Skip to content

[R-meta] About choosing the reference

2 messages · 英文科陳品誠, Wolfgang Viechtbauer

#
Dear all,

       When performing meta-regression analysis, I use this code to change
turn a variable into a factor:

dat$duration <- relevel
<http://stat.ethz.ch/R-manual/R-devel/library/stats/html/relevel.html>(factor
<http://stat.ethz.ch/R-manual/R-devel/library/base/html/factor.html>(dat$duration),
ref="S")

       In my data, the duration moderator has four values including S
(short), M (medium), L (long), and NA (not provided). And I tried to figure
out the interaction between this moderator and another moderator named
"Posttest type" with two values of M (meaning) and U (use). The problem I
face now is that with different ference I used,  the coefficiency will
differ, for example,
(a) I use M as reference for duration, and this is the result

                          estimate       se        zval          pval
     ci.lb           ci.ub

durS:posU  2.3309  0.8690 2.6824 0.0073  0.6278  4.0341   **


(b) I use N/A as reference for duration, and this is the result

                                                        estimate
 se             zval          pval           ci.lb            ci.ub

durS:posU  1.7890  1.4294 1.2516 0.2107 -1.0126  4.5906


(c) I use L as reference for duration, and this is the result


                            estimate       se              zval
 pval           ci.lb            ci.ub

durS:posU  1.2827  0.9275  1.3829 0.1667 -0.5352 3.1005


So, eventually, which one should I choose to be the final result? Is
there any criteria to choose between the three?


Nick


??? (Nick Chen)
Email: t571 at wlgsh.tp.edu.tw <t5741 at wlgsh.tp.edu.tw>
#
Dear Nick,

Please do not post in HTML format. See what this can do to posts -- makes things quite unreadable. See also the notes on the mailing list page: https://stat.ethz.ch/mailman/listinfo/r-sig-meta-analysis

All of these results just reflect different contrasts of the same two-way interaction. So there is not one that is inherently preferable. This is not an issue specific to meta-analysis. You would face the same issue in any regression model with such a two-way interaction. If the two-way interaction as a whole is significant (based on a test of all three interaction terms, so the chi^2- or F-test with df=3), then one could examine four simple contrasts (between M and U of pos for each of the four levels of dur).

Best,
Wolfgang