Notice: On March 31, it was **announced** that Statalist is moving from an email list to a **forum**. The old list will shut down on April 23, and its replacement, **statalist.org** is already up and running.

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

I believe the basic problem here is that you are looking at different kinds of quantities. Remember that the ZINB is a mixture model with a count and a binary component. Your treatment effect from the count component (1.treat = .783026) is conditional on y>0, whereas the marginal effect is the product of the probability of y>0 and the expected count, given y>0. This can easily be seen by doing some manual predictions. Below we set up some toy data with a binary treatment variable such that treated have a higher chance of having a zero count but, conditional on having a positive count, have a lower count than controls, which is similar to your data. So first we simulate some data. This is some modified code that Joseph Hilbe posted here: http://www.stata.com/statalist/archive/2011-06/msg00183.html *----------- *Toy data *----------- clear set obs `=1e4' set seed 1234 gen x1 = rnormal() gen treat = runiform()>.5 //Count component gen xb = 2 + 0.5*x1 - 0.5*treat gen a = .5 gen ia = 1/a gen exb = exp(xb) gen xg = rgamma(ia, a) gen xbg = exb * xg gen nby = rpoisson(xbg) //Binary component gen pi =1/(1+exp(-(0.5*x1 + 0.5*treat + 0.2))) gen bernoulli = runiform()>pi gen zy = bernoulli*nby rename zy y *--------------- *Predictions *--------------- Then we can fit the model: //Model fit zinb y x1 treat, inflate(x1 treat) and obtain marginal counts for the treatment arms: //Marginal counts for control and treatment, using both components margins, at(treat=(0 1)) which we can obtain manually by: //Manual predictions *Count component gen count_c=exp(1.991823 + .543534*x1) gen count_t=exp(1.991823 + .543534*x1 - .4944447) *Binary component gen p_c = 1 - ( exp(.2354296 + .4841011*x1 ) / (1+exp(.2354296 + .4841011*x1) ) ) gen p_t = 1 - ( exp(.2354296 + .4841011*x1 + .5088653 ) / (1+exp(.2354296 + .4841011*x1 + .5088653) ) ) *Product gen control = count_c * p_c gen trt = count_t * p_t sum control trt As we can see, we use parameters from both components to predict the counts. If we wanted to predict counts from only the count component we could type: *Marginal counts for control and treatment, using only count component margins, at(treat=(0 1)) expression(exp(predict(xb))) which is the same as: sum count_c count_t So it is not surprising that your two significance tests differ because in both instances you are dealing with quite different quantities. Joerg On Thu, Dec 13, 2012 at 3:48 PM, Ariel Linden, DrPH <ariel.linden@gmail.com> wrote: > Hi Fellow Listers, > > I am getting conflicting results from the regression output for a covariate > and when I subsequently run margins. More specifically, I ran -zinb- with > the primary covariate of interest being treatment and got a p value of > 0.152. I then ran margins to see the predicted values for treatment and > non-treatment, and then ran contrasts. The contrast indicates that the > difference between treatment and control is statistically significant (p > value = 0.0048). See below for the output... > > Could this be due to the different methods of estimating the p values > between the original regression and margins? If so, how do I reconcile the > two? Obviously, the covariate can't be both significant and not-significant > at the same time. > > Thanks! > > Ariel > > **** code **** > . zinb readmittot $xs [pw= attwt], inflate($xs) robust irr > > <output omitted? > > Zero-inflated negative binomial regression Number of obs = > 7264 > Nonzero obs = > 882 > Zero obs = > 6382 > > Inflation model = logit Wald chi2(9) = > 39.46 > Log pseudolikelihood = -801.1061 Prob > chi2 = > 0.0000 > > ---------------------------------------------------------------------------- > -- > | Robust > readmittot | IRR Std. Err. z P>|z| [95% Conf. > Interval] > -------------+-------------------------------------------------------------- > -- > readmittot | > 1.treat | .783026 .1337402 -1.43 0.152 .5602626 > 1.094361 //treatment is not significant > dxcg | 1.010933 .003087 3.56 0.000 1.004901 > 1.017002 > preadmits | .9912419 .0510683 -0.17 0.864 .8960375 > 1.096562 > ervisits | .9347274 .1392932 -0.45 0.651 .6979733 > 1.251789 > opvisits | .8799214 .1362863 -0.83 0.409 .6495399 > 1.192016 > age | 1.001473 .0055586 0.27 0.791 .9906373 > 1.012427 > sex | 1.045554 .1525712 0.31 0.760 .7854813 > 1.391736 > cci_score | .9187872 .0707482 -1.10 0.271 .7900797 > 1.068462 > gap | .9636616 .0135773 -2.63 0.009 .9374147 > .9906434 > _cons | .2499155 .1165989 -2.97 0.003 .1001526 > .6236259 > -------------+-------------------------------------------------------------- > -- > inflate | > 1.treat | .3987922 .4116639 0.97 0.333 -.4080542 > 1.205639 > dxcg | -.2056051 .0401806 -5.12 0.000 -.2843577 > -.1268526 > preadmits | -.2580604 .2825054 -0.91 0.361 -.8117609 > .2956401 > ervisits | -.3233463 .3895552 -0.83 0.407 -1.086861 > .4401679 > opvisits | -.1520791 .4182457 -0.36 0.716 -.9718256 > .6676674 > age | .0327528 .0177659 1.84 0.065 -.0020678 > .0675734 > sex | .2615432 .4124321 0.63 0.526 -.546809 > 1.069895 > cci_score | -.337497 .2392857 -1.41 0.158 -.8064884 > .1314944 > gap | -.1037818 .0423225 -2.45 0.014 -.1867325 > -.0208311 > _cons | 1.603676 1.316273 1.22 0.223 -.976172 > 4.183524 > -------------+-------------------------------------------------------------- > -- > /lnalpha | -1.116844 .5762305 -1.94 0.053 -2.246235 > .0125467 > -------------+-------------------------------------------------------------- > -- > alpha | .327311 .1886066 .1057968 > 1.012626 > ---------------------------------------------------------------------------- > -- > > . margins treat > > Predictive margins Number of obs = > 7264 > Model VCE : Robust > > Expression : Predicted number of events, predict() > > ---------------------------------------------------------------------------- > -- > | Delta-method > | Margin Std. Err. z P>|z| [95% Conf. > Interval] > -------------+-------------------------------------------------------------- > -- > treat | > 0 | .1214688 .0067881 17.89 0.000 .1081643 > .1347733 > 1 | .0858302 .0094371 9.09 0.000 .0673339 > .1043266 > ---------------------------------------------------------------------------- > -- > > . margins r.treat > > Contrasts of predictive margins > Model VCE : Robust > > Expression : Predicted number of events, predict() > > ------------------------------------------------ > | df chi2 P>chi2 > -------------+---------------------------------- > treat | 1 7.97 0.0048 //treatment is significant > ------------------------------------------------ > > -------------------------------------------------------------- > | Delta-method > | Contrast Std. Err. [95% Conf. Interval] > -------------+------------------------------------------------ > treat | > (1 vs 0) | -.0356385 .0126276 -.0603881 -.010889 > -------------------------------------------------------------- > **************end code********************** > . * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/faqs/resources/statalist-faq/ * http://www.ats.ucla.edu/stat/stata/

- Prev by Date:
**re: Re: st: Why does a non-statistically significant covariate in a a regression model become significant in margins?** - Next by Date:
**Re: st: issue with forvalues/foreach** - Previous by thread:
**st: how to estimate nonlinear MLE** - Next by thread:
**st: How to aggregate daily data in Stata** - Index(es):