We had estimated 600 samples for one study with national representation. In some districts due to the purpose of pilot intervention we collected additional 196 samples with same tools and methods. Would it be appropriate to include all the samples for national estimation?
As course, most data, means better, because most the quantity of information that will contribute to the model's quality. The primary data not necessarily have to be the result of a previous planning, but rather they can be the result of the normal operation of a process ( passive experiment). You would have an intermediate situation: almost a passive experiment.
Some data may be data redundant which can be taken in larger sphere of sampling. The smaller subset can be taken for initial sampling.
In a little improved English: As course, while more data, better,, because more quantity of information will contribute to the model's fitting. The primary data not necessarily have to be the result of a previous planning, but rather they can be the result of the normal operation of a process ( passive experiment). You would have an intermediate situation: almost a passive experiment.
Even in the case of executing an experimental plan, it is always advisable of carrying out at least. 2 or 3 replicas.
It depends on what type of analysis you are planning to do. If some patterns are specific for the oversampled area, you might wrongly conclude that these patterns are more general because they are overrepresented in the whole sample. See the topics of stratified resampling of databases.
Assuming the supplemental sample was from the same population, it certainly can be combined. How it is combined is completely determined by the details of the sampling design. Are more details about the design available?
Thank you all for your feedback and suggestions.
Dear William, our study population is drug shops. We don't have a sampling frame at the national level. However we wanted a nationally representative sample. We selected 25 out of 75 districts purposely to balance the ecological and regional allocation. In each district the sample framework is prepared by the field researchers. From each district we enroll 16 sample randomly i.e 400 in total. Of the 25 districts, 4 were oversampled district from where we included 49 additional sample with same methodology. This gives an additional 196 sample. The purpose for oversampling was for setting the benchmark indicators for the pilot intervention that we are planning to do.
Hi,
it seems that you already made a stratified sampling since you tried to balance regional variation by collecting the same number of samples from districts. Thus, I think that in order to exclude the possibility of overrepresentation of any unmeasured but spatially autocorrelated effect it would be better not to include the additional samples from the oversampled districts. Without the additional samples, potential regional bias due is perfectly controlled, no reviewer would criticise it, and I guess that you still have a sufficiently large sample.
@Deepak Joshi,
I think you might need to consider mixtures of distributions, because each of the two samples might have come from the same family of distributions, but with different parameters.
Good luck.
Dear Mohammed, samples do not overlap, this was ensured in the methodology.
I think it is better to use all data in the same time by using Big data techniques.
You (almost) have a straightforward stratified sample. Different sample sizes in different strata, but in general, that is not a problem at all. It might be a less than optimum, but that rarely happens in practice anyway. Your problem is that the strata sizes are apparently unknown. Are there any any approximations available?
Without any knowledge of the population, you won't know whether you have anything resembling a "representative" sample. And you have no way to construct a stratified estimate. Surely some strata size information can be found? How about using the population size in each of the strata areas -- may well be a usable proxy.
The concern is that the pilot samples are (in some way) sampling a different population than the full study. Maybe the pilot survey was done just prior to a very successful national ad campaign. The campaign changed the proportion of people that use that type of product. Say the survey was on those people who drink orange juice. The pilot survey was sampling a population where 80% of the people had at least one glass of orange juice per day. After the campaign, half the people that used to drink orange juice now drink cranberry juice. So the pilot and full surveys have very different populations.
One solution is to break the analysis into parts.
1) Analyze the full survey as one part.
2) Reanalyze the data with the pilot survey results included. Add a blocking variable coded as 0 if there is no pilot data for that area, 1 for pilot data, 2 for survey data where pilot data is also present. I would delete block=1, and ask if there was a significant block effect (is block=0 any different from block=2)? I would delete block=0 and ask if there was a significant block effect. If the answer in both cases is no, I would combine the data.
I definitely agree that blocking makes a lot of sense here. While at Virginia Tech, I attended another graduate student's defense, and was appalled at his conclusion that interspecific competition was important among his studied stream zoobenthos. Because N was very large, there was statistical significance, but low R2 and thus little (if any) biotic significance. I responded to the student with concerns about pseudoreplication, and would’ve preferred that such data were analyzed by blocking.
-Bob Vadas, Jr.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics