General rule (for any sampling method): the less information, the more uncertainty.

The stratified sampling reduces the costs for data collection but increases the work for planning and analysis for keeping the uncertainty within an acceptable level.

Yes, I think that stratified sampling is a very good method, specially when strata are very different in means.

Thanks a lot for your help!

Yes, I understand that I have to plan carefully to avoid mistakes. I will probably have to take care, that I do not over- or underrepresent one class (strata).

In case anyone is interested in this: I found this paper helpful:

S. V. Stehman and R. L. Czaplewski. Design and analysis for thematic map accuracy assessment: fundamental principles.

1998.

Hello

stratified sampling has the highest accuracy among sampling methods.

In research, this type of sampling is preferred to other methods.

Among its disadvantages are the following:

1) It takes more time than cluster sampling

2) This type of sampling is more expensive than cluster sampling because it has to be sampled from all floors.

Lisa -

You say you are taking a sample of your models? Does this mean you are making a collection, like a collection of chairs, not collecting quantitative data? The main advantage of stratified random sampling is that if you know enough about your data that you can stratify in such a way as to minimize variance within strata and maximize differences between strata, you can increase efficiency (i.e., lower overall variance). Also, i think a kind of "representativeness" might be more likely to be achieved, whereas simple random sampling is more vulnerable to collecting an "unfortunate" sample, not much like the population. The goal is to obtain aggregate level data most efficiently for the overall population. If you want to publish information on each stratum, you are not really doing stratified random sampling, just simple random sampling for each of a number of subpopulations.

The key for making stratified random sampling work for you is optimal allocation, so you need to know how to stratify best. But in your case, if you are not looking for quantitative data, it sounds like you just want to obtain a set of things which you think might be, in a way, "representative" of a population of things. Proportional allocation of a stratified random sample might do that well. However, you then just really need categories for strata. If you don't have categories, or even if you do, you might want to use sequential sampling rather than random sampling, at least for convenience, depending upon how your records are arranged.

So, are you talking about quantitative data or a list of 'things'?

Cheers - Jim

By the way, for quantitative data, stratified random sampling can be more efficient than simple random sampling which is more efficient than cluster sampling (efficient, meaning lowest variance for a given overall sample size, but likely not in terms of data collection), but for continuous data there is also unequal probability sampling to consider, and there is multi-stage sampling, and double sampling, and also if you have auxiliary/regressor data on the whole population, you can consider model-assisted and model-based approaches.