Common mistakes when describing a study population (the typical Table 1)

Reports of both observational and experimental studies typically include descriptive summaries of the group of people who were included in the study. This helps readers assess the generalizability of the study's results and how well they might apply to particular patients or groups of patients. Such description is usually provided in tabular form, and is usually the first table in the paper (Table 1). This is recommended in the CONSORT guidelines for reporting clinical trials and for most other types of clinical and translational research studies. Characteristics described usually include basic demographics such as age, sex, and race/ethnicity, along with others that are specifically relevant for the issues being investigated. An example layout is below


Variable

Treatment
(N=201)

Placebo
(N=199)

Age in years, mean\xB1SD

62\xB17

61\xB18

Sex, N (%) Male

99 (49%)

98 (49%)

Female

102 (51%)

101 (51%)

Hypertension, N (%)

27 (13%)

29 (15%)

...

...

...

This divides the study population according to a key predictor variable, in this case a randomized treatment assignment. This is typical for randomized controlled trials, and also for observational studies that have one categorical predictor variable that is of primary interest. For other studies, the table may only have one column, describing the entire study population.

There are two mistakes that are common in Table 1 summaries.

Reversing Outcome and Predictors

When a study has a dichotomous primary outcome variable, researchers often put this into the role that the dichotomous primary predictor variable plays in the example above. Unfortunately, this reverses the roles of predictor and outcome, which can lead to confusion. For example:

Variable

Survived
(N=200)

Died
(N=52)

P-value

Age (mean\xB1SD)

62\xB17

71\xB112

<0.0001*

Sex: Male

150 (68%)

45 (87%)

0.0096

Female

70 (32%)

7 (13%)

...

...

...

...

*Unpaired t-test with unequal variances

†Fisher's exact test

This shows a table that would be more appropriate if survived versus died were the predictor and the row variables were outcomes. For example, we see that those who died averaged 9 years older than those who survived, which would help us if we wanted to predict the ages of decedents and survivors. The table does not indicate what we would prefer to know, which is how older age influences the risk of death. Similarly, the percentages do not directly show how much higher the risk of death is for males compared to females.

Tables like this are often shown in observational studies where authors need something to put in place of the randomized treatment in the usual "Table 1" for randomized clinical trials. A better choice would be to simply include descriptive summaries of the entire population studied.

Another possibility is to combine some descriptive summaries with simple analyses that have the predictor and the outcome in their correct roles. For example:

Univariate logistic regression models


Predictor

# Died/N (%) or
Median (range)


OR (95% CI)


P-Value

Age (per decade)

65 (51 - 79)

2.2 (1.53-3.2)

<0.0001

Sex: Female

7/77 (9)

--reference--

Male

45/195 (23)

3.0 (1.25-8.3)

0.0096

...

...

...

...

This provides some descriptive summaries and shows how to analyze the same data the other way around, with the outcome and predictors in the right roles. We can see directly the estimated impact of age and sex on risk of death, including the more useful percentages, the 9% death rate in females versus 23% in males.

P-values for Randomized Treatments

The first example above does not show p-values for comparing treatment and placebo, because they are not relevant. Contrary to common belief, p-values do not indicate whether a difference is "real" or large enough to be important; they only indicate the strength of evidence against the null hypothesis that the difference arose by chance. Because we already know that all differences between randomized treatment arms are in fact due to chance, any such evidence is meaningless. The CONSORT guidelines document included an acknowledgement of this problem:
Unfortunately significance tests of baseline differences are still common; they were reported in half of 50 RCTs trials published in leading general journals in 1997. Such significance tests assess the probability that observed base\xADline differences could have occurred by chance; however, we already know that any differences are caused by chance.