Data Review and Assessment

Data Review and Quality Assessment

Always check to see if the range of the numeric values are reasonable and acceptable
- 10 point scale - are all 10 levels shown & how are they coded: is the range 0-9 or perhaps 1-10? [what are the advantages or disadvantages of starting at 0 versus starting at 1?]
- does the scale start at 0, if not why are there 0’s in your data?
- ZEROS are IMPORTANT and should always be investigated and confirmed
are there negative numbers, should there be?
are the numbers evenly distributed, should they be?
should these data be processed numerically or treated as ordinal/categorical levels (e.g. race, income levels, age groups) - does computing the mean make sense? (e.g. Likert scale)
are some of the values outside of your expected scale due to missing data or other coding you need to adjust for prior to analysis

check that numeric codes for missing have been treated as such
how much is missing per variable
how much is missing per subject
are there any obvious patterns
- people at end of study - missingness correlated with time
- other predictors of missing - depressed subjects more likely not to complete study
- variables/measures with more missing - sensitive items people refuse to answer, income, risky sexual behaviors, …
we’ll explore this more in future sessions
REMEMBER MISSING DATA IS DATA!
Pros and Cons of Missing Data imputation
- bias in the presence of missing data
- bias introduced by imputation

assuming a normal distributions, outliers often defined as outside +/- 2 standard deviations or +/- 3 standard deviations
but what if the distribution is NOT normal (income, biomarkers, count data)
extreme values (LVAD heart transplant cost) - can almost be treated like missing imputation - non-parametrics - bootstrapping - robustized stats
illogical values - possible typos

mathematical transformations
- square root, log, others (useful for right skewed tailed distributions)
- left skewed distributions use inverse (REF - value)
- approaches used to help “noramlize” the distribution
- pros and cons - nice distribution but complicated to interpret
distributions expected based on response patterns
- possible bimodal or other clusters defined by other factors like gender, race, etc.
zero-inflated or hurdle type questions
- symptom ratings 0-10, where 0=not present;
- 1-10 rates severity or frequency (really 2 questions in 1)
physical functioning:
- 0’s indicate cannot perform task,
- >0 measures how well/how far/how much they did activity
- again 2 measures/questions in 1
adherence & knowledge tests
- lots of zero’s, lots of 100’s
- will often have floor or ceiling effects
- could be bimodal
counts (“poisson” distribution)
- number of children;
- days in hospital (length of stay);
- even Age
time to event (“gamma” related distributions)
- time to disease onset or health event (heart attack, stroke)
- time to readmit

look for typos (2.6 versus 6.2; 11 instead of 1; 40 instead of 4)
unit mismatches (inches instead of cm; feet vs meters; kg versus pounds)
extreme values (biomarkers, costs, income)
income
- often yearly instead of monthly
- be sure common reporting units are used
negative numbers used to indicate missing
really high values to indicate missing
negative values or odd value occurring after performing date math (i.e. date of event happening before date of birth; date of discharge before date of admission; length of stay in years and you’re expecting days - possible year typo (often occurring in Dec-Jan change over))

always check to see if the data came in as a number if that what’s you’re expecting
- especially if the imported data accidently had text somewhere in the column
- for example, labs often report low values below detection as <x.xx or high values above the saturation limit of the detector as >x.xx. Look at for these leading (or following) text characters before or after the number of interest
in SPSS switch from string to numeric - watch for value deletion insertion of missing values
in R check for auto-creation of “factors”

often dates come in as text due to inconsistent formatting
watch for Excel conversion from date format to number format (the articles we discussed in previous lecture)
issues moving between Excel and other software
- the “REFERENCE” data van vary between software, i.s. number of days since January 1, 1900, see https://support.office.com/en-us/article/datevalue-function-df8b07d4-7761-4a93-bc33-b7471bbff252
- and this is different from EXCEL for MAC, see https://support.microsoft.com/en-us/help/214330/differences-between-the-1900-and-the-1904-date-system-in-excel
- SAS stored dates as the number of days since 01/01/1960, see https://communities.sas.com/t5/SAS-Enterprise-Guide/Date-conversion-to-integer/td-p/157849 and https://communities.sas.com/t5/SAS-Procedures/How-to-convert-date-to-numeric/td-p/295961
- and SPSS stores dates as the number of days since the beginning of the Gregorian calendar October 14, 1582, see this article discussing converting dates between SAS, Stata and SPSS, https://stats.idre.ucla.edu/spss/code/converting-sas-or-stata-dates-to-spss-dates/