I am planning to administer a survey to a number of students (female and male) and looking for the appropriate way to choose the sample. The students are grouped into sections course-wise by the system of admission and registration and they are considered homogeneous. My question is that is using cluster sampling would be suitable? I mean I choose not students but sections as students in the sections are considered homogeneous and after that I choose sections randomly as a sample of the study. If this technique is not suitable, what sampling technique would you suggest?
Thanks,
Ali -
If the clusters (sections) are homogeneous, I think you might want to sample just a few from each cluster (secondary sampling) so you can sample more clusters (sections, your primary unit). But that negates some of the convenience of cluster sampling.
If you could, I wonder if you might want to look at stratified random sampling? It might mean you have to spend more effort for a smaller overall sample, but likely a smaller overall sample would be needed. If you have any quantitative results, you might find it easier to estimate variances with stratified random sampling, though you could still do it with cluster sampling though it would be more complex.
I could be wrong, but I have an idea that if you do cluster sampling you might miss some important information, though cluster sampling is very good, and interesting, for many applications.
Cheers - Jim
Thanks a lot Jim for the useful answer. I understand that stratified sampling is for heterogeneous population where people are categorized into groups where they fit, according to similarity and any certain quality they share, then a sample is drawn out of the groups. Here, in my case, the sections are already existent and students are homogeneous. You correct me if mistaken.
Regards,
Ali -
I just saw this.
I see. Well, that sounds reasonable, but if you census each section, you will have a very inefficient sample, meaning its design effect (variance for the same sample size compared to simple random sampling - a concept by Leslie Kish) would not be favorable. But if doing a census for each section is easy, then the low efficiency becomes less of a problem.
As for stratification, I was just wondering, for example, if you could group your sections by some more general category so that you would sample from each of those groups or strata, as it sounded like your sections might vary a lot, and you could, by an unfortunate selection, end up with a 'bad'/'unrepresentative' sample. I did a lot of work with the model-based approach, and I know that when you have a strictly probability design-based approach, your biggest hazard may be an unfortunate selection. The more you know about the population so you can avoid that, the better. So here, for example, if you can avoid selecting all sections which differ greatly from a lot of sections which do not get selected at all, that could help tremendously. So, stratification is often very helpful.
However, it is just a thought. It sounds like you could then possibly have up to a three-stage sample, which could be too complex for the application. Also, I don't think it was stated as to whether or not a quantitative survey is being done. If so, you can find information in textbooks such as Cochran(1977), 3rd ed, Sampling Techniques, Wiley, and online, for single stage cluster sampling and two-stage sampling, but three-stage sampling starts to become hard to find information on.
Also, if you are doing a qualitative survey, where your thinking might be more in terms of each person being from a simple random sample, then that becomes further from the truth by cluster sampling. You'd have to think more at the primary sampling level (sections).
Do you expect no gender bias in the questions asked? If so I guess stratifying by female and male is unnecessary.
If you do separate the sections into categories, you might not have to call them strata. In stratified random sampling, one tries to get the best overall results, with the lowest overall sample size, and may allocate optimally based on standard deviations. But here, you might just think of it as several different subpopulations, where you are estimating for each one.
But this could still get too messy. If you are satisfied that the sections could be sampled, and by an unfortunate sample selection, it won't be a problem if some 'categories,' or 'types' of sectors are mostly included, and others are left out entirely, then you could go with a one-stage cluster sample, if a census of the selected sections is not difficult, and you use the appropriate variance estimations, etc., or if qualitative, you remember you might think of the sample at the section level.
Anyway, just some thoughts.
Cheers - Jim
Well, sorry to have to back off on the "that sounds reasonable" part a bit, but the difference here is that in stratified random sampling, each different homogeneous group does get sampling drawn from it. But in the cluster sampling case, the very different sections are not all sampled from. (I mean the sections are different from each other, and not all selected.)
In stratified random sampling, you want the variance within each stratum to be minimized, and the differences between strata to be maximized. Since you draw from each stratum, you minimize the chance of drawing an overall 'unfortunate' sample. You get something from each and every different stratum.
But with a cluster sample, if the number if clusters is large, you may not miss anything too important, but you could still miss some important category. The same is true of simple random sampling. Stratification is what helps with better representation, and also is more efficient than simple random sampling, not less efficient, as in the case of cluster sampling.
I think cluster sampling is interesting. Just be aware of the pitfalls.
- Badges
- Science topic
- Similar topics
- Biological Science
- Histology
- Sectioning