I would like to know if and how I can join data from a random and non-random sample.
I have responses from two surveys of Japanese municipalities (n = 49 and n = 30), where the total population is ~1700 cases/muncipalities. Response group A (49) resulted from a 300 case random sampling frame from the 1700. Response group B (30) resulted from a convenience sample in a prior study. If I can join them together in a statistically valid way, I could regress the results with greater confidence.
Is there any way to retroactively resample a random sampling frame that includes my response cases, or at least a large number of them?
See:
A Method to Combine Non-probability Sample Data with Probability Sample Data in Estimating Spatial Means of Environmental Variables
Brus, D.J. & de Gruijter, J.J. Environ Monit Assess (2003) 83: 303. https://doi.org/10.1023/A:1022618406507
Dear Hassan,
The question is not about sampling types, but it is about combining non-random data with a random sample.
Mohamed I. Riffi
Do you have regressor data on the entire population? If so, then you could try prediction. Best if you can stratify.
You might find this interesting:
Waksberg Award article:
Brewer, K.R.W. (2014), “Three controversies in the history of survey sampling,” Survey Methodology,
(December 2013/January 2014), Vol 39, No 2, pp. 249-262. Statistics Canada, Catalogue No. 12-001-X.
http://www.statcan.gc.ca/pub/12-001-x/2013002/article/11883-eng.htm
(The purely model-based example is for the classical ratio estimator, but Brewer used many other models in other work.)
Convenience sampling is likely to have missed whole areas (strata) of your populaton, so that would be a problem.
Random and Non-random,
Methodology seems quite different, do not merger the data.
I have the same issue. Arguably, random data mixed with non-random data could approximate randomness. While I agree with Hom that it is better to avoid, in some cases it can be unavoidable and in my case, the data cannot be parsed out again once mixed.
Timothy & David -
Yikes. I just noticed that even the 49 randomly drawn observations actually appear to be the result of an attempted sample size of 300, where there were apparently 251 nonresponses. Attempts at imputation or weight adjustment for that many nonresponses sounds futile, and you therefore don't really seem to have any random sampling.
Also, if the two samples were from different periods, how do you know the populations weren't substantially different?
You might consider this an exploratory study, though unfortunately haphazard in its ability to discover information. If you can find any regressor data on the population, and can stratify or group data into subpopulations, and find regressor data at that level, you might be able to obtain fair inferences. But otherwise you won't have a lot of information. Well, the problem is inference to the population.
If you can stratify/subdivide your population, then even the exploratory study might be more useful.
Best wishes - Jim
- Badges
- Science topic
- Similar topics
- Statistics
- Survey