Page 492 - Encyclopedia of Nursing Research
P. 492
SAmplIng n 459
associate degrees and those with baccalaure- all elements (or a relevant, random subset)
ate degrees separately. within each cluster. In contrast to stratified
For research purposes or gains, it is best sampling where one samples from all strata S
to select classification variables based on their of the classification variable, with cluster
assumed association with the dependent var- sampling one samples only some clusters,
iable. If more than one classification variable for example, some practice sites or some
is used, it also is advantageous if they are hospitals.
uncorrelated with each other. Stratified sam- Whereas the goal of stratified sampling
pling facilitates obtaining subgroup param- is to obtain homogeneous strata, when one
eter estimates and comparisons—especially does cluster sampling one wants the clusters
when some strata are rarer and stratification to be as heterogeneous as possible. To the
is used to ensure an adequate number of extent that the clusters are not heterogeneous,
cases in each stratum for valid comparisons. one loses precision and the cluster sample is
Stratified sampling also may increase the sta- less efficient than a simple random sample of
tistical efficiency of estimates if proportional the same size. At the extreme, if the cluster
allocation (as opposed to equal allocation) is is completely homogeneous, one achieves
used, and may be more convenient if sam- no gain from more than one case per cluster.
pling lists are organized according to the Cluster sampling generally is used for prag-
selected strata. matic purposes when there is no other way
The intent with stratified sampling is to to easily obtain the targeted sample than
decrease sampling variability by increasing through the identification of clusters.
the homogeneity of the strata. If one forced The last type of sample discussed here
equal numbers of cases in each stratum, it is convenience samples or nonprobability
is important to remember that the resulting samples. These are frequently used in nurs-
sample will not reflect the natural distribu- ing research, but their implications often
tion of the classification variable. In those are ignored. First, it is not possible to esti-
cases, one must assign weights to the cases to mate sampling errors with such samples.
reflect the known proportionate distribution Therefore, the validity of inferences drawn
of the strata in the population if one wishes to from nonprobability samples to the popula-
conduct analyses involving the classification tion remains unknown and whenever non-
variable in addition to analyses comparing random selection is used, the potential for
the strata within each classification variable. serious sample selection biases exists.
Stratified sampling, however, may be more lastly, it is important to note that sample
costly and complex. lastly, the control advan- selection bias may threaten internal as well
tages of using stratified sampling are limited as external validity (Berk, 1983). One way in
because stratification generally is applied to which this may happen is when investigators
some, but not all, variables of interest. inadvertently sample on their dependent var-
Cluster sampling is a fourth type of ran- iable by excluding cases at either the high or
dom sampling. With cluster sampling, the low end of values on the dependent variable.
elements of interest for the study and the For example, if one is studying the impact of
sampling units are not same. The sampling amputation on depression and quality of life,
unit, or cluster, is a convenient, practical, and but screens out all those currently diagnosed
economical grouping—for example, prac- with and on medications for depression, one
tice sites; hospitals—whereas the elements may obtain an erroneous or misspecified
of interest for the study may be the individ- model because those at one end of the depres-
ual patients obtained at the practice sites or sion continuum have been excluded from the
hospitals. With cluster sampling, one ran- sample. In a bivariate analysis, this misspeci-
domly samples the clusters and then takes fication will include either an attenuation or

