Statalist The Stata Listserver

[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

st: Re: Construct Null Datasets through Bootstrap Resampling

From   "Michael Blasnik" <>
To   <>
Subject   st: Re: Construct Null Datasets through Bootstrap Resampling
Date   Thu, 30 Nov 2006 14:54:22 -0500

It seems like you want something similar to -permute- except you want to permute all of the variables, not just one. Here's a short program that will scramble all of the variables in your dataset. You can usel this within your own program and perform a monte carlo simulation (check out -simulate-).

program define scramble
version 9.2
syntax [varlist]
tempvar hold order rand
gen long `order'=_n
foreach var of local varlist {
gen `rand'=uniform()
gen `hold'=`var'
sort `rand'
replace `var'=`hold'[`order']
drop `rand' `hold'

Michael Blasnik

----- Original Message ----- From: "Erik Ingelsson" <>
To: <>
Sent: Thursday, November 30, 2006 1:31 PM
Subject: st: Construct Null Datasets through Bootstrap Resampling

Dear Statalist users,

I am trying to construct null data sets through bootstrap resampling, to be able to account for multiple testing in genetic analyses. I would like to sample my genotypes and phenotypes randomly with replacement (without keeping linked them together as the original observations in my dataset), and then run regressions on these samples to evaluate a distribution of minimum probability values. Thereby, I will obtain empirical p-values by comparing the nominal p-values with the distribution of probability from the null data sets. I have seen this implemented in SAS, but I hope that it could be done also in STATA.

In my ?trail-and-error? approach, I have come so far, that I have learned how to use bootstrap sampling to get a new dataset with p-values from a set of regressions. However, these simulations are still using my original observations (although creating new samples), while I would like the observations to be randomly created from the available variables (not keeping them together as in the original dataset). Below is the code for what I have done so far. Variables linj001-linj004 are the genotypes, thus the important independent variables; the other variables are covariates to adjust for; stset and all definitions, etc are done above. In reality, I will have much more regressions to include in the simulation, but this is just for learning how to do it.

*Program with the commands to be run in all bootstrap samples*
capture program drop myboot
program myboot, rclass
stcox linj001 whoht70 adadiab70 ami70 vit70 z972 z290 z085 zekg_lvh
return scalar p1 = 2*(1-normal(abs(b[linj001]/_se[linj001])))
stcox linj002 whoht70 adadiab70 ami70 vit70 z972 z290 z085 zekg_lvh
return scalar p2 = 2*(1-normal(abs(b[linj002]/_se[linj002])))
stcox linj003 whoht70 adadiab70 ami70 vit70 z972 z290 z085 zekg_lvh
return scalar p3 = 2*(1-normal(abs(b[linj003]/_se[linj003])))
stcox linj004 whoht70 adadiab70 ami70 vit70 z972 z290 z085 zekg_lvh
return scalar p4 = 2*(1-normal(abs(b[linj004]/_se[linj004])))

*Run the program in the original sample*
ret list

*Bootstrapping in 10000 samples*
bootstrap "myboot" p1=r(p1) p2=r(p2) p3=r(p3) p4=r(p4), reps(10000) saving(C:\bootstrapsample) replace

This leaves me with a dataset (C:\bootstrapsample) which consists of the p-values from the 4 regressions derived from 10000 simulations. However, this is not exactly what I need, since the variables are still ?connected? in the original observations (and then randomly chosen for my simulated sets). I would like to get simulations with all variables scrambled, so that new observations with all variables scrambled are created in a number of bootstrap simulations, and then used for regressions. The present macro can give me 10000 simulated p-values for the regressions, based on samples with replacement, but these simulations are reusing the actual 2000 observations from the original dataset. Now I would like to create a ?null dataset?, in which I instead of sampling from the observations in the real dataset, I would like Stata to randomly ?make up? observations from the existing variables and values, so that I have 2000 fake observations (with random selection of all variables) to base the regressions on, in 10000 simulations.

I have read the help files, manual, Statalist, searched at Internet, even asked Technical Support (which helped me to come this far, but not the last part). I am using Stata 8.2 for Windows. Is there a way to do this? Did I explain what I want to do properly? Is there anyone who can help me with this?

Thanks a lot in advance,
Erik Ingelsson

Erik Ingelsson, MD, PhD

Current affiliation (until June 30, 2007):
Framingham Heart Study
73 Mt. Wayte Avenue, Suite 2
Framingham, MA 01702-5827
Phone: 508-935-3453
Fax: 508-626-1262
Cell: 508-202-8493

Permanent affiliation:
Uppsala University, Department of Public Health and Caring Sciences, Uppsala
Science Park, SE-751 85 Uppsala, SWEDEN.
Fax: +46-18-611 79 76

* For searches and help try:
*   For searches and help try:

© Copyright 1996–2024 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index