I am in a situation where sampling proportional to some weight might be very beneficial. In order to estimate $\tau$ the population total / sum I use the Hansen Hurwitz estimator. That is,

$$\hat{\tau}_p = \frac{1}{n}\sum_{i=1}^n\frac{y_i}{p_i}$$

with variance

$$\hat{Var}(\hat{\tau}_p) = \frac{\frac{1}{n}\sum_{i=1}^n(\frac{y_i}{p_i}-\hat{\tau}_p)^2}{n-1}$$

where $p_i$ are the sampling weights and $y_i$ the observations

How can I estimate my sample size? I have a set of $\hat{y}_i$ values that are extremely well correlated ($r>0.9$) with $y_i$. Normally I would just solve the variance equation for $n$, but I didn't manage to do that. Therefore, I simulated a few million sample draws and some variation around my estimates for $y_i$. However, for some large sample sizes (in my case 55-60) the power drops (from 0.97 to 0.9) and for $n>60$ rises to $1.0$. I am uncomfortable relying on the results with - what I feel are - such strong abnormalities.

Further, the weighting of the elements in the sample frame is very strong ($p_i$ ranging from $0.115$ to $1.9*10^{-6}$). Since I am sampling with replacement I am have several elements being selected multiple times. Up until now I just used a unique() function to get each element once. Does this bias the results?


You should first determine the desired CV. Then you can determine the n needed. It is not necessary to eliminate the duplicate units for the estimation (if the $\hat{y}_{i}$ are highly correlated with the $y_{i}$ is logical that the sample size is very small).


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.