Hello. I have serveral problems obtaining percentile intervals via bootstrap in panel data: Assuming a TxN panel with T observations across N groups and T>>N. I want to apply the two step Fama/MacBeth procedure (FM), so I have to look, if the mean of the estimated slopes over time is significant in the end of the second step (via t-stat with some strong assumptions). To get percentile intervals for the mean I want to bootstrap the whole panel. First problem here: Is that idea of resampling the whole panel to obtain the interval for the mean right or is it enough to resample the time series of slopes in the end of the second step? When I want to resample the whole panel I think resampling across time is better, because T>>N, so I will identify the block size l (Politis, White: Automatic Block Length Selection For Dependent Bootstrap, 2003) of each of the N-series of length T (l1, l2,...,lN). Second problem here: In the end of block-length selection I will have l1,...,lN block lengths, so what is the right block length for the whole table? - Is it max(l1,...,lN)? After obtaining a so called global block size for the whole panel I want to resample blocks along time with all N observations across groups. I use the global block size to sustain the dependency structure along time and across groups. After that resampling step I do my FM-stuff and store the mean and start the process again with resampling and so on and so on. In the end I have stored many many means and I can calculate my percentile intervals, standard errors and so on for the mean of the slopes. Third Problem here: Before obtaining the mean as my desired statistic for inference, I have to calculate other things (two step approach), so are my results useful in a bootstrap inference sense? - I mean is the resampling process directly connected to my desired statistic and so the bootstrap-interval reliable? I hope you can help me. Sincerely, Andreas P.s.: Paper which handles a little bit the bootstrap-panel-stuff, but not explicitly the global block size stuff: Kapetanios, A Bootstrap Procedure For Panel Data Sets With Many Cross-Sectional Units, 2008