Critical-review questions become easier to approach when you decide what a study can establish before considering the answer choices. Options often contain recognisable statistical language, but familiarity is not the same as relevance. A correct calculation can still support an incorrect conclusion if you overlook the design, denominator, outcome or follow-up period.
The aim is to build an interpretation in a predictable order: identify the question, describe the data, calculate the relevant effect, inspect uncertainty, and judge the claim. This method works across a range of Paper B research tasks and gives you a precise way to review mistakes.
Start with the research question
Before calculating, identify the population, intervention or exposure, comparator, outcome and time period. A number describing symptom response after six weeks cannot automatically support a conclusion about hospital admission over a year. Similarly, a result in a highly selected study population may need further consideration before it informs another setting.
Write a one-sentence summary in plain English. For example: "This trial compares two approaches for preventing relapse over six months in adults who meet the stated eligibility criteria." That sentence gives every subsequent calculation a context. If the answer option changes the population or outcome, you can see the mismatch immediately.
It also helps distinguish a question about effectiveness from one about diagnostic accuracy, prognosis or association. The appropriate denominator follows the task.
Work through an illustrative trial
Imagine an original teaching example in which 200 participants are allocated to each of two groups. During six months, relapse occurs in 30 participants receiving intervention A and 50 receiving intervention B. Assume for this calculation that outcomes are available for everyone and that the reported numbers refer to the same definition of relapse.
| Group | Relapse | No relapse | Total |
|---|---|---|---|
| Intervention A | 30 | 170 | 200 |
| Intervention B | 50 | 150 | 200 |
The observed risk is 15% with A and 25% with B. The absolute risk reduction is 10 percentage points. The risk ratio is 0.15 divided by 0.25, or 0.60. The relative risk reduction is 40%. Each figure describes a different aspect of the same data; none should be substituted for another without explanation.
The Cochrane Handbook chapter on effect measures explains why risk ratios, odds ratios and risk differences have distinct interpretations. In this example, the odds ratio is approximately 0.53, calculated from 30/170 divided by 50/150. Calling that value a risk ratio would exaggerate the relative reduction compared with the risk-ratio calculation.
Attach a time period to the number needed to treat
The reciprocal of the absolute risk reduction, 1 divided by 0.10, gives an illustrative number needed to treat of 10. The interpretation needs the outcome and follow-up: treating ten comparable participants with A rather than B for six months would correspond to one fewer relapse on average, if the observed difference reflects the treatment effect.
It would be wrong to describe this as ten patients needing lifelong treatment to prevent one admission. That statement changes both the outcome and time period. The arithmetic may remain recognisable while the clinical meaning has disappeared.
When an examination option gives a number needed to treat, check the direction of benefit, the comparator and the outcome. If the underlying difference represents harm, the wording and interpretation must change accordingly.
Uncertainty is a separate question
An effect estimate is not a complete account of the evidence. Precision matters. A confidence interval tells you something about the uncertainty around the estimate under the statistical model; it is not a probability that a fixed true effect lies inside the particular interval you calculated.
For an additional illustrative comparison, suppose two different studies both report a risk ratio of 0.80. One has a 95% confidence interval of 0.72 to 0.89; the other, 0.45 to 1.42. Their point estimates are identical, but the second is compatible with a much wider range of effects. These intervals are separate teaching examples, not calculations from the table above.
The Cochrane guidance on interpreting results emphasises examining effect size and uncertainty. Avoid turning a threshold into the whole conclusion. A result that is not statistically significant does not demonstrate that the interventions are equivalent.
Ask whether the design supports the claim
Return to the relapse trial. The numbers alone do not tell you whether allocation was concealed, how outcomes were assessed, whether the groups received comparable follow-up or whether the analysis followed a prespecified plan. These features affect how confidently the difference can be attributed to the intervention.
Now imagine the same event counts came from people who chose their own treatment. Differences in illness severity, preferences or access to care might contribute to the result. The arithmetic is unchanged, while the causal interpretation becomes more difficult.
This is why "the calculation is correct" is an incomplete answer review. Identify what kind of claim the option makes and whether the study design supports it. An association can be informative without proving that changing the exposure will produce the observed difference.
Missing outcomes can change the interpretation
Suppose the study reports relapse only among participants who attended the final appointment, and attendance differs between groups. Ask why outcomes are missing and whether missingness could relate to relapse or treatment experience. A convenient denominator may conceal an important bias.
Do not invent the missing outcomes or assume they all favoured one group. Instead, state the uncertainty and look for how the investigators addressed it. The relevant question is whether the conclusion remains credible under plausible explanations for the missing data.
For revision, practise describing this problem without statistical jargon first. "We may be comparing different subsets of the original groups" is often the key insight. Technical terminology should clarify that insight rather than replace it.
Review your errors by category
After a question, classify the mistake. Was the denominator wrong? Did you confuse risk with odds? Did you forget the follow-up period? Did you infer equivalence from a non-significant result? Did you treat an observational comparison as a randomised experiment?
Each category suggests a different next task. Arithmetic errors need a fresh calculation. Interpretation errors need a contrasting example. Design errors need a short account of how the study generated its groups and outcomes. Repeating an entire statistics chapter after every miss rarely identifies the specific problem.
iatroX's Paper B practice and Socratic Tutor can support this targeted approach. State your interpretation before requesting help, then use the discussion to test the weakest link. Return later to a new set of numbers so that successful retrieval reflects understanding rather than recognition of the original answer.
Worked dataset: absolute effect, relative effect and uncertainty
This fictional randomised study follows 400 people for six months. An adverse outcome occurs in 20 of 200 participants assigned intervention and 40 of 200 assigned comparison. Assume complete follow-up for this arithmetic exercise.
| Group | Outcome | No outcome | Total |
|---|---|---|---|
| Intervention | 20 | 180 | 200 |
| Comparison | 40 | 160 | 200 |
The risks are 10% and 20%. The absolute risk reduction is 10 percentage points; the risk ratio is 0.5 and relative risk reduction is 50%. The number needed to treat is 1/0.10 = 10 over the stated six months. Reporting “a 50% reduction” without the absolute risks and follow-up loses useful context.
Suppose a separate estimate of the risk ratio has a 95% confidence interval of 0.30 to 0.84. The null value for a ratio is 1, which lies outside that interval. That does not measure the probability that this particular treatment is right for a particular patient, establish the absence of bias or answer whether the effect is clinically worthwhile. A confidence interval for a difference has a different null value: zero.
For a delayed retest, change both event counts while preserving the ratio. If the comparison risk becomes 2% and intervention risk 1%, the relative reduction remains 50%, but the absolute reduction is one percentage point and NNT becomes 100. Explain why the original NNT cannot simply be carried across.
Use the Paper A and B resource-access guide to distinguish MRCPsychiatrist's learning targets from an independently validated prediction of passing. For examination scope, use the RCPsych syllabus that applies from January 2027.
Frequently asked questions
Should I read the options before working through the data?
A brief independent interpretation is often useful. It helps prevent an appealing option from determining which calculation or conclusion you pursue. You can then compare your interpretation with the choices.
Is a relative risk reduction more informative than an absolute reduction?
They answer different questions. Relative effects describe proportional change; absolute effects show the observed difference in event risk. Clinical interpretation usually benefits from understanding both.
Does a non-significant result prove two treatments are equivalent?
No. It may reflect limited precision or other features of the study. Equivalence requires an appropriate design and analysis, rather than simply failing to detect a difference.
Strengthen your research interpretation with the iatroX MRCPsych Paper B question bank.