Data preparation & presentation Gary Collins EQUATOR Network, - - PowerPoint PPT Presentation

Data preparation & presentation

Gary Collins

EQUATOR Network, Centre for Statistics in Medicine NDORMS, University of Oxford EQUATOR Network – OUCAGS training course 24 October 2015

2

Data preparation

• Prior to ANY data analysis it is important to check

your data

– often more time is spent ‘cleaning’ and examining the data than the actual analysis – reliable results rely on accurate and reliable data

• garbage in  garbage out
• Data accuracy can be improved by prospectively

thinking about the data before collecting it

– produce a data collection sheet (also systematic reviews) – set up database – is possible, include basic checks in the database

• often done for clinical trials

3

Prospective data collection

4

Data Definitions (CRASH dataset)

Variable Label Comments Max length Type Codes Patient ID Six digit unique identifier treatment box-pack number 7 string Sex Gender of the patient 1 Number 0=Male 1=Female DOB Date of birth of the patient DD/MM/YYYY 10 Date

• 7303=DOB not

known TRAND Time of randomisation HH:MM:SSS 8 Time GCS_EYE Glasgow Coma Scale: Eye

• pening

1 Number 4=spontaneous 3=to sound 2=to pain 1=none

5

Data preparation (after collection)

• For data types

– discrete (e.g. number of children): check for non-integer values – continuous (e.g. height): check values line in plausible range – binary (e.g. male/female) – nominal (e.g. blood type) – ordinal (e.g. cancer stage) – dates: check for valid dates, date of event before date of birth – most statistical software doesn’t by default differentiate between upper and lower case (e.g. yes, Yes, yES, yeS, yES, Yes, YeS, YES)

• Missing data: consistency in how this is recorded

– how much is missing

Check for unlikely values

6

Van den Broeck et al, PLoS Med 2005

7

Implausible values

Cancer Male Female Total Breast 45 45 Colorectal 23 17 40 Lung 8 7 15 Prostate 12 1 13 Total 43 70 113

8

Outliers

• Outliers will seem incompatible with the rest of the

data

• May deviate from the main body of the data
• Usually extreme values (high or low)
• Can be genuine observations
• Can have considerable influence on the results
• Any suspicious values should be checked
• The decision to include or exclude outliers should

be made with caution

• Provided all the measurements are valid, often useful to

analyse with and without the outlier

9

Data presentation (exploration/presentation)

• Plot the data, plot the data, plot the data !!!

– prior to any statistical analysis – look at the data! – get familiar with the data – what does it look like – can highlight/indicate any errors (spurious values) in the data part of data cleaning) – what is the distribution of the data (univariate)?

• do they have an approximate Normal distribution?
• if not, can may be transformed (e.g. log transformation)?
• Plot the data, plot the data, plot the data !!!

10

Unfortunately, not all data look like this

11

Positively skewed data

Bland & Altman. BMJ 1996

12

CRASH-2 trial

(distribution of baseline Glasgow Coma Score)

13

14

15

Results: Graphs

• Statisticians like to (ideally) see the raw data!
• Use graphs to describe results, for example

– Dotplots (or boxplots for large data sets)

• good for comparing groups

– Scatter plots

• good to accompany correlation analyses

– Survival curves

• time-to-event analyses

– Forest plots

• meta-analyses

16

What can you tell me about these data?

For each dataset would you believe that the: Mean of x is 9 Variance of x is 11 Mean of y is 7.5 Variance of y is 4.122 Correlation between x and y is 0.816 Regression line: y = 3 + 0.5x Called ‘Anscombe’s Quartet’ Illustrates the importance of showing your data

17

Data presentation (exploration/presentation)

• Where possible, plot the raw values

– there is NO reason to hide it – let the data tell the story

• Beware of ‘chartjunk’ (Edward Tufte)

– anything that distracts the viewer (including you) from the information that graph is intended to present – let the data speak for themselves, don’t clutter the plot

18

Dynamite plots

19

Dynamite plots

20

Dynamite plots

Schriger & Cooper. Ann Emerg Med 2001 Mean=81.905 mm Mean=77.445 mm

21

The same data presented differently

Schriger & Cooper. Ann Emerg Med 2001

22

23

Summary

• Think upfront about data collection

– How are the data coded? – Be consistent throughout – Have data checks to flag implausible values

• Plot the data, plot the data, plot the data!

– Get to know your data – Check for outliers

• Avoid ‘chartjunk’

– Maximise data:ink ratio