What is missing data? Edit
Survey responses may have missing data because:
- A respondent did not answer the question (intentionally or unintentionally)
- For hard copy surveys:
- The answer provided did not follow the requested format
- More than two answers are provided
- A response is illegible
In addition, missing data can be introduced, usually unintentionally, if the:
- Data entry person does not enter the data
- Data analyst deletes data (intentionally or unintentionally)
- For numeric variables, the default for missing data is a decimal point or full-stop in the cell, which will appear automatically. Alternatively a specific value, such as -1 or 99, can be used and specified in the Variable View to indicate missing values.
- For string variables, missing data is indicated by a blank cell
- Missing is indicated by a blank cell
- Coded missing values can be specified through Variables
Missing data in a cell will mean that a variable and a case each have some missing data.
Dealing with missing data Edit
The presence of missing data should be identified through data screening.
Strategies for dealing with missing data should be decided prior to data analysis.
One strategy for dealing with missing data is listwise. This means that all cases with even a single piece of missing data (for the variables of interest) will not be used e.g.,:
DESCRIPTIVES VARIABLES=VAR00001 /STATISTICS=MEAN STDDEV MIN MAX /MISSING=LISTWISE.
In other words, to be used in an analysis, each case must be complete (have no missing data) for the variables of interest.
This approach has the advantage of only working with complete data, but it risks removing a lot of potentially useful data.
Alternatively, missing data can be dealt with pairwise. This means that all available data is used, even from cases with some missing data.
This approach includes more data, but it can mean that there are different Ns for different analyses which can get confusing.
An alternative approach is imputation. Imputation involves predicting or "filling in" the missing data.
The simplest form of imputation is mean replacement (i.e., replace the missing data with the mean score for other cases for the same variable). However, this will exaggerate the central tendency.
When composite scores are computed using responses to two or more variables, tolerance for some missing data can be allowed. In this way, a composite score can be created even when a respondent has some missing data.
More sophisticated imputation uses regression-based prediction (i.e., by using scores on other related variables to predict the missing value). Two common methods are estimation maximalisation or multiple imputation.
See also Edit
- Advanced handling of missing data (Workshop slides, Nicole Janz)
- The treatment of missing data (pdf; Howell)
- Missing data treatment (Howell)
- Statistical analysis with missing data: Table of contents (Little & Rubin, 2019)