This case is, however, retained based on the following rationale:
the 5 missing values are spread across 4 latent constructs;
there is no violation of the ‘more than one third of the total values per latent
construct’ condition described above;
10.9 per cent is very close the guideline figure of 10 per cent in rule of thumb
(C); and, significantly
considering rule of thumb (D), removing the case does not significantly
improve the extent of the missing data (changes to all missing data points and
missing data per variable are negligible).
Cases with... Number of cases Per cent of all
casesPer cent missing
values per casea^
No missing values 177 83.1^0
One missing value 24 11.3^ 2.^2
Two missing values 10 4.7^ 4.^3
Three missing values 1 0.5^ 6.5^
Five missing values 1 0.5^ 10.^9
Total cases 213 - -
a – based on 46 item statement variablesTable 6-4 Distribution of missing values by case
Step 3 – Randomness of missing data
Step 2 has allowed for the removal of concentrated occurrences of missing values
and established that the data set is now within acceptable limits for calculating
replacement values and moving on to the SEM analyses.
Step 3 and 4 of Hair et al.’s four-step procedure are somewhat redundant in this
example. Specifically, step 3 characterises the degree of randomness of the
missing data and this informs the selection of imputation method in step 4. In
this example, the preferred method of data imputation is the model-based
method, and this method is robust even where data are missing not at random.
Nevertheless, it is a useful exercise to move through step 3 in order to clarify the
characteristics of the differing degrees of randomness and to fully report the
characteristics of the missing data in this research.