Hello;
> First off, am I correct to assume that the 95% confidence intervals
> should overlap if the predicted medians are not significantly different?
> Or can a non-significant variable have 95% confidence intervals that
> don't overlap (i.e. even though gender was not significant, the 95% confidence intervals can still NOT overlap)?
With means, if confidence intervals do not overlap, you can be sure
that predicted means are significantly different, but confidence
intervals can overlap and predicted means still be significantly
different.
I have no experience with quantile regression, but in standard
regression formulas for the standard errors of beta coefficients (from
where one establishes that the effect of the variable is significant
or not) and standard errrors of the predicted means are not the same,
so it comes as no surprise that the confidence interval of the beta
coefficient includes 0 (the coefficient is non-significant) but the
confidence intervals of the predicted means do not overlap. You could
check standard errors formulas in quantile regression.
Best regards,
Angel Rodriguez-Laso
2009/1/14 I M <[email protected]>:
> Hello,
>
>
>
> Thank you in advance.
>
> I have run some analyses on my data using quantile
> regression (essentially, quite similar to linear regression, except that
> it uses least absolute values to calculate the medians rather than means.)
>
>
>
> Suppose I have a categorical variable in my models (e.g. female versus
> male), and in the regression models, the coefficients are not
> significant (i.e. Female gender is not significantly associated with
> worse outcomes compared to males, p-value is greater than 0.05).
>
>
>
> Now, I then ran the predict command to give me the adjusted values for
> this regression model. When I determine the 95% confidence intervals of
> the median for these predicted values, I find that the 95% confidence
> intervals do not overlap.
>
>
>
> First off, am I correct to assume that the 95% confidence intervals
> should overlap if the predicted medians are not significantly different?
> Or can a non-significant variable have 95% confidence intervals that
> don't overlap (i.e. even though gender was not significant, the 95% confidence intervals can still NOT overlap)?
>
>
> If I am doing things wrong completely, then please correct me!
>
>
>
> ---------------
>
>
>
> What I did:
>
> qreg [outcome variable] [predictor 1, e.g. female versus male]
> [predictor2] [predictor3] // this was my quantile regression model
> including my covariates
>
> predict, xb // this is the linear prediction command that gave me my
> adjusted values for the outcome variable
>
> by [predictor1], sort: centile xb // this gave the median and 95% CI for
> my adjusted values for the outcome variable, sorted by predictor 1 (e.g.
> by female versus male)
>
>
>
> Interpretation:
>
> the qreg command gave me the preregression model and coefficients.
> Assume that predictor1 (e.g. female versus male) was not significant.
>
> the predict command gives me the adjusted medians.
>
>
>
> Is this the correct way to obtain adjusted values according to my
> regression model, and the correct way to obtain 95% CI?
>
>
>
> Thank you immensely!
>
> *
> * For searches and help try:
> * http://www.stata.com/help.cgi?search
> * http://www.stata.com/support/statalist/faq
> * http://www.ats.ucla.edu/stat/stata/
>
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
