Confusion of Odds Ratio (comparison of OR from epidemiology table and logreg)

[Zhianni Yang]

Hello.

Suppose that I have a set of data where I want to know the association between Sex and Social Media Disorder (SMD). My codebook for Sex and SMD is this way:

Coding	Variable
0	Female
1	Male
0	SMD
1	Not SMD

and my epidemiology, 2x2 table looks like this one:

Sex	SMD	NOT SMD
Female	90	198
Male	38	93

The odds ratio should be: Female/Male= 1.112, right? (manually calculated), meaning that the odds for female to have SMD is 1.112 times higher than male

However, when I try to use logistic regression under this command, I found that something is not right

logistic RECODE_KATSMD ib1.SEX

where RECODE_KATSMD is my variable for SMD

which yields the result as below:

logistic RECODE_KATSMD ib1.SEX

Logistic regression        Number of obs     =    419
        LR chi2(1)        =    0.21
        Prob > chi2       =    0.6432
Log likelihood =  -257.7637        Pseudo R2         =    0.0004

            
RECODE_KATSMD  Odds Ratio   Std. Err.    z    P>z     [95% Conf.    Interval]
            
SEX 
Perempuan     .8989249   .2073968    -0.46    0.644     .5719225    1.412894
_cons       2.447368     .471196       4.65    0.000     1.678093    3.569296
            
Note: _cons estimates baseline odds.

Which is a little bit confusing because I use male as a references, but the odds ratio for female is 0.898 (different than the one I calculated manually from the epidemiology table)

I try to do:

adjust, by(SEX) exp

it returns:

   
SEX          exp(xb)
    
Female       2.2
Male        2.44737
    
Key:    exp(xb)  =    exp(xb)

meaning that the odds for women= 2.2, and men= 2.447 and when i calculate odds ratio female/male= 2.22/2.44737 is 0.898, which is also different from the OR we calculated manually before

I am now confused and feeling dumb.

Thank you very much

[Ken Chui]

I agree with the hand calculation, the output is not matching your calculation. However, there isn't enough Stata code for me to decide if you have coded something incorrectly. But the code is where I'd scrutinize.

Check this yourself.

clear
input nosmd male freq
0 1 38
1 1 93
0 0 90
1 0 198
end
generate smd = 1 - nosmd

label define l_male 1 "Male" 0 "Female"
label define l_nosmd 1 "No SMD" 0 "Yes SMD"
label define l_smd 1 "Yes SMD" 0 "No SMD"

foreach x in male nosmd smd{
    label values `x' l_`x'
}

* NOSMD as dependent
* Female to male
logistic nosmd ib1.male [fweight=freq], base

* Male to female
logistic nosmd ib0.male [fweight=freq], base


* SMD as dependent
* Female to male
logistic smd ib1.male [fweight=freq], base

* Male to female
logistic smd ib0.male [fweight=freq], base

My guess is probably due to sloppy recoding when creating that "RECODE_KATSMD", you didn't show that part of the code, so I can't be sure.

There are a few tips:

• Do not play mental gymnastics with yourself. If you mean for SMD yes to be the outcome, code that as 1; not 0. Do not even entertain the idea of having a "no smd" variable where no = 1 and yes = 0. Those double-double negative can be extremely confusing.
• Don't name your variable "sex". Instead, call it either male (male = 1 and female = 0) or female (female = 1 and male = 0). That way you will not get confused.
• Use label scheme, and then use the ", base" option in regression command to also print the reference group.
• Beware if you have implemented any self-looping code. Such as:

generate nosmd = smd
replace nosmd = 1 - nosmd

In this case, if the analyst carelessly just ran the second line again, it'd flip the coding by accident. Use more secure recoding that wouldn't loop back to itself, such as:

recode smd (1 = 0)(0 = 1), gen(nosmd)
* OR:
generate nosmd = 1 - smd

[Paul Rathouz]

I think the problem is not with your predictor, but with your response. You have "1" corresponding to NOT SMD, so -logistic- is modeling the log odds for "not SMD" instead of the log odds of "SMD". That's my guess in any case. -- P

[Zhianni Yang]

Hello Ken and Paul, thanks for noticing. I tried to recode again for the variable SMD as follows:

recode 1=0 0=1

and that gives me a codebook of:

                           128         0  NOT SMD
                           291         1  SMD

I tried to run logreg again, but the difference does not exist between my previous and current code:

logistic RECODE_KATSMD ib1.Male, base

Logistic regression                             Number of obs     =        419
                                                LR chi2(1)        =       0.21
                                                Prob > chi2       =     0.6432
Log likelihood =  -257.7637                     Pseudo R2         =     0.0004

------------------------------------------------------------------------------
RECODE_KAT~D | Odds Ratio   Std. Err.      z    P>|z|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
        Male |
 Female  |   .8989249   .2073968    -0.46   0.644     .5719225    1.412894
  Male  |          1  (base)
             |
       _cons |   2.447368    .471196     4.65   0.000     1.678093    3.569296
------------------------------------------------------------------------------
Note: _cons estimates baseline odds.

Which also, gives no differences.

For RECODE_KATSMD, I recode that one in SPSS and move it to STATA, so I dont us coding

[Zhianni Yang]

I just run some code and I thought the problem was that STATA use (0) as the baseline for dependen variables

I tried to run this:

mlogit RECODE_KATSMD ib1.Male, base baseoutcome(1)

Iteration 0:   log likelihood = -257.87097  
Iteration 1:   log likelihood = -257.76372  
Iteration 2:   log likelihood =  -257.7637  

Multinomial logistic regression                 Number of obs     =    419
LR chi2(1)        =    0.21
Prob > chi2       =    0.6432
Log likelihood =  -257.7637                     Pseudo R2         =    0.0004

    
RECODE_KAT~D       Coef.   Std. Err.      z    P>z     [95% Conf.    Interval]
    
NOT_SMD      
Male 
Perempuan     .1065557   .2307165     0.46   0.644    -.3456402    .5587517
Laki-Laki            0  (base)

_cons   -.8950131   .1925317    -4.65   0.000    -1.272368    -.5176578
    
SMD            (base outcome)

and later I run again my code

logistic KAT_SMD ib1.Male, base

Logistic regression        Number of obs     =    419
        LR chi2(1)        =    0.21
        Prob > chi2       =    0.6432
Log likelihood =  -257.7637        Pseudo R2         =    0.0004

            
KAT_SMD  Odds Ratio   Std. Err.    z    P>z     [95% Conf.    Interval]
            
Male 
Perempuan      1.11244   .2566582    0.46    0.644     .7077671    1.748489
Laki-Laki            1  (base)

_cons    .4086022   .0786689    -4.65    0.000     .2801673    .5959147
            
Note: _cons estimates baseline odds.

[Zhianni Yang]

Edit: now i tried to close my stata and run the code again and it returns to the previous one. I feel like I am getting deceived

[Zhianni Yang]

everyone sorry for causing confusion. After consulting to my fellow researcher I was so dumb and did a mistake in my codification of Sex, hence explain why the hand calculation does not make sense with my stata output. I learnt this the hardest way. Thank you everyone for helping (@Paul, @Ken).