kimhenry posted on Thursday, August 28, 2003 - 3:10 pm
I'm working on a dataset that has two sampling weights. The first weight adds up to the nominal sample size of 1,200 while the second weight adds up to an effective sample size of 860 because it takes into account the extra variance in estimates due to the variability of the weights. I've been told that if you don't account for the extra variability due to weighting the estimates of the standard errors are too low. In the mplus manual, it states that the weights should add up to the total number of observations. Should I therefore use the first weight that adds up to 1200 (the total number of observations)? If so, will my standard errors be underestimated?
If the (unweighted) sample size is 1200, that's what Mplus will use regardless of which weights are used (I assume that the two sets of weights are proportional). If the weights truly have some measurement error and the user wants to adjust for that by reducing the sample size to 860, the user will have to do this by hand. Each SE estimate should be multiplied by square root of 1200/860.
Jinni Su posted on Tuesday, August 21, 2012 - 2:22 pm
I ran a multilevel regression model with a sample size of over 13,000 in my dataset, however, in the output file, it says that the number of observations is 6,488 (the model estimation terminated normally). I did not use any command such as useobs to restrict my sample for analysis. Could you privide some insight regarding how/why would this happen?