Getting started with tables

The title (or legend) of a table should enable the reader to understand its content, so a clear, concise description of the contents of the table is required. The specific details needed for the title will vary according to the type of table. For example, titles for tables of characteristics should give details of the study population being summarised and indicate whether separate columns are presented for particular characteristics, such as sex. For tables of main findings, the title should include the details of the type of statistics presented or the analytical method. Ideally the table title should enable the table to be examined and understood without reference to the rest of the article, and so information on study, time and place needs to be included. Footnotes may be required to amplify particular points, but should be kept to a minimum. Often they will be used to explain abbreviations or symbols used in the table or to list confounding factors for which adjustment has been made in the analysis.

Clear headings for rows and columns are also required and the format of the table needs careful consideration, not least in regard to the appropriateness and number of rows and columns included within the table. Generally it is better to present tables with more rows than columns; it is usually easier to read down a table than across it, and page sizes currently in use are longer than they are wide. Very large tables can be hard to absorb and make the reader’s work more onerous, but can be useful for those who require extra detail. Getting the balance right needs care.

Many research articles present a summary of the characteristics of the study population in the first table. The purpose of these tables is to provide information on the key characteristics of the study participants, and allow the reader to assess the generalisability of the findings. Typically, age and sex will be presented along with various characteristics pertinent to the study in question, for example smoking prevalence, socio-economic position, educational attainment, height, and body mass index. A single summary column may be presented or perhaps more than one column split according to major characteristics such as sex (i.e. separate columns for males and females) or, for trials, the intervention and control groups.

Subsequent tables generally present details of the associations identified in the main analyses. Sometimes these include results that are unadjusted or ‘crude’ (i.e. don’t take account of other variables that might influence the association) often followed by results from adjusted models taking account of other factors.

Other types of tables occur in some papers. For example, systematic review papers contain tables giving the inclusion and exclusion criteria for the review as well as tables that summarise the characteristics and results of each study included in the review; such tables can be extremely large if the review covers many studies. Qualitative studies often provide tables describing the characteristics of the study participants in a more narrative format than is used for quantitative studies. This paper however, focuses on tables that present numerical data.

The main summary statistics provided within a table depend on the type of outcome under investigation in the study. If the variable is continuous (i.e. can take any numerical value, between a minimum and a maximum, such as blood pressure, height, birth weight), then means and standard deviations (SD) tend to be given when the distribution is symmetrical, and particularly when it follows the classical bell shaped curve known as a Normal or Gaussian distribution (see Fig. 1a). The mean is the usual arithmetic average and the SD is an indication of the spread of the values. Roughly speaking, the SD is about a quarter of the difference between the largest and the smallest value excluding 5% of values at the extreme ends. So, if the mean is 100 and the SD is 20 we would expect 95% of the values in our data to be between about 60 (i.e. 100–2×20) and 140 (100 + 2×40).

The median and inter-quartile range (IQR) are usually provided when the data are not symmetrical as in Fig. 1b, which gives an example of data that are skewed, such that if the values are plotted in a histogram there are many values at one end of the distribution but fewer at the other end [7]. If all the values of the variable were listed in order, the median would be the middle value and the IQR would be the values a quarter and three-quarters of the way through the list. Sometimes the lower value of the IQR is labelled Q1 (quartile 1), the median is Q2, and the upper value is Q3. For categorical variables, frequencies and percentages are used.

Common statistics for associations between continuous outcomes include differences in means, regression coefficients and correlation coefficients. For these statistics, values of zero indicate no association between the exposure and outcome of interest. A correlation coefficient of 0 indicates no association, while a value of 1 or −1 would indicate perfect positive or negative correlation; values outside the range −1 to 1 are not possible. Regression coefficients can take any positive or negative value depending on the units of measurement of the exposure and outcome.

For binary outcome measures that only take two possible values (e.g. diseased versus not, dead versus alive, obese versus not obese) the results are commonly presented in the form of relative measures. These include any measure with the word ‘relative’ or ‘ratio’ in their name, such as odds ratios, relative risks, prevalence ratios, incidence rate ratios and hazard ratios. All are interpreted in much the same way: values above 1 indicate an elevated risk of the outcome associated with the exposure under study, whereas below 1 implies a protective effect. No association between the outcome and exposure is apparent if the ratio is 1.

If a 95% CI around a result for a continuous variable (difference in means, regression or correlation coefficient) includes 0 then it is unlikely that there is a real association between exposure and outcome whereas, for a binary outcome, a real association is unlikely if the 95% CI around a relative measure, such as a hazard or odds ratio, includes 1.

The p-value is the probability that the finding we have observed could have occurred by chance, and therefore there is no identifiable association between the exposure of interest and the outcome measure in the wider population. If the p-value is very small, then we are more convinced that we have found an association that is not explained by chance (though it may be due to bias or confounding in our study). Traditionally a p-value of less than 0.05 (sometimes expressed as 5%) has been considered as ‘statistically significant’ but this is an arbitrary value and the smaller the p-value the less likely the result is simply due to chance [8].

Frequently, data within tables are presented with 95% CIs but without p-values or vice versa. If the 95% CI includes 0 (for a continuous outcome measure) or 1 (for a binary outcome), then generally the p-value will be greater than 0.05, whereas if it does not include 0 or 1 respectively, then the p-value will be less than 0.05 [9]. Generally, 95% CIs are more informative than p-values; providing both may affect the readability of a table and so preference should generally be given to 95% CIs. Sometimes, rather than giving exact p-values, they are indicated by symbols that are explained in a footnote; commonly one star (*) indicates p < 0.05, two stars (**) indicates p < 0.01.

Results in tables can only be interpreted if the units of measurement are clearly given. For example, mean or median age could be in days, weeks, months or years if infants and children are being considered, and 365, 52, 12 or 1 for a mean age of 1 year could all be presented, as long the unit of measurement is provided. Standard deviations should be quoted in the same units as the mean to which they refer. Relative measures, such as odds ratios, and correlation coefficients do not have units of measurement, but for regression coefficients the unit of measurement of the outcome variable is required, and also of the exposure variable if it is continuous.

The examples are all drawn from recent articles in Archives of Public Health. They were chosen to represent a variety of types of tables seen in research publications.

The table of characteristics in Table 1 is from a study assessing knowledge and practice in relation to tuberculosis control among in Ethiopian health workers [10]. The authors have presented the characteristics of the health workers who participated in the study. Summary statistics are based on categories of the characteristics, so numbers (frequencies) in each category and the percentages of the total study population within each category are presented for each characteristic. From this, the reader can see that:

The table of characteristics in Table 2 is from a study of the relationship between distorted body image and lifestyle in adolescents in Japan [11]. Here the presentation is split into separate columns for boys and girls. The first four characteristics are continuous variables, not split into categories but, instead, presented as means, with the SDs given in brackets. The three characteristics in the lower part of the table are categorical variables and, similar to Table 1, the frequency/numbers and percentages in each category are presented. The p-values indicate that boys and girls differ on some of the characteristics, notably height, self-perceived weight status and body image perception.

In Table 3, considerable detail is given for continuous variables in the table. This comes from an article describing the relationship between mid-upper-arm circumference (MUAC) and weight changes in young children admitted to hospital with severe acute malnutrition from three countries [12]. For each country, the categorical characteristic of sex is presented as in the previous two examples, but more detail is given for the continuous variables of age, MUAC and height. The mean is provided as in Table 2, though without a standard deviation, but we are also given the minimum value, the 25th percentile (labelled Q1 – for quartile 1), the median (the middle value), the 75th percentile (labelled Q2, here though correctly it should be Q3 – see above) and the maximum value. The table shows:

It is unusual to present as much detail on continuous characteristics as is given in Table 3 . Usually, for each characteristic, either (a) mean and SD or (b) median and IQR would be given, but not both.

Many results tables are simple summaries and look similar to tables presenting characteristics, as described above. Sometimes the initial table of characteristics includes some basic comparisons that indicate the main results of the study. Table 4 shows part of a large table of characteristics for a study of risk factors for acute lower respiratory infections (ALRI) among young children in Rwanda [13]. In addition to presenting the numbers of children in each category of a variety of characteristics, it also shows the percentage in each category among those who suffered ALRI in the previous two weeks, and provides p-values for the differences between the categories among those who did and did not suffer from ALRI. Thus only 2.9% of older children (24–59 months) within the study suffered from ALRI, compared with about 5% in the two youngest categories. The p-value of 0.001, well below 0.05, indicates that this difference is statistically significant. The other finding of some interest is that children who took vitamin A supplements appeared to be less likely to suffer from ALRI than those who did not, but the p-value of 0.04 is close to 0.05 so not as remarkable a finding as for the difference between the age groups.

Table 5 shows a summary table of average life expectancy in British Columbia by socioeconomic status [14]. The average life expectancy at birth and the associated 95% CIs are given according to level of socio-economic status for the total population (column 1), followed by males and females separately. The study is large so the 95% CIs are quite narrow, and the table indicates that there are considerable differences in life expectancy between the three socioeconomic groups, with the lowest category having the poorest life expectancy. The gap in life expectancy between the lowest and highest category is more than three years, as shown in the final row.

Continuous outcome measures can be analysed in a variety of ways, depending on the purpose of the study and whether the measure of the exposure is continuous, categorical or binary.

Table 6 shows an example of correlation coefficients indicating the degree of association between the exposure of interest (cognitive test scores) and the outcome measure (academic performance) [15]. No confidence intervals are presented, but the results show that almost all the particular cognitive test scores are statistically significantly associated (p-value < 0.05) with the two measures of academic performance. Note that this table is an example of where a footnote is used to give information about the p-values. Not surprisingly, all the correlations are positive; one would expect that as cognitive score increase so too would academic performance. The numbers labelled “N” give the number of children who contributed data to each correlation coefficient.

Table 7 is quite a complex table, but one that bears examination. It presents regression coefficients from an analysis of pregnancy exposure to nitrogen dioxide (NO₂) and birth weight of the baby in a large study of four areas in Norway; more than 17,000 women-baby pairs contributed to the complete crude analysis [16]. Regression coefficients are presented and labelled “Beta”, the usual name for such coefficients, though the Greek letter β, B or b are sometimes used. They are interpreted as follows: for one unit increase in the exposure variable then the outcome measure increases by the amount of the regression coefficient. Regression coefficients of zero indicate no association. In this table, the Beta in the top left of the table indicates that as NO₂ exposure of the mother increases by 1 unit (a ‘unit’ in this analysis is 10 μg/m³, see the footnote in the table, which gives the units of measurement used for the regression coefficients: grams per 10 μg/m³ NO₂) then the birth weight of her baby decreases (because the Beta is negative) by 37.9 g. The 95% CI does not include zero and the p-value is small (<0.001) implying that the association is not due solely to chance.

However, reading across the columns of the table gives a different story. The successive sets of columns include adjustment for increasing numbers of factors that might affect the association. While model 1 still indicates a negative association between NO₂ and birth weight that is highly significant (p < 0.001), models 2 and 3 do not. Inclusion of adjustment for parity or area and maternal weight has reduced the association such that the Betas have shrunk in magnitude to be closer to 0, with 95% CIs including 0 and p-values >0.05.

The table has multiple rows, with each one providing information on a different subset of the data, so the numbers in the analyses are all smaller than in the first row. The second row restricts the analysis to women who did not move address during pregnancy, an important consideration in estimating NO₂ exposure from home addresses. The third row restricts the analysis to those whose gestational age was based on the last menstrual period. These second two rows present ‘sensitivity analyses’, performed to check that the results were not due to potential biases resulting from women moving house or having uncertain gestational ages. The remaining rows in the table present stratified analyses, with results given for each category of various variables of interest, namely geographical area, maternal smoking, parity, baby’s sex, mother’s educational level and season of birth. Only one row of this table has a statistically significant result for models 2 and 3, namely babies born in spring, but this finding is not discussed in the paper. Note the gap in the table in the model 2 column as it is not possible to adjust for area (one of the adjustment factors in model 2) when the analysis is being presented for each area separately.

Table 8 presents results from a study assessing whether children’s eating styles are associated with having a waist-hip ratio greater or equal to 0.5 (the latter being the outcome variable expressed in binary form – ≥0.5 versus <0.5) [17]. Results for boys and girls are presented separately, along with the number of children in each of the eating style categories. The main results are presented as crude and adjusted odds ratios (ORs). The adjusted ORs take account of age, exercise, skipping breakfast and having a snack after dinner, all of these being variables thought to affect the association between eating style and waist-hip ratio. Looking at the crude OR column, the value of 2.04 in the first row indicates that, among boys, those who report eating quickly have around twice the odds of having a high waist-hip ratio than those who do not eat quickly (not eating quickly is the baseline category, with an odds ratio given as 1.00). The 95% CI for the crude OR for eating quickly is 1.31 – 3.18. This interval does not include 1, indicating that the elevated OR for eating quickly is unlikely to be a chance finding and that there is a 95% probability that the range of 1.31 – 3.18 includes the true OR. The p-value is 0.002, considerably smaller than 0.05, indicating that this finding is ‘statistically significant’. The other ORs can be considered in the same way, but note that, for both boys and girls, the ORs for eating until full are greater than 1 but their 95% CIs include 1 and the p-values are considerably greater than 0.05, so not ‘statistically significant’, indicating chance findings.