caused the higher win rate” adds a claim the bars do not test. Rosters, injuries, opponents, and scheduling could also differ. A before-and-after comparison would still need to address other changes occurring at the same time.

A causal question asks what would have happened under an alternative decision. Random assignment can help create a credible comparison when feasible; observational analysis needs explicit assumptions and a suitable design. Merely adding more variables does not guarantee that those assumptions hold.[10]

Separate “the reported rates differ” from “this decision caused the difference.”The second statement requires additional evidence.

Suppose you try many date windows, opponent groups, and performance measures, then report only the comparison with a small p-value. The reader sees one apparently decisive result and never sees how many opportunities you gave chance to produce it. This is the mechanism behind p-hacking and selective reporting.[9]

Exploration is useful. The misleading step is presenting a discovery from that search as if it were a single prediction specified beforehand. Report what you tried, distinguish exploratory from confirmatory claims, and use an appropriate correction or fresh evidence to evaluate selected findings.