You said, "we adjust our results for large number of factors".  That makes me wonder if you have a problem with multicollinearity.  One effect of multicollinearity is inflation of the standard errors (which translates to wider CIs).  What is the width of the CI for the crude odds ratio (i.e., with no adjustment for any other variables)? 

You also said "good sample size also result in such type of results".  But it's having too small a sample size that yields large SEs and wide CIs. (Maybe I misunderstood what you were trying to say there.)  Anyway, what is your sample size?  And how many degrees of freedom (df) are used up by the predictors in your full model (with all potential confounders included)?

HTH.

Dear Wilhelm and Weaver

Thanks a lot for your valuable suggestions

I am not good at statistics so am apprehensive for the results with large intervals. We are carrying large epidemiological study with cases and controls of more than 2000 sample size but when I analyze some epidemiological data in relation with different genes sample size decreases as you stated sometimes because of the rare condition. Tell me, do I have to ignore such results.

For example

We are analyzing some of the xenobiotic genes in combination with environmental risk factors

this is the example I am providing.. see the last row of each table how CI changes drastically.

Results before adjustments:

------------------------------------------------------------------------------

        case | Odds Ratio   Std. Err.      z    P>|z|     [95% Conf. Interval]

-------------+----------------------------------------------------------------

 _Iepx4cig_2 |   1.296509   .1992926     1.69   0.091     .9592531    1.752337

 _Iepx4cig_3 |   .7754223   .3377649    -0.58   0.559     .3301907    1.821007

 _Iepx4cig_4 |   .4594805     .12694    -2.81   0.005     .2673664    .7896366

 _Iepx4cig_5 |   .9782273   .3596762    -0.06   0.952     .4758484    2.010995

 _Iepx4cig_6 |   .6778495   .9812753    -0.27   0.788     .0397108    11.57066

------------------------------------------------------------------------------

------------------------------------------------------------------------------

        case | Odds Ratio   Std. Err.      z    P>|z|     [95% Conf. Interval]

-------------+----------------------------------------------------------------

     epx4cig |   .9127948   .0572526    -1.45   0.146      .807205    1.032197

------------------------------------------------------------------------------

After adjusting our results with variables like age, sex, residence, religion, education,  ethinicity,  alcohol,  income,  nass,  mca cat,  fruitveg.

   Iepx4cig_2 |   1.447175   .3549335     1.51   0.132     .8948618     2.34038

 _Iepx4cig_3 |   .6563501   .4767597    -0.58   0.562     .1580664    2.725408

 _Iepx4cig_4 |   .7337014   .3564449    -0.64   0.524     .2831332    1.901288

 _Iepx4cig_5 |   1.759279   1.095768     0.91   0.364     .5189955    5.963565

 _Iepx4cig_6 |   .3288945   .9888205    -0.37   0.711     .0009077    119.1769

------------------------------------------------------------------------------

        case | Odds Ratio   Std. Err.      z    P>|z|     [95% Conf. Interval]

-------------+----------------------------------------------------------------

     epx4cig |   1.050009    .110431     0.46   0.643      .854419    1.290373

such type are results are very rare in our analysis, so tell me could we include such type of results or drop down them.

Thank you

Sumaiya,

  It sounds like you have some uncertainty with respect to the predictive model you should use.

  One option in your case might be to model average the candidate models using Burnham and Anderson's (2004) method.  Given the level of instability in the estimates you show above, the "smoothing" averaging can provide could be helpful.

  Model averaging will probably not reduce the standard error for the ORs, but will incorporate the model selection uncertainty into your estimates and, arguably, produce more realistic estimates and standard errors than relying on any one model.  Thus, you don't have to drop any models - you might be better served incorporating aspects of many models (see Burnham and Anderson, 2004).

 You can get model averaged coefficients using the miinc package on Stata's SSC which works with logistic (which is the command I assume you're using).

- joe

Burnham, K. P., & Anderson, D. R. (2004). Multimodel inference understanding AIC and BIC in model selection. Sociological methods & research, 33(2), 261-304.

Hello Sumaiya.  Have you looked at any multicollinearity diagnostics?  Judging from the output you pasted, you are using Stata.  The page linked below has a section on multicollinearity.  Note that multcollinearity is concerned with relationships among the explanatory variables only, so even though you are using logistic regression for your main analysis, you can use linear regression (if necessary) before issuing the vif post-estimation command as shown on that page.  (I don't use Stata enough to know if vif works after logistic regression, or only after linear regression.)

The second page linked below suggests another possibility:  It says you can use the program collin, and provides a download link. 

HTH.

http://www.ats.ucla.edu/stat/stata/webbooks/reg/chapter2/statareg2.htm

http://www.ats.ucla.edu/stat/stata/webbooks/logistic/chapter3/statalog3.htm

Thank you all.

I will read the supplementary material suggested by you people.