How to read a clinical research paper

Correct answer: It buys what the data cannot rule out: with an upper bound of 1.932, something close to a doubling of risk remains on the table

A p value answers how far an estimate sits from the null by compressing both ends of the interval into one number, which makes too little data to exclude anything look the same as plenty of data around a genuinely tiny effect. The 0.805 is the first of those. Reading the interval first shows which one it is: an upper bound of 1.932 leaves something close to a doubling of risk within what the data allow. The point estimate of 1.077 is not enough to declare a direction either, because when the interval spans the null the direction is undetermined, and calling it harmful is as unfounded as calling it protective.

Correct answer: In looking only at ordering. The lowest-risk group was predicted 0.287 and observed 0.364, while the C-index is entirely immune to a shift of the whole risk scale

The C-index asks only whether the model ranks higher-risk people above lower-risk ones, and shifting or stretching everyone's predicted value leaves the ordering, and therefore the C-index, untouched. The gap between 0.642 and 0.666 accordingly says nothing about whether the predicted numbers are right. The real problem is calibration: the lowest-risk group was predicted 0.287 and observed 0.364, and it is the predicted value that a threshold rule uses. The optimism invoked by the second answer is near zero in this example, because the development cohort has an ample number of events relative to its parameters and almost nothing to overfit, so that is not the explanation here. As for the threshold the first answer appeals to, there is no universal pass mark for a C-index. What counts as acceptable depends on how predictable the outcome is in the first place and on what the model is for, since ranking a waiting list and deciding whether to add adjuvant treatment do not need the same discrimination. And even if 0.642 were judged sufficient, the failure on this page would still be calibration rather than discrimination; recasting it as a threshold question only swaps one discrimination measure for another. A prediction model paper with a C-index and no calibration has reported half of what matters.

Correct answer: Respecting the matching moves the odds ratio only from 1.07 to 1.09; what flips the conclusion is something else, controlling a confounder

Only a comparison that differs in one thing supports attribution. With the same variables and only the matched structure restored, the odds ratio moves from 1.07 to 1.09, which is hardly at all, and the 4.15 arrives only once the number of spontaneous abortions is added, a step that changes two things at once. Booking the whole difference to the matching credits the wrong cause. None of this makes ignoring the matching harmless, since it distorts the standard error and the interval and the direction of its bias depends on how the matching variables relate to exposure and outcome. The answer built on 0.76 treats a tiny change in a p value as evidence, and comparing p values between two estimates that are both far from significance yields nothing.

Correct answer: Apparent sensitivity climbs to 0.776 while apparent specificity falls, because the people skipped for verification hold far more true negatives than false negatives

Apparent sensitivity climbs from 0.634 to 0.776 while apparent specificity falls, so the two move in opposite directions. The mechanism is that the skipped negatives contain far more true negatives than false negatives: dropping them strips true negatives out of the denominator of specificity and false negatives out of the denominator of sensitivity. So saving money without much cost has it backwards, since the test is made to look both more sensitive and less specific, and by no small margin, with specificity down to 0.293 when only a tenth of negatives are verified. This bias never labels itself in a paper and has to be inferred from how many people received the reference standard, which is what the flow diagram is for.

Correct answer: That every estimate afterwards actually has a denominator of 299 rather than the dataset total

Every estimate afterwards runs on 299 people, not 316. The difference is those excluded for missing values, and for a complete-case analysis to hold the missingness has to be benign enough, which need not be true if people with missing values were followed more loosely in the first place. The 48 events do set the precision, but precision cannot answer whom an estimate represents. Most statistical misreadings come down to the denominator in the end: a baseline p value reflects sample size rather than difference, a case-control denominator is chosen by the investigators, verification bias works on the denominator, and a prediction model's calibration fails because the new population's denominator is built differently.

Correct answer: Because those 165 controls were selected by the investigators, so the denominator corresponds to no population

Whether incidence can be computed depends on whether the denominator is a defined population observed over time, not on how large the sample is. The 165 controls here were picked to satisfy matching criteria, and their chance of entering the study was not governed by natural incidence, so any proportion built on them corresponds to nothing in the world outside. The number of cases, 83, affects precision alone, and the total of 248 was assembled in the same way. Nor is the absence of recorded follow-up time the reason: a nested case-control study is still a case-control study, yet because its controls are sampled from a defined cohort it can estimate incidence. What decides is where the controls came from, not how much data there is.