Probability, Hypothesis Testing & Statistical Tests
On this page
🎯 Learning Outcomes
By the end of this session, you should be able to:
- Explain the purpose of hypothesis testing in health research and nursing practice.
- Write clear null (H₀) and alternative (H₁) hypotheses for simple clinical questions.
- Use p-values and significance levels correctly without overclaiming.
- Select chi-square, paired t-test, or independent t-test appropriately based on data structure.
- Interpret test results in clear nursing and public health language — not just statistical jargon.
Why Hypothesis Testing Matters for Nurses
In real clinical and community practice, we rarely measure everybody. We take a sample a group of patients, a ward, a village and use that sample to make a careful, evidence-based judgement about the wider population.
The Inference Pathway
Sample Result ➔ Statistical Test ➔ Evidence-Based Conclusion
Real Nursing Examples
- Hand Hygiene: Does a 2-hour hand hygiene training session improve compliance rates among student nurses on the maternity ward?
- Malaria Prevention: Is malaria positivity different between children who sleep under an insecticide-treated net and those who do not?
- Patient Education: Do mothers' knowledge scores about ORS preparation increase after a health education session?
- Clinic Efficiency: Is the average waiting time at Clinic A significantly shorter than at Clinic B?
- Nutrition: Is there an association between a child's sex and their nutritional status (wasted vs. not wasted)?
Statistical Inference:
Statistical inference is the bridge between what we observe in our sample and what we believe is true in the entire population.
| Component | What It Is | Nursing Example |
|---|---|---|
| KNOWN Sample observed data | The actual numbers we collected from our study group. | We measured Hb levels in 120 pregnant mothers attending ANC at Hospital X. |
| METHOD Probability and tests | The mathematical rules (t-tests, chi-square, p-values) that judge whether our sample result is meaningful. | We run a statistical test to compare the mean Hb before and after iron supplementation. |
| UNKNOWN Population truth | The real situation in the entire population that we can never measure perfectly. | The true mean Hb among ALL pregnant mothers in the district, not just our 120. |
💡 Core Question of Inference: "Is the observed difference likely to be real, or could it be due to chance?"
A test does not prove truth; it measures how compatible the data are with the assumption of "no effect" (H₀). We are essentially asking: "If nothing were really happening, how surprised would I be to see these results?"
Key Terms You Must Know
| Term | Simple Meaning | Nursing Example |
|---|---|---|
| Parameter | The true value in the entire population. Usually unknown. | The true mean haemoglobin (Hb) level among all pregnant mothers in Uganda. |
| Statistic | A value calculated from your sample data. This is what you actually know. | The mean Hb in the 120 mothers you actually tested at your clinic. |
| Sampling Error | The natural difference between a sample statistic and the true population parameter. It exists simply because we did not measure everyone. | Your sample mean Hb is 11.2 g/dL, but the true population mean might be 11.5 g/dL. That gap is sampling error. |
| Statistical Test | A formal mathematical rule for judging whether sample evidence is strong enough to support a conclusion. | Chi-square test, paired t-test, independent t-test, ANOVA. |
| Hypothesis | A clear, testable statement about a population that can be supported or refuted by data. | "Mothers who receive nutrition counselling have higher mean Hb than those who do not." |
Mnemonic: "Parameter = Population" and "Statistic = Sample". Both start with the same letter. If you know the whole population, it's a parameter. If you only have a sample, it's a statistic.
From Research Question to Hypothesis
Every statistical test begins with a question. But a research question is usually written in plain English. To test it statistically, we must translate it into two formal statements: the null hypothesis (H₀) and the alternative hypothesis (H₁).
| Research Question | Null Hypothesis (H₀) | Alternative Hypothesis (H₁) |
|---|---|---|
| Does training improve knowledge scores among nurses? | Mean score before training = Mean score after training. (No difference) | Mean score before training ≠ Mean score after training. (There is a difference) |
| Is malaria positivity related to net use? | There is no association between net use and malaria positivity. | There is an association between net use and malaria positivity. |
| Do two clinics have different waiting times? | Mean waiting time in Clinic A = Mean waiting time in Clinic B. | Mean waiting time in Clinic A ≠ Mean waiting time in Clinic B. |
| Does iron supplementation reduce anaemia? | Mean Hb after supplementation = Mean Hb before supplementation. | Mean Hb after supplementation > Mean Hb before supplementation. |
| Is there a link between ward type and infection rate? | Ward type and infection rate are independent (no association). | Ward type and infection rate are associated. |
❌ Common Student Error: Writing H₀ as "The training will improve scores." That is actually H₁! H₀ must always be the "no effect / no difference / no association" statement. Think of H₀ as the skeptic's position — it assumes nothing interesting is happening until proven otherwise.
Null and Alternative Hypotheses in Depth
H₀ — The Null Hypothesis
- It is the "no difference" or "no association" statement.
- It represents the status quo — what we would expect if chance alone were operating.
- We test H₀ because random variation can easily mislead us, especially with small samples.
- We never "prove" H₀ is true. We either reject it or fail to reject it.
H₁ — The Alternative Hypothesis
- It is the statement the researcher thinks may be supported by evidence.
- It represents the new idea, the intervention effect, or the association we suspect.
- It is what we hope to find evidence for by finding evidence against H₀.
Think of H₀ as "innocent until proven guilty." The null hypothesis claims "nothing happened." Our data is the evidence. If the evidence is strong enough, we reject the claim of innocence (reject H₀). If the evidence is weak, we do not declare innocence — we simply say "not guilty enough to convict" (fail to reject H₀).
A Statistically Significant Result
When we say a result is "statistically significant," we mean: "The data we observed would be very unlikely if H₀ were actually true." Therefore, we have reasonable grounds to doubt H₀ and consider H₁ instead.
✅ Correct: "We reject H₀" or "We fail to reject H₀."
❌ Incorrect: "We accept H₀" or "H₀ is true." — The test only tells us whether evidence is strong enough to discard H₀, not whether H₀ is definitively true.
One-Tailed vs. Two-Tailed Tests
| Type | Question It Answers | When to Use |
|---|---|---|
| Two-tailed | "Is there any difference?" (Could be higher OR lower) | Most classroom and research situations. Use this as your default. It tests for a difference in either direction. |
| One-tailed | "Is one group specifically higher or lower?" | Only when the direction is justified before analysis based on strong theory or prior evidence. E.g., "We expect training to increase scores, not decrease them." |
🛡️ Safe Rule for Students: Use a two-tailed test unless the study clearly planned a directional test before seeing the data. Never choose a one-tailed test just because it makes your p-value smaller — that is scientific misconduct.
Example: A new drug for hypertension might lower BP, raise BP, or do nothing. Since we are not sure of the direction, we use a two-tailed test. But if we are testing whether a proven anti-hypertensive drug reduces BP (and we have no reason to think it would raise it), a one-tailed test might be justified.
Significance Level: α (Alpha)
Alpha (α) is the cut-off point for deciding whether our evidence is strong enough to be called "statistically significant." It is the threshold of "surprise" — how unlikely must our results be under H₀ before we reject H₀?
The Convention
α = 0.05
This means we are willing to accept a 5% chance of wrongly rejecting H₀ when it is actually true.
Decision Rule
- If p ≤ 0.05: Reject H₀. The result is statistically significant. The evidence is strong enough to support H₁.
- If p > 0.05: Fail to reject H₀. The evidence is not strong enough. We do not have sufficient proof to support H₁.
⚠️ Critical Warning: α = 0.05 is a convention, not a law of nature. It is an arbitrary threshold agreed upon by scientists. A result with p = 0.051 is not magically "false," and p = 0.049 is not magically "true." The difference is tiny. Always interpret p-values as a continuum of evidence, not a binary switch.
What a p-value Actually Means
A p-value is the probability of observing results as extreme as (or more extreme than) our sample result, assuming H₀ is true.
In plain nursing language: "If the intervention really did nothing, how lucky (or unlucky) would I have to be to see this result just by chance?"
| p-value | Statistical Interpretation | Nursing Language |
|---|---|---|
| p = 0.40 | Observed result is quite compatible with H₀. | "The difference we saw could easily happen by chance. No strong evidence of a real effect." |
| p = 0.06 | Evidence is suggestive, but not significant at α = 0.05. | "There is a hint of an effect, but we cannot be confident. Maybe the sample was too small." |
| p = 0.03 | Observed result is unlikely if H₀ is true. Reject H₀. | "This result would be surprising if the training had no effect. We have reasonable evidence it works." |
| p = 0.001 | Observed result is very unlikely if H₀ is true. Strong evidence. | "This would almost never happen by chance alone. We are very confident the effect is real." |
🚨 What p-values Do NOT Tell You:
- They do NOT measure the probability that H₀ is true or false.
- They do NOT measure clinical importance or practical significance.
- They do NOT show the size of the effect (a tiny effect in a huge sample can have p < 0.001).
- They do NOT prove that the exposure caused the outcome — only that they are associated.
- They should ALWAYS be interpreted alongside confidence intervals, effect sizes, and study design.
Example: A drug lowers BP by 0.5 mmHg with p = 0.001. Statistically significant? Yes. Clinically important? No — no doctor would change practice for half a millimetre.
Confidence Intervals and Tests
A confidence interval (CI) gives us a range of plausible values for the true population parameter. When combined with a p-value, it tells us both whether an effect exists and how large it might be.
| Result | How to Read It |
|---|---|
| Mean difference = 8.4 points, 95% CI: 7.0 to 9.8 | Likely improvement. The CI does not include 0, so the difference is statistically significant. We are 95% confident the true improvement is between 7.0 and 9.8 points. |
| Risk ratio = 1.20, 95% CI: 0.90 to 1.60 | Not statistically clear. The CI includes 1 (the "no effect" value for ratios), meaning the true effect could be no effect at all. The sample size may be too small. |
| Mean difference = 2.1 kg, 95% CI: −0.5 to 4.7 | Not significant. CI includes 0. The intervention might help a little, hurt a little, or do nothing. We cannot tell from this study. |
For differences (subtraction), "no effect" = 0. If the CI includes 0, the result is not significant.
For ratios (division), "no effect" = 1. If the CI includes 1, the result is not significant.
Mnemonic: "Difference = Zero, Ratio = One."
Type I and Type II Errors
Because we are making decisions based on sample data (not the whole population), we can make mistakes. There are two types of errors we must understand:
| Reality | Your Decision | Error Name | Nursing Analogy |
|---|---|---|---|
| H₀ is true (No real effect) | You reject H₀ | Type I Error False Positive α = 0.05 | You think a new training works, but it actually doesn't. You waste resources rolling it out. |
| H₀ is false (Real effect exists) | You fail to reject H₀ | Type II Error False Negative β (beta) | You think a life-saving drug doesn't work, but it actually does. You deny patients effective treatment. |
Power of a Study
Power = 1 − β. It is the probability that a study will correctly detect a real effect when one truly exists. Power improves when:
- Sample size is larger (more data = more precision).
- Measurement tools are accurate and reliable (less random noise).
- The true effect size is large (big differences are easier to detect).
- Variability in the data is low (consistent measurements).
Mnemonic: "Type I = Innocent person convicted" (false positive — you found an effect that isn't there).
Mnemonic: "Type II = Ignored criminal" (false negative — you missed an effect that is there).
The 6-Step Testing Framework
Do not start by calculating. Start by understanding the data and the question. Follow this framework for every hypothesis test:
- Ask a clear question. What exactly are you trying to find out? Make it specific.
- State H₀ and H₁. Translate your question into formal null and alternative hypotheses.
- Choose α. Usually 0.05. Decide before you see the data.
- Select the right test. This depends on your data type and structure (see Section 12).
- Calculate the test statistic and p-value. Use software, calculator, or manual formulas.
- Interpret in context. Translate the numbers back into nursing or public health language.
QHATCI — "Quite Hard Always To Calculate, Initially"
Choosing the Right Test: The Decision Tree
The most common mistake students make is using the wrong test. The right test follows from two questions:
- What type of outcome variable do you have? (Categorical vs. Continuous)
- How many groups are you comparing, and are they related or independent?
| Outcome Variable | Groups / Comparison | Suitable Test | Nursing Example |
|---|---|---|---|
| Categorical (counts, proportions, yes/no) | Two categorical variables (association) | Chi-square (χ²) test | Is net use associated with malaria positivity? |
| Continuous (mean scores, weight, BP, time) | Same people measured twice (before/after) | Paired t-test | Did knowledge scores increase after training? |
| Continuous | Two separate, unrelated groups | Independent two-sample t-test | Is waiting time different in Clinic A vs. Clinic B? |
Part 1: Chi-Square (χ²) Test
Used for relationships between two categorical variables.
Purpose
- Tests whether two categorical variables are associated (linked) or independent (unrelated).
- Compares observed counts with the counts we would expect if H₀ were true.
- Useful for tables of frequencies — not means.
Chi-Square Formula
χ² = Σ (Observed − Expected)² / Expected
Sum across all cells in the table.
Worked Example: Net Use and Malaria
Research Question: Is sleeping under an insecticide-treated net (ITN) associated with malaria test result among febrile children?
- H₀: There is no association between net use and malaria test result.
- H₁: There is an association between net use and malaria test result.
Step 1: Observed Counts (Contingency Table)
| Net Use | Malaria Positive | Malaria Negative | Total |
|---|---|---|---|
| Yes (slept under net) | 18 | 42 | 60 |
| No (did not sleep under net) | 32 | 28 | 60 |
| Total | 50 | 70 | 120 |
The counts show fewer positive tests among children who slept under a net (18/60 = 30%) compared to those who did not (32/60 = 53.3%). But is this difference likely due to chance?
Step 2: Calculate Expected Counts
Formula: Expected count = (Row Total × Column Total) ÷ Grand Total
| Cell | Calculation | Expected Count |
|---|---|---|
| Yes & Positive | (60 × 50) ÷ 120 | 25 |
| Yes & Negative | (60 × 70) ÷ 120 | 35 |
| No & Positive | (60 × 50) ÷ 120 | 25 |
| No & Negative | (60 × 70) ÷ 120 | 35 |
Expected counts represent what we would see if net use and malaria were completely unrelated (H₀ true). Notice the totals still add up to 60 and 60.
Step 3: Calculate χ²
| Cell | Observed (O) | Expected (E) | O − E | (O − E)² | (O − E)² / E |
|---|---|---|---|---|---|
| Yes & Positive | 18 | 25 | −7 | 49 | 1.96 |
| Yes & Negative | 42 | 35 | +7 | 49 | 1.40 |
| No & Positive | 32 | 25 | +7 | 49 | 1.96 |
| No & Negative | 28 | 35 | −7 | 49 | 1.40 |
| Total χ² | 6.72 | ||||
Step 4: Interpretation
- χ² = 6.72, degrees of freedom (df) = 1, p ≈ 0.010.
- Since p = 0.010 ≤ 0.05, we reject H₀.
- Conclusion: Malaria test result is statistically associated with net use in this sample.
- Context wording: Children who slept under a net had lower observed malaria positivity (30% vs. 53.3%).
❌ Avoid saying: "Net use caused the lower malaria rate."
✅ Correct: "Net use was associated with lower malaria positivity." Causation requires stronger study designs (like RCTs), not just a chi-square test.
Chi-Square Assumptions
- Data are counts or frequencies, not percentages alone. (You can have percentages, but the test works on the raw counts behind them.)
- Categories are mutually exclusive: each person belongs in one cell only.
- Observations are independent — one person's result does not affect another's.
- Expected counts are generally at least 5 in most cells (some say 80% of cells). If not, use Fisher's exact test.
- Sample is collected in a way that represents the target population.
Practice Task: Handwashing and Diarrhoea
Scenario: A nurse investigates whether availability of handwashing stations at school is associated with diarrhoea among pupils.
| Handwashing Station | Diarrhoea: Yes | Diarrhoea: No | Total |
|---|---|---|---|
| Available | 12 | 48 | 60 |
| Not Available | 24 | 36 | 60 |
| Total | 36 | 84 | 120 |
Your Tasks:
- State H₀ and H₁.
- Calculate one expected count (e.g., Available & Yes).
- Decide whether chi-square is appropriate (check expected counts).
✅ Solution:
- H₀: There is no association between handwashing station availability and diarrhoea.
- H₁: There is an association between handwashing station availability and diarrhoea.
- Expected count (Available & Yes): (60 × 36) ÷ 120 = 18
- All expected counts: 18, 42, 18, 42 — all ≥ 5, so chi-square is appropriate.
- χ² calculation: (12−18)²/18 + (48−42)²/42 + (24−18)²/18 + (36−42)²/42 = 2.0 + 0.86 + 2.0 + 0.86 = 5.72
- Conclusion: p-value ≈ 0.017 (df=1). Since p < 0.05, reject H₀. Handwashing station availability is associated with diarrhoea risk.
Part 2: Paired t-Test
Used when the same person is measured twice (before/after).
Purpose
- Compares two related means.
- Common design: before-and-after measurement on the same participants.
- The test analyses the difference within each pair, not the raw scores.
- It answers: "Is the mean difference significantly different from zero?"
Paired t-Test Formula
t = mean difference ÷ (SD of differences / √n)
Or simply: t = mean difference ÷ standard error of the difference
When Data Are Paired
Paired data means the two measurements are linked or dependent. They are not independent observations.
| Paired Situation | Why It Is Paired |
|---|---|
| Knowledge score before and after training | Same learner measured twice. The "after" score depends on the "before" score. |
| Blood pressure before and after treatment | Same patient. Their baseline BP influences their post-treatment BP. |
| Left eye and right eye of same person | Two measurements linked to one individual's genetics and environment. |
| Mother's weight before and after nutrition counselling | Same mother. Her starting weight affects how much she can gain or lose. |
| Pain score before and after analgesia | Same patient. Their pain tolerance and condition are consistent across both measurements. |
Worked Example: Knowledge Scores Before and After Health Education
Scenario: A nurse educator measures knowledge scores about infant nutrition before and after a 1-hour education session among 8 mothers.
The Raw Data
| Mother | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 |
|---|---|---|---|---|---|---|---|---|
| Before | 45 | 50 | 55 | 60 | 52 | 48 | 58 | 62 |
| After | 55 | 58 | 63 | 70 | 59 | 56 | 64 | 72 |
| Difference (After − Before) | 10 | 8 | 8 | 10 | 7 | 8 | 6 | 10 |
Step-by-Step Calculation
- Calculate the mean difference:
Mean difference = (10 + 8 + 8 + 10 + 7 + 8 + 6 + 10) ÷ 8 = 67 ÷ 8 = 8.38 points - Calculate the standard deviation (SD) of the differences:
Using the formula for sample SD, SD of differences ≈ 1.51 - Calculate the standard error (SE) of the mean difference:
SE = SD / √n = 1.51 / √8 = 1.51 / 2.828 = 0.534 - Calculate the t-statistic:
t = mean difference / SE = 8.38 / 0.534 = 15.73 - Determine degrees of freedom and p-value:
df = n − 1 = 8 − 1 = 7. With t = 15.73 and df = 7, p < 0.001.
Interpretation
- Since p < 0.001 ≤ 0.05, we reject H₀.
- The mean increase in knowledge score (8.38 points) is statistically significant.
- Direction: Scores increased after health education. (Always state the direction!)
- Clinical context: The education session was effective at improving mothers' nutrition knowledge in this sample.
Paired t-Test Assumptions
- The outcome variable is continuous or approximately continuous (e.g., scores, weight, BP).
- Pairs are genuinely linked: same person, matched pair, or before/after design.
- The differences are approximately normally distributed. (With n ≥ 30, this is less critical due to the Central Limit Theorem.)
- No extreme outliers in the differences. One wildly different pair can distort the mean.
- The measurement method is consistent before and after. (Don't change the scoring rubric mid-study!)
How to Report a Paired t-Test
- ✅ Good Report: "Mean knowledge score increased from 53.8 before training to 62.1 after training. The mean paired increase was 8.38 points and was statistically significant, t(7) = 15.73, p < 0.001."
- ❌ Poor Report: "The training worked because p was less than 0.05."
Why it's poor: No means, no direction, no context. The reader has no idea how big the improvement was or whether it matters clinically.
Practice: Pain Scores After Counselling
Scenario: Eight patients have pain scores (0–10 scale) before and after counselling. The differences (After − Before) are: −2, −1, −3, 0, −2, −1, −4, −2.
Answer these:
- What is the outcome variable? Pain score (continuous, 0–10 scale).
- Why is a paired t-test appropriate? Same patient measured twice — before and after counselling.
- What does a negative difference mean? Pain score decreased after counselling. (Negative = After is lower than Before = improvement!)
- What would H₀ say? The mean difference in pain scores = 0 (counselling has no effect on pain).
- What would H₁ say? The mean difference ≠ 0 (counselling changes pain scores).
- Quick calculation: Mean difference = −1.875, SD ≈ 1.25, SE ≈ 0.44, t ≈ −4.26, p ≈ 0.004. Reject H₀. Counselling significantly reduced pain scores.
Part 3: Independent Two-Sample t-Test
Used when comparing means from two separate, unrelated groups.
Purpose
- Compares the mean of a continuous outcome in two independent groups.
- Independent means one participant belongs to only one group — there is no natural pairing.
- Example: Mean waiting time in Clinic A versus Clinic B.
- It asks: "Is the difference between group means larger than what we would expect by chance alone?"
Independent vs. Paired — The Critical Distinction
| Question | Data Structure | Correct Test |
|---|---|---|
| Before and after training in the same learners | Same learners measured twice | Paired t-test |
| Clinic A patients vs. Clinic B patients | Different patients in two clinics | Independent t-test |
| Men vs. women Hb level | Different people in two groups | Independent t-test |
| Vaccinated children vs. unvaccinated children (immunity levels) | Two separate groups of children | Independent t-test |
| Mother's weight before and after nutrition program | Same mothers measured twice | Paired t-test |
Worked Example: Waiting Time in Two Clinics
Scenario: A district health officer wants to know if mean waiting time differs between Clinic A (new digital system) and Clinic B (paper-based system).
The Raw Data (waiting time in minutes)
| Clinic | Wait Times | |||||||
|---|---|---|---|---|---|---|---|---|
| Clinic A (Digital) | 40 | 45 | 50 | 42 | 48 | 46 | 44 | 47 |
| Clinic B (Paper) | 55 | 58 | 60 | 52 | 57 | 61 | 54 | 59 |
Descriptive statistics:
- Clinic A mean = 45.25 minutes
- Clinic B mean = 57.00 minutes
- Mean difference = 45.25 − 57.00 = −11.75 minutes
Inferential statistics:
- t ≈ −7.39
- df ≈ 14 (n₁ + n₂ − 2 = 8 + 8 − 2)
- p < 0.001
Decision: Since p < 0.05, we reject H₀. Mean waiting time differs significantly between the two clinics.
Context: Clinic B (paper-based) had significantly longer waiting times than Clinic A (digital). The digital system appears to improve patient flow.
Independent t-Test Assumptions
- The outcome variable is continuous or approximately continuous.
- The two groups are independent — no person appears in both groups.
- The outcome is approximately normally distributed within each group. (Less critical with large samples, n > 30 per group.)
- No extreme outliers that would distort the mean.
- Variability (variance) in the two groups is reasonably similar. If not, use Welch's t-test (a modified version that handles unequal variances).
How to Report an Independent t-Test
✅ Good Report: "Mean waiting time was 45.25 minutes in Clinic A and 57.00 minutes in Clinic B. The mean difference was −11.75 minutes and was statistically significant, t(14) = −7.39, p < 0.001."
Key addition: Always name the groups and the direction of the difference. Don't just say "significant" — say which group was higher/lower.
Choosing Tests: Mini Scenarios
| Scenario | Correct Test | Why? |
|---|---|---|
| Sex of child vs. immunisation status (yes/no) | Chi-square | Two categorical variables. |
| Weight before and after nutrition counselling in same mothers | Paired t-test | Same people, before/after. |
| Mean Hb in two separate wards | Independent t-test | Two separate groups, continuous outcome. |
| Facility ownership (govt/private) vs. stock-out status (yes/no) | Chi-square | Two categorical variables. |
| Pain score before and after analgesia in same patients | Paired t-test | Same patients measured twice. |
| Mean BMI between urban and rural adolescents | Independent t-test | Two separate groups, continuous outcome. |
| Smoking status (yes/no) vs. TB status (yes/no) | Chi-square | Two categorical variables. |
| Knowledge scores of nurses trained online vs. nurses trained in-person | Independent t-test | Two different groups of nurses, continuous score. |
Common Mistakes to Avoid
| Mistake | Why It Is Wrong | What To Do Instead |
|---|---|---|
| Using percentages instead of counts for chi-square | The formula requires raw counts. Percentages lose the sample size information. | Always construct the table with actual frequencies (n). |
| Using independent t-test for before-and-after data | You ignore the natural pairing, which increases error and reduces power. | Use a paired t-test when the same people are measured twice. |
| Reporting only p-values without means or proportions | A p-value tells you significance but not effect size or direction. | Always report means/SDs or proportions alongside p-values. |
| Claiming causation from a simple association | Association does not prove causation. Confounding variables may explain the link. | Use cautious language: "associated with," "linked to," not "caused." |
| Ignoring assumptions and outliers | Outliers can make means misleading. Non-normal data can invalidate t-tests. | Check your data first. Plot it. Look for outliers. Consider transformations or non-parametric tests. |
| Choosing α after seeing the p-value | This is data dredging / p-hacking. You are moving the goalposts. | Set α = 0.05 before any analysis. Stick to it. |
| Saying "accept H₀" when p > 0.05 | You never prove the null hypothesis. You only fail to find evidence against it. | Say "fail to reject H₀" or "no statistically significant difference was found." |
Practical Work Example
Scenario: A nurse educator measures infection prevention knowledge before and after a 2-hour training session among 20 student nurses. Scores range from 0 to 100.
Your Tasks & Solutions:
- Identify the outcome variable. Infection prevention knowledge score (continuous, 0–100).
- State H₀ and H₁.
- H₀: Mean knowledge score before training = Mean knowledge score after training.
- H₁: Mean knowledge score before training ≠ Mean knowledge score after training.
- Choose the most appropriate test. Paired t-test — same 20 nurses measured before and after.
- Explain why the test is appropriate. The data are continuous, and each participant serves as their own control. The two measurements are dependent (paired).
- Write one sentence showing how you would report a significant result.
"Mean infection prevention knowledge increased from 48.2 before training to 71.5 after training; the mean paired difference of 23.3 points was statistically significant, t(19) = 8.45, p < 0.001."
Group Exercise: Test Selection
Complete the table. For each question, identify the data types and choose the correct test. Then justify your choice in one sentence.
| Question | Data Type | Test | Justification |
|---|---|---|---|
| Is nutrition status associated with sex? | Categorical + Categorical | Chi-square | Both variables are categories; we are testing association. |
| Did average temperature fall after treatment? | Continuous, paired | Paired t-test | Same patients measured before and after treatment. |
| Are mean ages different between two wards? | Continuous, independent | Independent t-test | Two separate groups of patients; age is continuous. |
| Is HIV testing uptake associated with age group? | Categorical + Categorical | Chi-square | Both uptake (yes/no) and age group (e.g., 15–24, 25–34) are categories. |
| Is mean blood pressure lower after a 6-month exercise program? | Continuous, paired | Paired t-test | Same hypertensive patients measured at baseline and 6 months. |
| Are recovery rates different between Hospital X and Hospital Y? | Categorical + Categorical | Chi-square | Recovery (recovered/not recovered) and hospital are both categorical. |
Quick Quiz (Self-Check)
Cover the answers and test yourself. Use one clear sentence per answer.
- What does H₀ usually state?
Answer: H₀ states there is no difference, no association, or no effect — the "status quo" position. - What does p = 0.03 mean at α = 0.05?
Answer: The probability of seeing this result (or more extreme) if H₀ were true is 3%. Since 0.03 < 0.05, we reject H₀ and conclude a statistically significant result. - When do we use a chi-square test?
Answer: When we want to test whether two categorical variables are associated, using frequency counts in a contingency table. - What makes a t-test "paired"?
Answer: The same individual is measured twice (before/after), or two measurements are naturally linked (e.g., left eye and right eye). - Why should clinical meaning be reported alongside statistical significance?
Answer: A result can be statistically significant but clinically trivial (tiny effect in a huge sample), or clinically important but not significant (small sample). Both dimensions matter for patient care. - What is a Type I error?
Answer: Rejecting H₀ when it is actually true — a false positive. We think an effect exists when it does not. - What is a Type II error?
Answer: Failing to reject H₀ when it is actually false — a false negative. We miss a real effect. - What does a 95% confidence interval that includes 0 tell us about a mean difference?
Answer: The difference is not statistically significant at α = 0.05. The true effect could be zero.
Key Take-Away & Decision Summary
The right statistical test follows the research question and the data structure. Do not memorise tests by topic — memorise them by variable type and group relationship.
The Decision Tree
- Categorical association (e.g., net use vs. malaria status) ➔ Chi-square test
- Same people before and after (e.g., knowledge scores pre/post training) ➔ Paired t-test
- Two separate groups with continuous outcomes (e.g., waiting time Clinic A vs. Clinic B) ➔ Independent t-test
- Always interpret in health context, not only by p-value.
- Always state the direction of the effect (which group was higher/lower).
- Never claim causation from a single test — say "associated with."
- Always check assumptions before choosing a test.
- Report means, SDs, and CIs alongside p-values.
References
- Polit, D. F., & Beck, C. T. (2020). Nursing Research: Generating and Assessing Evidence for Nursing Practice (11th ed.). Wolters Kluwer.
- Grove, S. K., & Gray, J. R. (2018). Understanding Nursing Research: Building an Evidence-Based Practice (7th ed.). Saunders.
- Field, A. (2018). Discovering Statistics Using IBM SPSS Statistics (5th ed.). SAGE Publications.
- Daniel, W. W., & Cross, C. L. (2018). Biostatistics: A Foundation for Analysis in the Health Sciences (11th ed.). Wiley.
Quick Quiz
Probability, Hypothesis Testing & Statistical Tests Quiz
Epidemiology and Biostatistics
Privacy: Your details are used only for quiz tracking and certificates.
Probability, Hypothesis Testing & Statistical Tests Quiz
Epidemiology and Biostatistics
Preparing questions...
Choose your answer and keep your streak alive.
Great effort.
Here is your quick performance summary.
