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

From |
Joseph Coveney <jcoveney@bigplanet.com> |

To |
Statalist <statalist@hsphsun2.harvard.edu> |

Subject |
Re: st: pkcross vs. proc mixed |

Date |
Sun, 03 Apr 2005 14:34:59 +0900 |

SR Millis wrote: I've been using SAS proc mixed to analyze AB/AB cross-over designs (both mixed and fixed effects models). Out of curiosity, I compared results with Stata's pkcross--using the Chow & Liu data (used in the Reference manual example). I'm getting very different results in SAS vs Stata, eg, SAS is finding a significant treatment effect while Stata isn't. What could be accounting for the differences? -------------------------------------------------------------------------------- Perhaps if you posted your SAS code for this example, someone might be able to assist. The dataset is balanced, so PROC MIXED (with the default REML method) should be giving you the same result as PROC GLM and PROC ANOVA. Just by inspection, there doesn't seem to be a treatment effect. Joseph Coveney clear set more off use http://www.stata-press.com/data/r8/chowliu.dta * Pooling for inspection generate delta = period2 - period1 replace delta = -delta if seq == 2 ttest delta = 0, level(90) // P = 0.54, when pooled bysort seq: ttest delta = 0, level(90) // Nothing hiding here drop delta * pkshape id seq period1 period2, order(ab ba) * Inspection redux--compare treat 1 period 1 with treat 2 period 2 * and treat 2 period 1 with treat 1 period 2 table treat period, contents(mean outcome sd outcome n outcome) format(%3.0f) * forvalues parameterization = 1/4 { pkcross outcome, param(`parameterization') sequence(sequence) /// treatment(treat) carryover(carry) period(period) id(id) } * Using the third parameterization, anova outcome sequence / id|sequence treat treat*sequence * Or, equivalently, anova outcome sequence / id|sequence treat period lincom _b[treat[2]] - _b[treat[1]], level(90) * Random effects, FGLS (REML, here) xi: xtreg outcome i.sequence i.period i.treat, i(id) re sa theta adjust _Isequence_2 = 0 _Iperiod_2 = 0, by( _Itreat_2) xb se lincom _b[_Itreat_2], level(90) // Note z = t above; use as // t test statistic (with df), and P will be the same * Random effects, full maximum likelihood xi: xtreg outcome i.sequence i.period i.treat, i(id) mle nolog estimates store A adjust _Isequence_2 = 0 _Iperiod_2 = 0, by( _Itreat_2) xb se quietly xi: xtreg outcome i.sequence i.period, i(id) mle nolog lrtest A ., stats // Slightly anticonservative, as expected, but . . . * estimates drop _all keep id sequence outcome treat reshape wide outcome, i(id) j(treat) xi: mvreg outcome1 outcome2 = i.sequence lincom ([outcome2]_cons + 0.5 * [outcome2]_Isequence_2) - /// ([outcome1]_cons + 0.5 * [outcome1]_Isequence_2), level(90) // Same * Equivalently, quietly manova outcome1 outcome2 = sequence matrix Representative_sequence = (1, 0.5, 0.5) matrix Treatment_difference = (1, -1) manovatest , test(Representative_sequence) ytransform(Treatment_difference) exit * * For searches and help try: * http://www.stata.com/support/faqs/res/findit.html * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/

- Prev by Date:
**st: re: count variable** - Next by Date:
**Re: st: re: count variable** - Previous by thread:
**st: pkcross vs. proc mixed** - Next by thread:
**Re: st: pkcross vs. proc mixed** - Index(es):

© Copyright 1996–2016 StataCorp LP | Terms of use | Privacy | Contact us | What's new | Site index |