HS 326 common mistakes
HS 326 · Understand the traps and how to avoid them.
When a question stem reports a p-value of 0.03 and asks whether the result is clinically meaningful, the candidate answers that it is, treating statistical significance as proof of practical importance.
Candidates conflate the null-hypothesis decision with effect size and precision. p-values are computed under a specified null and say nothing about the magnitude of the effect or its relevance to patients, yet significance language ('significant') is read as 'important' in everyday usage.
What to do instead
Before answering, restate the effect estimate with its confidence interval and ask whether the interval excludes values that matter clinically. Practice classifying results into a 2x2 of statistically significant/not by clinically important/not, using items where the two diverge.
The candidate reads only the abstract of a trial and answers questions about internal validity (allocation concealment, blinding, attrition handling) from abstract wording rather than checking whether the methods are described at all.
Abstracts are optimized to report results, not methods, and time pressure rewards the fastest available text. Candidates then infer quality from the presence of words like 'randomized' without evidence that the randomization was concealed or that outcome assessors were blinded.
What to do instead
Adopt a fixed appraisal checklist (randomization and concealment, blinding by outcome, completeness of follow-up, pre-specified outcomes, funding) and answer method questions only from the methods section. When only an abstract is given, mark method items as unclear rather than assuming adequacy.
On a question about a screening test, the candidate computes positive predictive value from sensitivity and specificity alone, without using prevalence.
Sensitivity and specificity are the numbers usually quoted and memorized, and they are prevalence-independent, so candidates assume PPV is too. The dependence of predictive values on pre-test probability is the single most commonly missed quantitative relationship.
What to do instead
Redraw a 2x2 table with a hypothetical 1,000 patients at the stated prevalence for every screening question, fill in true and false positives and negatives, and read PPV and NPV directly off the table. Verify that changing prevalence while holding sensitivity and specificity fixed changes PPV.
The candidate answers a question about whether a study supports causation by citing the strength of association or a small p-value, ignoring confounding, temporality, and selection of participants.
Candidates substitute statistical strength for causal reasoning, and observational designs are mentally filed as 'weaker' without a structured framework for why. Without a checklist, the judgment becomes impressionistic.
What to do instead
Apply an explicit causal checklist to every aetiology question: temporality, plausibility, dose-response, consistency across studies, specificity, reversibility, and strength, plus a named confounder and whether it was measured or adjusted. State which criterion the study fails before choosing an answer.
The candidate designs or critiques a study and proposes a sample size 'to be representative' or 'to reach significance' without specifying the outcome, the expected effect, the variance, alpha, and power.
Sample-size justification is taught as a formula to memorize rather than as a consequence of the research question. Candidates also confuse representativeness (a sampling issue) with precision (a sample-size issue).
What to do instead
For any design question, write the five inputs before the number: primary outcome and its scale, the smallest effect worth detecting, the assumed variability or event rate, one- or two-sided alpha, and power. Then explain what increasing the sample changes (precision and power) and what it does not (bias).
The candidate selects a matching or stratification strategy to control confounding but then analyzes the matched data as if it were an independent unmatched sample, or adjusts for a variable that lies on the causal pathway.
Candidates learn that matching and adjustment 'control confounding' without learning the analysis consequences (paired or conditional analysis) or the distinction between confounders and mediators. Overadjustment is invisible unless the causal structure is drawn.
What to do instead
Sketch a directed acyclic graph for every adjustment question, mark exposure, outcome, confounders and mediators, and select the minimal sufficient adjustment set. If matching was used, choose the paired or conditional analysis; if a variable is a mediator, do not adjust for it when estimating the total effect.
The candidate answers an evidence-based-practice question by naming a guideline or 'the literature' as the basis for a decision, without specifying the patient's values and preferences, the local resource context, or the strength of the recommendation.
Guidelines are memorized as answers rather than as one input into a decision. Candidates also conflate strong and conditional recommendations, so they apply a conditional recommendation as a mandate for every patient.
What to do instead
Structure each answer as question, best available evidence with its grade, patient values and circumstances, and clinical judgment. When citing a recommendation, state whether it is strong or conditional and what that implies for a patient whose values conflict with it.
The candidate reports a relative risk reduction as the benefit to communicate to a patient, or describes a change in risk without the baseline, so that a reduction from 2 in 1,000 to 1 in 1,000 is presented as 'a 50% reduction'.
Relative measures are larger, more impressive, and more frequently reported in abstracts and promotional material. Absolute measures require the baseline risk, which candidates often do not compute.
What to do instead
Convert every effect to absolute terms before answering: calculate the absolute risk reduction, then the number needed to treat as its reciprocal, and state the baseline risk and the time horizon. Present both relative and absolute figures when communicating, and check that the NNT denominator matches the outcome and follow-up period.
The candidate treats a conflict of interest disclosure, institutional review board approval, or a data-sharing statement as a substitute for answering the underlying integrity question about authorship, fabrication, or selective reporting.
Regulatory language is familiar and sounds exculpatory, so candidates pick the option containing an oversight body. Approval certifies that a plan was reviewed, not that the conduct or reporting was honest.
What to do instead
Separate the three integrity questions: was the study permissible (approval, consent), was it conducted and reported truthfully (fabrication, falsification, selective outcome reporting), and was credit allocated fairly (authorship criteria, disclosure). Choose the option that addresses the specific failure described in the stem.
The candidate answers a shared-decision-making or interprofessional item by describing what the clinician will tell the patient or the team, rather than how information will be elicited, checked for understanding, and integrated.
Communication items are read as knowledge tests about the correct management option, so candidates pick the clinically right plan and overlook the process skills the item is testing, such as teach-back, exploring values, or clarifying roles.
What to do instead
For communication items, identify the function the response must serve (elicit, inform, verify understanding, negotiate, or coordinate) and select the option that performs that function with a specific technique. Reject options that only assert a plan or reassure without checking comprehension.