I have a logistic regression model as follows:
dependent ~ var1 + var2 + var3 + ... + var24
The results of the model show some significant beta estimates, but the upper and lower bound for the beta estimates are in opposite directions for almost all the variables.
For example, for var4, the estimate is -0.118, the upper estimate is 1.709, and the lower estimate is -1.999. The directions are opposite, and they imply completely different effects of the variable.
What is going on here? The only pattern I noticed is that when the p-value is < 0.05, the upper and lower estimate are in the same direction, and not necessarily so when the p-value is >= 0.05.
Variable | Estimate | Std. Error | z value | P value | 2.5 % | 97.5 % |
---|---|---|---|---|---|---|
var1 | -10.5633455339251 | 0.759594371480191 | -13.906561094365 | 5.77920204964665e-44 | -12.0565995298244 | -9.07846678375401 |
var2 | 0.834244972702608 | 0.0597828513393095 | 13.9545865413425 | 2.95007544494396e-44 | 0.717795996175918 | 0.95218596536246 |
var3 | 0.670686357092985 | 0.21666036968767 | 3.09556546063234 | 0.0019643800732445 | 0.247189216918815 | 1.09664152539369 |
var4 | -0.11827182131671 | 0.944136983515969 | -0.125269768456973 | 0.900309984899174 | -1.99985360814008 | 1.70919144266964 |
var5 | -0.838742349613382 | 1.44030127660015 | -0.582338128306907 | 0.560338947926719 | -3.99745084094769 | 1.7778193195088 |
var6 | 0.00351026911347644 | 0.00375769899895975 | 0.934153883652786 | 0.350224520205816 | -0.0035456482198036 | 0.0113524222954051 |
var7 | -0.000731414874584177 | 0.00487809942760052 | -0.149938492529651 | 0.88081314233484 | -0.00975234521520494 | 0.00976533653881936 |
var8 | -1.19552692531913e-05 | 0.00673205576541774 | -0.0017758719876631 | 0.998583059903854 | -0.0116320385616292 | 0.0163799646989692 |
var9 | 0.000137595011787968 | 0.00472745479443718 | 0.0291055161330949 | 0.976780436425101 | -0.008764713087581 | 0.00983780537950456 |
var10 | -0.137314541163203 | 0.0393174449097 | -3.49245841072767 | 0.00047859611846024 | -0.213673700147479 | -0.0594908796696738 |
var11 | 0.0279370642051698 | 0.0238522441522973 | 1.17125516688454 | 0.241496226532665 | -0.0188823270944542 | 0.0746333490175093 |
var12 | -0.00395247678315438 | 0.00214161113078487 | -1.84556230883328 | 0.0649557834759472 | -0.00817664917521618 | 0.000219701951803874 |
var13 | 1.21509909187451 | 0.0984290947580135 | 12.3449178808544 | 5.18871561184112e-35 | 1.02156762761588 | 1.40749581880397 |
var14 | -0.00451544526175419 | 0.00394966580859281 | -1.14324742410623 | 0.252935877243301 | -0.0122986061668043 | 0.00318802869388361 |
var15 | 0.0135134262292429 | 0.0126865181256071 | 1.06518006717436 | 0.286794451982887 | -0.0112553514129102 | 0.0384819312617267 |
var16 | -0.0138442704104665 | 0.00732732910516601 | -1.88940201972173 | 0.0588379804731442 | -0.028234664566532 | 0.00049389410510635 |
var17 | 0.0111374372976773 | 0.0203408759664438 | 0.547539708518485 | 0.584007997795982 | -0.0288601526385988 | 0.0508878526572194 |
var18 | 0.0149310101117362 | 0.00811568787987997 | 1.8397713579834 | 0.0658018121433485 | -0.000971969252617372 | 0.0308463266255572 |
var19 | 0.00311471128042081 | 0.00948300727316004 | 0.328451849787824 | 0.74257004491251 | -0.0155133744502076 | 0.0216618395316565 |
var20 | 0.00585217320183963 | 0.0143809113296191 | 0.406940357791264 | 0.684051793713323 | -0.0222225404766296 | 0.0341636722917656 |
var21 | 0.0068148700357023 | 0.0120561599031068 | 0.565260422097266 | 0.571896644381092 | -0.0164317427744267 | 0.0308539253978295 |
var22 | 0.835607905844263 | 0.104779189340344 | 7.97494150417639 | 1.52453009251868e-15 | 0.629139007916138 | 1.03997790608976 |
var23 | -0.0073298137446894 | 0.00467548343417065 | -1.56771248318829 | 0.116948247068106 | -0.0165766110799228 | 0.00175730977838965 |
var24 | -0.000266749353306101 | 0.00806360445460414 | -0.0330806595000816 | 0.973610265791254 | -0.0161678639143166 | 0.0154513944600894 |
Best Answer
Your "upper/lower bounds" form 95% confidence intervals. Their interpretation is: if you were to repeat your experiment many times, calculating 95% CIs each time, then 95% of these CIs would contain the true parameter (assuming your model specification is correct). Yes, that is a very cumbersome definition, but it is the best that frequentist statistics can do.
There indeed is a direct connection between p values (from standard two-sided t tests) and CIs: the CI does not contain zero exactly if $p<.05$.
In your case, most of your parameters have CIs that overlap zero. In classical parlance, the parameter estimates are not significantly different from zero. That is: you cannot reject the null hypothesis that the true value of these parameters is zero.
Also, you have a huge model. If you estimate no less than 24 parameters, I truly hope you have multiple thousands of data points. Anything less will be very overparameterized.