In my point of view, considering non-response rate in final sample size calculation for longitudinal study is important because refusal of participants may occur at any stage of the study and ensuring adequate number of sample size is important by considering non-response rate (it can decrease time, economy and energy wastage). However, in case of cross sectional study, considering non-response rate for final sample calculation is not recommended because we can go for the next if the first participant refuses to participant.
Importantly, I saw that some studies add the 10% non response rate to the effective sample size i.e effective sample size = 384 and non response rate =10%, so final sample size = 384*0.1 +384 =422.4. Is it right? Look the link it says different. Article Sample Size estimation and Power analysis for Clinical resea...
I need a concrete conclusion regarding my question.
Your conclusion should be based on previous studies.Concrete justification is key.If you find that this to be very subjective,then report it as a limitation of your study.
The suggested article was retracted. You can obviously go for the next if the first participant refuses to participate. But this could induce selection bias. Therefore, the revised sample size is n* = n / response rate. E.g., n* = 384/0.90 = 427 based on your calculation.
Here is a discussion on it: https://www.researchgate.net/post/How-to-adjust-sample-size-for-non-response-in-cross-sectional-studies
Your conclusion should be based on previous studies.Concrete justification is key.If you find that this to be very subjective,then report it as a limitation of your study.
Taklo -
Replacing a nonrespondent by someone willing to respond means that your sample will represent the population of all those willing to respond, not the actual complete population. You could try weighting responses by considering "response propensity" groups. I think the US Census Bureau, at least at some time, has replaced households they could not contact by counting a similar household twice, where similarity is based on a number of characteristics. But if your nonresponse rate is high, you are still going to have dubious results. This is especially true if you have few, or worse, no respondents representing a particular segment of the population. (Even in cutoff sampling with 'prediction,' you do have regressor data on everyone.)
You could research "weighting for nonresponse," such as the case of "response propensity," and "imputation for nonresponse," such as the "hot deck" method. If you have regressor data on the entire population, you could use "prediction" for imputation, which would also give you the estimated variance of the prediction error. (Note that "prediction" does not mean "forecasting." It just means estimation when you are using regression, and thus you have a random variable.)
Best wishes - Jim
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Sampling Design
- Sample Size