Research Methods, Evidence-Based Practice and Biostatistics
This domain tests whether you can read a paper critically: identify the design, judge the risk of bias, and apply the result to your patient. It also covers the diagnostic and inferential statistics used in physiotherapy literature.
Designs and their questions
Randomised controlled trial answers treatment effect; cohort study answers prognosis and incidence; case–control study suits rare conditions; cross-sectional study measures prevalence; case reports generate hypotheses. Systematic reviews with meta-analysis sit at the top of the evidence hierarchy, expert opinion near the bottom. The evidence-based practice cycle runs: ask a focused question, acquire evidence, appraise it, apply it, and assess the outcome.
Bias control
- Selection bias: prevented by randomisation and allocation concealment.
- Performance and detection bias: prevented by blinding participants, clinicians and assessors.
- Attrition bias: handled by intention-to-treat analysis, which keeps everyone in the group they were allocated to.
- Reporting bias: publishing only favourable outcomes. Internal validity concerns whether the study is sound; external validity concerns whether the findings transfer to your patient.
Diagnostic statistics
Sensitivity = TP/(TP + FN); specificity = TN/(TN + FP). Positive likelihood ratio = sensitivity/(1 − specificity); negative likelihood ratio = (1 − sensitivity)/specificity. A likelihood ratio above 10, or below 0.1, substantially shifts your thinking; values between 2 and 5 shift it modestly. 'SpPin' - a highly specific test, when positive, rules the condition in; 'SnNout' - a highly sensitive test, when negative, rules it out.
Inference and effect size
A p-value is the probability of the observed data, or something more extreme, if the null hypothesis is true; it is not the probability that the null is true. A 95% confidence interval is the estimate plus or minus 1.96 standard errors. Use paired tests for within-group change, independent tests for two groups, ANOVA for more than two, chi-square for categorical data, and non-parametric equivalents when assumptions fail. Number needed to treat = 1/absolute risk reduction; Cohen's d of 0.2, 0.5 and 0.8 represents small, moderate and large effects.
Worked example
A test has 80% sensitivity and 90% specificity, and prevalence in your clinic is 10%, with 1000 patients. TP = 80, FN = 20, FP = 90, TN = 810. PPV = 80/170 ≈ 47%. The same test looks far better in a high-prevalence specialist setting, which is why pretest probability must drive interpretation.
Traps that cost marks
- Reading a p-value as the probability the treatment works.
- Ignoring prevalence when interpreting a predictive value.
- Calling a study randomised when allocation used day of attendance.
Evidence-Based Practice and Clinical Biostatistics Applied
This section focuses on applying evidence at the bedside: converting trial results into patient-relevant numbers, choosing the right statistical test, and knowing when a result should change practice. It complements the research-methods section by emphasising interpretation rather than design.
From trial result to your patient
Absolute risk reduction = control event rate − experimental event rate. Number needed to treat = 1/absolute risk reduction. Relative risk reduction can look large while the absolute difference stays small, so always compute both. For harms, calculate the number needed to harm the same way; a treatment is only worth using when benefits clearly outweigh harms for this patient.
Reading a confidence interval
A 95% confidence interval gives the range of values compatible with the data. If a trial reports a mean difference of 4 points (95% CI 1 to 7) on a scale whose minimal clinically important difference is 6 points, the result is statistically significant but not clinically important. If the interval crosses the line of no effect, the study is inconclusive rather than neutral.
Matching the test to the data
- Paired t-test for before-and-after scores in one group; Wilcoxon signed-rank test when normality fails.
- Independent t-test for two unrelated groups; Mann–Whitney U as its non-parametric counterpart.
- One-way ANOVA for three or more groups; Kruskal–Wallis as the non-parametric alternative.
- Chi-square for associations between categorical variables; Pearson correlation for the strength of a linear relationship between two continuous variables, remembering that correlation is not causation.
- Type I error is a false positive (alpha, usually 0.05); type II error is a false negative, and power is 1 − beta.
Applying evidence to a real patient
Use the PICO frame: population, intervention, comparison, outcome. Then judge applicability: is your patient similar to the trial population, is the intervention feasible in your setting, and does the outcome matter to the patient? GRADE separates strong from weak recommendations and downgrades evidence for risk of bias, inconsistency, indirectness, imprecision and publication bias.
Worked example
A trial in knee osteoarthritis reports that 30% of the control group and 20% of the exercise group had persistent pain at 12 weeks. ARR = 0.30 − 0.20 = 0.10, so NNT = 10. Reasoning: you must treat ten patients with the exercise programme for one additional patient to avoid persistent pain. That is a clinically worthwhile number for a low-risk intervention, and it should be quoted alongside any harms and the patient's own preferences.
Traps that cost marks
- Reporting relative risk reduction alone and implying a large benefit.
- Claiming benefit when the confidence interval crosses zero.
- Choosing a parametric test for clearly skewed ordinal data.