Where the numbers come from decides what you are allowed to conclude from them.
You can almost never measure everyone. So you measure a sample, describe it with statistics, and use those to estimate the parameters of the whole population. Every later chapter is about how far that estimate can be trusted.
Everyone you want to know about. All patients admitted with sepsis in the US this year. Its size is N.
The part you actually measure. 200 sepsis charts pulled from three hospitals. Its size is n.
A number that describes the population: mean μ, SD σ, proportion p. Usually unknown.
A number that describes the sample: mean x̄, SD s, proportion p̂. You compute it.
Counted. Whole numbers with gaps. Number of falls this month, number of meds on the MAR, number of children, number of patients on the unit.
Ask: could it be 2.5? If not, it is discrete.
Measured. Any value in a range. Temperature 98.63 °F, weight 72.4 kg, time to first dose 17.5 min, hemoglobin 11.8 g/dL.
The only limit is how precise your instrument is.
| Level | What you can do | Nursing example | Trap |
|---|---|---|---|
| Nominal | name / categorize only | blood type, unit, diagnosis code, sex | no order — O is not “less than” AB |
| Ordinal | rank in order | pain 0–10, Braden score, cancer stage, triage level | the gaps are not equal: pain 8 is not twice pain 4 |
| Interval | order + equal gaps, no true zero | temperature in °F or °C, calendar year | 0 °C is not “no heat”, so ratios are meaningless |
| Ratio | everything, true zero | weight, blood pressure, urine output, heart rate, dose | 0 mL really means none — 200 mL is twice 100 mL |
| Method | How | Example |
|---|---|---|
| Simple random | every set of n has the same chance (random number generator on the whole list) | pick 50 MRNs at random from all 900 admissions |
| Stratified | split into groups that matter, sample within each | 10 from each age band so every band is represented |
| Cluster | split into natural groups, pick whole groups | randomly pick 3 units and survey everyone on them |
| Systematic | every k-th person from a random start | every 10th chart in the file |
| Convenience | whoever is nearby | the patients you happen to have today — biased |
Researchers watch and record; nobody is assigned anything. Chart review of pressure-injury rates by turning schedule. Shows association only.
Retrospective looks back at existing records. Prospective follows people forward. Cross-sectional is one snapshot in time.
Researchers assign the treatment. Randomize 200 post-op patients to ice packs vs no ice packs and compare pain scores. The only design that can show cause.
Explanatory variable = the treatment. Response variable = the outcome you measure.
Coin-flip assignment spreads unknown differences evenly between groups, so the groups start alike.
Single-blind: subjects do not know their group. Double-blind: neither do the people measuring them. A placebo (sugar pill, saline) hides who got the real thing.
The sample was not picked the way the population looks. Surveying only day-shift nurses about fatigue.
The people who answered differ from those who did not. Only the angriest patients return the satisfaction survey.
People answer wrong on purpose or by accident. Patients under-report drinks per week.
A hidden third factor drives both things. Coffee drinkers have more heart disease — because more of them smoke.
Peer-reviewed journal > professional guideline (CDC, AHA) > textbook > news article > blog. Ask: who collected the data, how big was the sample, was it random, who paid for it, and is the full method published? A survey on a company’s own website that sells the product is not evidence.