P-values are sometimes misinterpreted to mean “the probability that the results occurred by chance”. This is problematic on at least two counts: the probability of any particular result occurring by chance will be extremely low, and also “by chance” requires further definition to be meaningful. A more technical definition is that the p-value is the probability of finding a result at least as extreme as the observed result if the null hypothesis (usually “no difference”) is correct and all assumptions used to compute the p-value are met.

Errors in p-value interpretation usually involve confusing the following two probabilities:

• The probability that the treatment is ineffective given the observed evidence (the misinterpretation of a p-value)
• The probability of the observed evidence if the treatment were ineffective (what the p-value provides)
  • For more information on this common inference mistake (known as the “The Prosecutor’s Fallacy”) see Westreich D et al.

By convention, a p-value ≤0.05 is considered statistically significant, though this is increasingly recognized as an oversimplification and ignores consideration of clinical importance.

The most important takeaway from this discussion is that the typical understanding of a p-value is incorrect and such misunderstanding can lead to erroneous conclusions. For a further discussion of p-value misinterpretation in medical literature see Price R et al. For a more advanced discussion on why research findings are often false despite statistically significant results see Ioannidis JPA.

The technical definition of a 95% confidence interval (CI) is: If we were to repeat the study an infinite number of times, 95% of 95% CIs would contain the true effect, if all assumptions used to calculate the interval are correct. Consequently, a 95% CI does not entail “there is a 95% chance the true value is within this range” (a common misinterpretation). As with p-values, the true meaning is more nuanced. See here for visual CI simulations as illustrative examples.

The 95% CI provides all of the information of a p-value (and is derived using the same information), but also adds information on a plausible range of the effect size. When interpreting CIs, it is important to examine both ends of the CI and judge whether there is a clinically important difference between them. For example, a point estimate of 5% absolute risk reduction in stroke risk over 5 years with a 95% CI of 3% to 7% will include a narrow range that many clinicians would consider clinically important difference at both ends of the interval. This examination (along with considerations of bias) will help establish the degree of uncertainty in the result. See McCormack J et al. for a discussion of how considering only statistical significance without proper regard for CIs can cause confusion.

By convention, a 95% CI that does not include the null (e.g. a relative effect 95% CI that includes 1.0 or an absolute risk difference 95% CI that includes 0%) is considered statistically significant (i.e. consistent with p<0.05).

It is sometimes believed that sample size (i.e. how many participants were included in the study) is a determinant of internal validity. However, a review (Kjaergard LL et al.) found that smaller trials (<1,000 participants) only exaggerated treatment effects compared with larger trials (≥1,000 participants) when they had inadequate randomization, allocation concealment, or blinding. As such, sample size itself is not indicative of bias. Furthermore, if there were too few participants enrolled to detect a difference between groups this will be reflected in the corresponding wide CI (see Confidence Intervals – How precise were the estimates of treatment effect? for a discussion of wide CIs and how they illustrate precision).

Absolute Risk Difference

The absolute risk difference between groups refers to the risk of an event in one group minus the risk in another group. Consider the following example of a theoretical 2 year trial examining insomnia rates:

In this case, the absolute difference was calculated by subtracting the intervention group event rate (15%) by the comparator group event rate (5%), which equals 10% (15% – 5%). This difference is “absolute” because the number (e.g. +10% risk of insomnia over 2 years) is independently meaningful.

Absolute differences also need to be communicated in the context of time. For example, a 1% absolute risk reduction over 1 month is quite different from a 1% absolute risk reduction over 10 years. As such, absolute differences should be stated as a __% increase/decrease over [timeframe].

Relative Measures of Effect

This contrasts with relative effect measures. One example of a relative effect is relative risk (RR), which is calculated by dividing the risk of event in the intervention group by that in the comparator group.

In this case, the RR was calculated by dividing the intervention group rate (15%) by the comparator group event rate (5%), which equals 3.0 (15% ÷ 5%). This difference is “relative” because the number (e.g. a 3.0 RR of experiencing insomnia) is dependent on the risk in the comparator group to be meaningful. RR 3.0 means that the risk has tripled, but without knowing the baseline risk that is being tripled, then the number is not fully interpretable.

This dependence can be problematic if not properly considered. Consider the following example, where the RR is identical in both cases:

As demonstrated, RR considered in isolation lacks crucial information. The same concept is relevant when (responsibly) buying a product during a sale. Knowing that a particular product is 50% off is not sufficient for a rational choice, as there needs to also be information about the original price (e.g. $20 versus $20,000) before deciding if the purchase is desirable.

Note: RR is just one relative measure – see discussion below for information on relative risks, odds ratios, and hazard ratios.

Both number needed to treat (NNT) and number needed to harm (NNH) are measures of how many patients have to receive the treatment of interest for one additional person to experience the outcome of interest (NNT being for beneficial outcomes, and NNH for harmful outcomes).

It is calculated as: 100 ÷ Absolute risk difference, with the result always rounded up.

For example, if a treatment has a 7% absolute risk increase of causing urinary retention over 3 months then the NNH is 15 (100 ÷ 7 = 14.3, then round up to 15). This means that 15 patients will have to be treated for one of them to have urinary retention over the next 3 months (always including timeframe, as with absolute risk differences).

This is an alternative way to understand absolute risk differences that may be more intuitive to some (though it is more poorly understood by patients than other measures, as discussed here).

Before delving into the details of each type of relative effect it should be noted that all of them have the following features:
Any relative measure = 1.0 means there was no difference between groups
Any relative measure > 1.0 means the outcome was more likely with the intervention than the comparator
Any relative measure < 1.0 means the outcome was less likely with the intervention than the comparator

To demonstrate the differences between these measures of relative effect, consider the following table:

Relative Risk (RR)

Calculating the RR consists of dividing the risk of event in the aspirin group by risk of event in the placebo group.

Using the above table:
The risk of event in the treatment group: A ÷ (A+B)
The risk of event in the comparator group: C ÷ (C+D)
With numbers imputed: 10 ÷ 100 = 0.1 (or 10%) in the aspirin group and 20 ÷ 100 = 0.2 (or 20%) in the placebo group.
The RR is then 0.1 ÷ 0.2 = 0.5.

Odds Ratio (OR)

Calculating the OR consists of dividing the odds of an event in the aspirin group by the odds of an event in the placebo group.

Using the above table:
The odds of event in intervention group: A ÷ B
The odds of event in the comparator group: C ÷ D
With numbers imputed: 10 ÷ 90 = 0.11 in the aspirin group and 20 ÷ 80 = 0.25 in the placebo group.
The OR is then 0.11 ÷ 0.25 = 0.44.

OR are similar to RR when events are rare (A ÷ (A+B) ≈ A ÷ B when A is very small) (Holcomb WL et al.). As events become more common, these measures diverge and ORs will overestimate RRs (such as in this example where RR=0.5 and OR=0.44). The ClinCalc tool can be used to convert OR to RR.

Hazard Ratio (HR)

Hazard ratios (HRs) represent the average of the instantaneous incidence rate at every point during a trial. Consider an example of a 5-year trial that has a HR of 0.70 for the outcome of death comparing an intervention against some comparator. This means that a participant assigned to intervention will be 30% less likely to die relative to the comparator at any point during the trial:

• Year 1: If 5% have died in the comparator group, then 3.5% are expected to have died in the intervention group (5% * 0.70 = 3.5%)
• Year 2: If 10% have died in the comparator group, then 7% are expected to have died in the intervention group (10% * 0.70 = 7%)
• Year 5: If 20% have died in the comparator group, 14% are expected to have died in the intervention group (20% * 0.70 = 14%)

The same is approximately true at any given timepoint during the trial follow-up. These are all approximations as the HR is an average, it (almost certainly) will not be exactly true at every time point. For instance, the final HR might be 0.70, but it could be 0.80 during the first half of the trial and 0.60 during the latter half.

For example, consider the unadjusted analysis from an observational study (Turgeon RD, Koshman SL, et al.) that compared the use of ticagrelor vs. clopidogrel in patients who had undergone percutaneous coronary intervention following acute coronary syndrome (ACS). For the outcome of survival without major adverse coronary events (MACE) the HR was 0.84 before adjustment for potential confounding variables, depicted visually below:

While not exactly true at every time point, past the first 100 days the cumulative proportion patients experiencing death or MACE in the ticagrelor group appears to be roughly 84% of the cumulative proportion in the clopidogrel group fairly consistently (see “Kaplan Meier Curves” for more information below on how to interpret these types of graphs). This coheres with the HR of 0.84 discussed above.

HRs are usually similar to RRs (Sutradhar R et al.). HRs examine multiple timepoints over trial follow-up, whereas RRs evaluate cumulative proportions at the end of the trial (or at another single timepoint). HRs can account for differential follow-up times, and contain more information than RRs/ORs since they include the added dimension of time (Guyatt G et al.). HRs are limited in their ability to convey fluctuations in effect over time, as a HR of 1.0 could mean that there was consistently no effect, or it could mean that there was beneficial effect during the first half and a proportional detrimental effect during the latter half (Hernán MA). However, some of these limitations can be overcome by combining a HR with the use of a Kaplan Meier curve, as discussed below.

Cumulative Hazards

Kaplan Meier curves are graphical representations comparing event accrual between two groups over time. Consider another example from the aforementioned study comparing ticagrelor vs. clopidogrel:

Each curve displays the cumulative proportion of patients in that group who have experienced the outcome of interest. As time passes more participants experience the outcome and the curve progresses downwards. The relative differences in outcome accumulation can be expressed as a HR, as discussed above. Note that, while not displayed, each curve has a CI surrounding it at every time point.

Onset of Benefit

Consider the above Kaplan Meier curve. During the first 50 days, the curves for the two groups overlap. However, after this point the two curves begin to separate. This curve provides insight into the onset of benefit of the intervention and if the benefit is sustained over time. In this case, onset of benefit begins after approximately 50 days and is sustained as time elapses.

Course of Condition

The graph also gives insight into event rates over time. As can be seen above, the curve is steepest initially – indicating that the risk of death or ACS is highest immediately following the intervention. The slope of the curve then flattens and remains relatively stable – indicating the event rate after the initial period is relatively constant during the first year. This demonstrates how Kaplan Meier curves can be useful to understand the course of a condition over time.

Total at Risk (or Number at Risk)

Consider another curve from the same study, this one examining survival without major bleeds:

As depicted, sometimes there is also a “Total at risk” (sometimes called “Number at risk”) table beneath the curve. This can give additional information regarding the participants as they progressed through the trial. All participants begin “at risk”, but as the study progresses the number decreases. Below are the following reasons the number may decrease, as well as possible implications if the decreases are not balanced between groups:

Thus the total at risk table can serve as a clue that further examination should be undertaken to see if there is bias.

Forest plots are used in meta-analyses to graphically depict the effects of an intervention across multiple studies. Consider the labeled example below from the “Beta-blockers for hypertension” Cochrane Review (Wiysonge CS et al.):

Plot 4. Forest plot of beta-blockers vs. placebo in patients with hypertension for the outcome of mortality.

As showcased, forest plots show information about each individual study included for that outcome, and also the combined results. This information is displayed visually as well as numerically. Trials with more events or participants are generally given greater weight. Heterogeneity is typically measured via I², which is 0% in this case. For more information on heterogeneity see here.

The standardized mean difference (SMD) is a method of combining multiple continuous outcome scoring systems into one measurement. For example, when performing a meta-analysis on the effects of antidepressants on depression symptom reduction, trials may use different scales to rate depression symptoms (HAM-D, PHQ-9, etc.). SMD will allow the aggregation of the results of all these studies. Notably, using SMD assumes that differences between studies are due to differences in scales (and not in intervention/population characteristics). Arbitrary “rule-of-thumb” cutoffs (e.g. SMD of 0.2 = “small effect”) may not reflect the minimal important difference.

An alternative approach is to transform the SMD into a more familiar scale (Higgins JPT et al.). Multiply the SMD by the standard deviation (SD) of the largest trial to convert to its scale.

Ideally the review will present this information along with a comparison of the proportion of participants within each group who experienced clinically important improvement or decline (the so-called “responder analysis”). This is useful because calculating the mean alone will not provide any information about the distribution. The responder analysis unfortunately cannot typically be calculated by readers if it is not already reported by the reviewers.

While this section has focused on statistical fundamentals it is important to emphasize that statistics are only one aspect of critical appraisal. Even if a study shows statistical significance, there needs to be considerations of bias, clinical significance, and generalizability.

Bias: P-values and CIs are contingent on all the assumptions being used to calculate them being correct. In other words, they assume there is absolutely no bias present. As such, if the study is poorly conducted (e.g. a RCT without adequate allocation concealment and blinding) this will not be reflected in the statistical analysis of the results. This is why it is necessary to appraise the conduct of the trial to evaluate the credibility of the results.

Clinical significance: Even if a result is statistically significant, it may be too small of an effect to matter to a patient. For instance, with enough participants, a 1-point reduction in pain on a 100-point scale could be statistically significant, but very unlikely to be felt by an individual patient.

Generalizability: Even if the results are unbiased and clinically significant, they will only be useful if they can be applied to practice. If there are substantial differences between the features of the trial and your own practice, then the result may not be applicable.

Other sections of this resource will go into these concepts in more depth, but these are the fundamental reasons why a comprehensive approach to appraisal is necessary, and simply looking at statistical significance in the results section will not be sufficient to understand the clinical implications of a trial.