Tip for data extraction for meta-analysis - 25

What if you’re missing a mean and only a similar statistical statistic is given?

Kathy Taylor

Previously, I highlighted a list of ways where, when extracting data for meta-analysis of
continuous outcomes, you might find that a summary statistic that you want is missing. In this
post I’ll give some examples of the 3rd way - a similar summary statistic is reported, but it’s
not the statistical measure that you want - when you have missing means.

Finding a median reported

You may find that instead of a mean, a median is reported. A median is a different type of
average. The reporting of medians indicates that the distribution of outcome data is skewed.
The median and mean are equal if the distribution of the data is perfectly symmetrical (Figure
1). When the distribution is skewed, the mean and median will differ, and the difference
between them will depend on the degree of skewness.

Figure 1. The relative position of the mean and median depending on the data’s distribution

The Cochrane Handbook (section 6.5.2.9) highlights three papers which provide equations for
estimating means from other summary statistics.
One is the paper by Hozo et al who concluded that even with skewed data, the sample mean
can be estimated by the median. They provided estimates for the mean based on the sample
size and range (min and max are the bounds of the range):

𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 ≈ 𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 if n ≥25

𝑚𝑚𝑚𝑚𝑚𝑚+2𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚+𝑚𝑚𝑚𝑚𝑚𝑚
𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 ≈ if n<25
4
The curved equal sign means ‘approximately equal’. The estimates were tested using
simulation, and drawing samples from normal and skewed distributions.

The second paper is by Bland who uses more information by providing estimates of the
sample mean based on the median, range, sample size and interquartile range (q1 to q2):

𝑚𝑚𝑚𝑚𝑚𝑚(𝑛𝑛 + 3) + 2(𝑛𝑛 − 1)(𝑞𝑞1 + 𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 + 𝑞𝑞3 ) + max(𝑛𝑛 + 3)

8𝑛𝑛
When n is large, this equation simplifies to

𝑚𝑚𝑚𝑚𝑚𝑚 + 2(𝑞𝑞1 + 𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 + 𝑞𝑞3 ) + max

8
He tested his estimates on three real data sets with simulated data drawing samples from
normal and skewed distributions.

The formulae of Bland and Hozo et al both work better with small samples.

The third paper is by Wan et al who provide estimates of the sample mean based on the
median and interquartile range. This has the advantage of not being influenced by extreme
values.
𝑞𝑞1 + 𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚 + 𝑞𝑞3
3
Using simulation, they also tested their estimates, drawing samples from normal and skewed
distributions and found smaller relative errors compared to Bland’s approach. Wan et al also
provide a very useful spreadsheet which you can use to calculate and compare their estimated
means with those of Bland and Hozo et al.

Another paper by Luo et al provide improved estimates of the sample mean based on the
sample size, median, range and interquartile range. Their estimates use a weighted
formulation. They consider three scenarios.
Firstly, when the sample size, median and range are reported.

4 𝑚𝑚𝑚𝑚𝑚𝑚 + 𝑚𝑚𝑚𝑚𝑚𝑚 𝑛𝑛0.75

� �� �+� � 𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚
4 + 𝑛𝑛0.75 2 4 + 𝑛𝑛0.75

Secondly, when the sample size, median and interquartile range are reported.

0.39 𝑞𝑞1 + 𝑞𝑞3 0.39

�0.7 + �� � + �0.3 − � 𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚
𝑛𝑛 2 𝑛𝑛

Thirdly, when the sample size, median, range and interquartile range are reported.

2.2 𝑚𝑚𝑚𝑚𝑚𝑚 + 𝑚𝑚𝑚𝑚𝑚𝑚 0.72 𝑞𝑞1 + 𝑞𝑞3 0.72 2.2

� 0.75
�� � + �0.7 − 0.55 � � � + �0.3 + 0.55 − � 𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚𝑚
2.2 + 𝑛𝑛 2 𝑛𝑛 2 𝑛𝑛 2.2 + 𝑛𝑛0.75

These are approximations of their more complicated formulae that are reported in their
paper. They demonstrate the accuracy of their estimates using simulation and they provide
an excel spreadsheet.

Finding a geometric mean reported

Sometimes a geometric mean is reported. This is another type of average, which arises from
the analysis of skewed data which have been log-transformed and then back-transformed
(using the exponential function) when presenting results. With small samples, skewed data is
often log-transformed, before analysis, because standard inferences on the means of skewed
data is only acceptable for large samples. With large samples we assume that the means of
outcome measurements are approximately normally distributed due to the central limit
theorem.

So, instead of means (which are more formally known as arithmetic means) and standard
deviations (SD), geometric means are reported, either with confidence intervals (CIs), the
exponential of the SD of the log-transformed values (often referred to as the tolerance factor
or the inappropriately named as the ‘SD of the geometric mean’), or the exponential of the
standard error (SE) of the log-transformed values. Geometric means and arithmetic means
should not be pooled. If most of your studies report arithmetic means, you will want to
convert geometric mean summary data to arithmetic mean summary data. Pooling is possible
by using the conversion equations of Higgins et al. It’s a two-stage process (Figure 2) as the
geometric mean data has to be log-transformed first.
 Figure 2. From geometric means to arithmetic means

Using the following notation:

g is the geometric mean of a treatment arm

(glower to gupper ) is the confidence interval of g
𝑒𝑒 𝑠𝑠𝑧𝑧 is the incorrectly named SD of g
𝑒𝑒 𝑆𝑆𝑆𝑆𝑧𝑧 is the incorrectly named SE of g

STEP 1

Calculate the log-transformed measurements (𝑧𝑧̅ 𝑎𝑎𝑎𝑎𝑎𝑎 𝑠𝑠𝑧𝑧 ) from the geometric mean data.

𝑧𝑧̅ = 𝑙𝑙𝑙𝑙(𝑔𝑔) and

�ln�𝑔𝑔𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢 �−ln(𝑔𝑔𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 )�√𝑛𝑛

𝑠𝑠𝑧𝑧 = OR
2𝑡𝑡
𝑠𝑠𝑧𝑧 = 𝑙𝑙𝑙𝑙(𝑒𝑒 𝑠𝑠𝑧𝑧 ) in cases where 𝑒𝑒 𝑠𝑠𝑧𝑧 has been reported OR
𝑠𝑠𝑧𝑧 = 𝑙𝑙𝑙𝑙�𝑒𝑒 √𝑛𝑛×𝑆𝑆𝑆𝑆𝑧𝑧 �
where t is the 97.5 percentage point of the t-distribution with (n-1) degrees of freedom.

STEP 2

Apply the conversion equations to the log-transformed data to calculate the arithmetic mean
summary data (𝑥𝑥̅ 𝑎𝑎𝑎𝑎𝑑𝑑 𝑠𝑠𝑥𝑥 ). There are two sets of equations depending on the similarities
between the SDs of the two treatment arms. Higgins et al recommend comparing the SDs on
the log scale as it’s more plausible. If the SDs are different, use Method 1. If the SDs are similar,
use Method 2.
Method 1

For each treatment arm, calculate

𝑠𝑠𝑧𝑧2
𝑥𝑥̅ = 𝑒𝑒𝑒𝑒𝑒𝑒 �𝑧𝑧̅ + �
2
𝑠𝑠𝑥𝑥 = �(𝑒𝑒𝑒𝑒𝑒𝑒(𝑠𝑠𝑧𝑧2 ) − 1)𝑒𝑒𝑒𝑒𝑒𝑒(2𝑧𝑧̅ + 𝑠𝑠𝑧𝑧2 )

Method 2

First calculate
2
(𝑛𝑛1 − 1)𝑠𝑠𝑧𝑧,1 2
+ (𝑛𝑛2 − 1)𝑠𝑠𝑧𝑧,2
𝑠𝑠𝑧𝑧,𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝 =�
𝑛𝑛1 + 𝑛𝑛2 − 2

Then, for each treatment arm, calculate

𝑠𝑠𝑧𝑧,𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝
𝑥𝑥̅ = 𝑒𝑒𝑒𝑒𝑒𝑒 �𝑧𝑧̅ + �
2
2 2
𝑠𝑠𝑥𝑥 = �(𝑒𝑒𝑒𝑒𝑒𝑒(𝑠𝑠𝑧𝑧,𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝 ) − 1)𝑒𝑒𝑒𝑒𝑒𝑒(2𝑧𝑧̅ + 𝑠𝑠𝑧𝑧,𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝 )

Higgins et al also provide equations to convert the other way, from arithmetic means to
geometric means (Figure 3). You might want this if the majority of your included studies
report geometric means.

Figure 3. From arithmetic means to geometric means

STEP 1
Convert the arithmetic mean summary data to log summary data.
𝑥𝑥̅ and 𝑠𝑠𝑥𝑥 are the arithmetic mean and SD
𝑧𝑧̅ and 𝑠𝑠𝑧𝑧 are the mean and SD of the log data

If the SDs are different, use Method 1. If the SDs are similar, use Method 2.
Method 1
For each treatment arm, calculate
1 𝑠𝑠𝑥𝑥2
𝑧𝑧̅ = 𝑙𝑙𝑙𝑙(𝑥𝑥̅ ) − 𝑙𝑙𝑙𝑙 � 2 + 1�
2 𝑥𝑥̅
𝑠𝑠2𝑥𝑥
𝑠𝑠𝑧𝑧 = �𝑙𝑙𝑙𝑙 � + 1�
�2
𝑥𝑥

Method 2
For each treatment arm, calculate

𝑠𝑠2𝑥𝑥
𝑠𝑠𝑧𝑧 = �𝑙𝑙𝑙𝑙 � + 1�
�2
𝑥𝑥

Then calculate
2
(𝑛𝑛1 − 1)𝑠𝑠𝑧𝑧,1 2
+ (𝑛𝑛2 − 1)𝑠𝑠𝑧𝑧,2
𝑠𝑠𝑧𝑧,𝑝𝑝𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 = �
𝑛𝑛1 + 𝑛𝑛2 − 2

Then for each treatment arm, calculate

1 2
𝑧𝑧̅ = 𝑙𝑙𝑙𝑙(𝑥𝑥̅ ) − 𝑠𝑠𝑧𝑧,𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝
2

STEP 2

Back-transforming (exponentiating) the log data calculates the geometric mean data.
𝑧𝑧̅ and 𝑠𝑠𝑧𝑧 are the log data.
g = 𝑒𝑒 𝑧𝑧̅ and 𝑒𝑒 𝑠𝑠𝑧𝑧 are the geometric mean data.

Meta-analysis can be carried out on the log scale and SD for the log values can be calculated
using the following equation:
(𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢𝑢 𝐶𝐶𝐶𝐶 − 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙 𝐶𝐶𝐶𝐶)
𝑆𝑆𝑆𝑆 = √𝑛𝑛
3.92

Let me show you an example of converting geometric mean data to arithmetic mean data.

For the review that I worked on we used these conversion equations when we extracted data
from the study by Romero et al. They reported microalbuminuria (albumin excretion rate) at
6 months with geometric mean (95% CI). Data for the intervention group, treated with
Captopril, was converted from 60 (35 to 104) mg/24hr to mean (SD) of 90 (101) mg/24hr, and
data for the untreated group was converted from 91 (58 to 141) mg/24hr was converted to
119 (101) mg/24hr (Table). I applied the equations from Method 1.

Table. Calculating arithmetic mean data

Captopril Untreated
n 13 13
gm 60 91
g lower 35 58
g upper 104 141
z=Ln(g) 4.09 4.51
dof 12 12
t-value 2.18 2.18
Sz 0.90 0.74
x 90.05 119.22
Sx 100.77 100.91

See below to find out where the t-value came from. Higgins et al highlight that their estimates
are likely to be biased in small sample studies. As this study had small samples, it was removed
as part of a sensitivity analysis.

Other methods

Other approaches of dealing with missing means highlighted by the review of Weir et al
include the simulation-based approximate Bayesian computation (ABC) approach of Kwon
and Reis.

If a large proportion of studies have missing means, pooling is not recommended.

Here’s a tip…
You can derive estimates of means from
other reported summary statistics.
In my next blog post I’ll give some more examples of a similar summary statistic is reported,
but it’s not the statistical measure that you want when you have missing SDs.

Where did the equations and t-value come from?

Converting geometric means to arithmetic means and the reverse
Higgins et al derive their equations in their paper.

Calculating means from medians, range and interquartile range

Hozo et al, Bland, Wan et al and Luo et al also derive their equations in their respective papers.

Calculating an SD from a 95% confidence interval:

This was derived earlier and in my next blog post I will give more details.

What about the t-value?

This came from a t-distribution table (Figure 4).

Figure 4. t-distribution table

The first column of the table shows the degrees of freedom (dof) and the area probabilities
(also known as percentages or p-values) are shown in the first row. As indicated in the t-
distribution curve above the table, the p-values represent the area under the t-distribution
curve in the tail, from the t-value to infinity (shaded black) for different dofs.

p values
dof

t-values

In the example I gave, the Captopril group had n=13 so dof=n-1=12.

The area probability for the 97.5 percentage point of the t-distribution using the above table
=1-0.975=0.025. The corresponding t-value is 2.179 (Figure 5).

Figure 5. t-distribution table for dof=12 and p-value=0.025

The same t-value can also be calculated using EXCEL by typing in an EXCEL cell
=ABS(T.INV(one sided p-value, dof)). With my example this is =ABS((T.INV(0.025,12).

Note that I’m using EXCEL 2016. Earlier versions use the term TINV(two-sided p-value, dof)
i.e. TINV(0.05,12)=ABS(T.INV(0.025,12)=2.178813.
Dr Kathy Taylor teaches data extraction in Meta-analysis,
https://www.conted.ox.ac.uk/courses/meta-analysis</link> This is a short course that is
also available as part of our MSc in Evidence-Based Health Care
https://www.conted.ox.ac.uk/about/msc-in-evidence-based-health-care, MSc in Medical
Statistics
https://www.conted.ox.ac.uk/about/msc-in-ebhc-medical-statistics, and MSc in
Systematic Reviews
https://www.conted.ox.ac.uk/about/msc-in-ebhc-systematic-reviews

Follow me on Twitter @dataextips for updates on my blog, related news, and to find out
about further examples where others, like me, are trying to make statistics more broadly
accessible.