My ordinal regression results in very large odds ratios

[Hans Sadeghi]

Hey there,
I'm working on data from my patients (n=150)
I have almost 30 variables and an ordinal outcome with 4 categories (liver steatosis) (I also have 2 other similar outcomes)

I don't know why when I perform an ordinal regression with backward elimination, I receive ORs as large as even 8000!

Does anybody know what am I doing wrong?

Here is a sample of my data:

---

* Example generated by -dataex-. For more info, type help dataex
clear
input byte(Gender Married Employment University_degree Age_group BMI_cat) double(ALK_p INR) byte(Emotionaleater Sono_Grade Bx_Steatohepatitis)
1 2 1 1 3 5   175 1.01 2 2 1
1 . . . 1 6   217    1 . 3 0
1 . . . 1 6     .    . . 0 0
1 2 1 1 1 6     .    1 1 1 0
1 . . . 2 6   181    1 . 2 1
1 2 1 1 2 6   197  1.2 2 2 2
1 1 1 2 1 6    92    1 1 0 1
1 2 1 1 3 6   210    1 2 2 1
2 2 2 1 2 4   203    . 1 0 1
1 . . . 3 5   144    . . 2 0
1 . . . 2 6   170    . . 2 0
1 . . . 2 4   141    1 . 1 0
2 1 2 1 1 6   209  1.1 1 1 1
1 2 1 1 3 6   124    1 2 0 1
2 1 2 1 1 6   202    1 1 3 1
1 2 1 1 3 6   171    1 2 2 2
1 2 1 1 2 6   217    1 2 1 0
1 1 1 1 1 6   176    1 2 1 0
1 . . . 2 6    77  1.1 . 2 0
2 . . . 3 6   178  1.3 . 2 1
1 2 2 1 2 5   150    1 2 3 1
1 1 2 2 2 4    17    1 2 1 1
1 . . . 1 6   163    1 . 3 1
1 . . . 3 6   158  1.2 . 3 .
1 . . . 2 5   261    1 . 1 0
1 2 1 1 1 4   142 1.24 2 2 1
2 . . . 1 6     . 1.13 . 2 0
1 2 1 1 1 6   188    . 2 0 0
1 2 1 1 3 6   151    1 2 1 1
1 . . . 2 4   274 1.31 . 0 1
1 . . . 1 5   219    . . 0 .
1 2 1 1 2 6   144    1 2 0 1
1 . . . 1 6   224    1 . 2 0
2 1 2 1 2 6   194  1.1 1 2 1
1 . . . 4 5   177    . . 1 .
1 . . . 1 6   156 1.05 . 0 0
2 1 2 1 1 6   266 1.23 1 3 1
1 1 1 1 1 6     .    . 2 0 .
1 2 1 1 3 6   211    1 2 3 1
1 2 2 2 4 5   129   .1 2 2 1
2 1 2 1 2 5   128    1 2 3 0
2 . . . 1 5   137    . . 2 1
1 . . . 2 6   146  1.5 . 1 .
1 2 1 1 1 6   228    1 2 2 1
2 2 2 1 4 6   305    1 1 2 0
2 . . . 2 6   252  1.4 . 3 .
1 2 1 1 2 5   122    1 1 2 1
2 2 2 1 2 6     . 1.08 2 3 .
2 2 2 1 2 6     .    1 1 2 1
2 2 2 1 3 5   171    1 2 0 1
2 2 2 1 3 6   133    . 1 3 .
1 2 1 1 4 5   181    1 1 3 1
1 2 1 1 3 4     .    . 2 3 .
1 2 1 1 4 6     .    . 1 3 .
2 2 2 1 2 4     .  1.1 2 2 0
2 2 2 1 2 6   245    1 2 3 1
1 1 2 1 1 4   121    1 2 2 1
2 2 2 1 3 6     . 1.11 1 3 1
1 2 1 1 3 5    54  .86 2 2 1
1 2 2 2 3 5   246    . 1 2 .
1 1 2 1 1 6   134    1 2 1 .
2 2 2 1 1 6    93    1 2 3 1
1 1 2 1 1 5   171    1 2 2 1
1 2 1 2 2 5    66    . 2 2 .
2 2 2 2 2 5   106    1 1 2 .
2 1 2 1 2 5   245  1.1 2 3 .
2 1 2 1 2 6   179    1 1 3 .
2 1 2 1 2 5     .    1 1 2 .
1 1 2 1 1 5   124    1 2 1 1
1 1 1 1 1 5   150    . 1 1 1
1 1 1 1 4 6   250    1 1 2 .
1 2 1 1 4 6   190    1 1 2 2
1 1 1 1 1 6   246    1 2 3 .
1 2 1 1 4 6     .    1 1 1 0
1 1 1 1 2 6   136    . 2 2 .
2 2 2 1 2 5    83    1 2 2 .
2 2 2 1 2 6   177 1.17 2 0 1
1 2 1 1 3 5   194    1 2 2 1
1 2 2 1 2 6     .    1 2 3 1
1 2 2 2 2 6   185    1 1 1 0
1 2 2 1 1 5   176    1 2 2 0
1 2 1 1 2 6   240    1 2 2 1
1 2 1 1 1 6   127    1 2 1 1
2 2 2 1 3 6   181    1 1 3 0
1 2 1 1 1 5 252.5    1 2 2 0
2 1 2 1 2 6   142  1.1 2 1 .
1 1 1 2 1 5   197    1 1 0 .
2 1 2 1 2 6   254    1 1 2 0
1 1 1 1 4 6   179  1.1 2 1 .
1 2 1 1 3 6   249    1 2 2 .
1 2 2 1 2 6   131    1 2 2 .
1 2 1 1 4 6   316  1.1 2 0 .
1 2 1 1 4 6   179    1 2 2 2
1 2 1 1 4 5   231    1 2 2 1
1 1 2 1 2 6   154    1 2 0 .
1 2 1 2 2 6    18    . 1 1 1
1 2 2 2 2 6   137    1 1 1 .
2 2 2 1 3 6   189    1 2 2 .
1 1 1 1 2 6   194    1 2 2 .
1 1 2 1 2 6   176    1 1 1 .
end
label values Gender labels0
label def labels0 1 "Female", modify
label def labels0 2 "Male", modify
label values Married labels1
label values University_degree labels1
label def labels1 1 "No", modify
label def labels1 2 "Yes", modify
label values Employment labels2
label def labels2 1 "No", modify
label def labels2 2 "Yes", modify
label values Age_group labels3
label def labels3 1 "19-30", modify
label def labels3 2 "31-40", modify
label def labels3 3 "41-50", modify
label def labels3 4 ">50", modify
label values BMI_cat labels4
label def labels4 4 "Obesity class I", modify
label def labels4 5 "Obesity class II", modify
label def labels4 6 "Obesity class III", modify
label values Emotionaleater labels14
label def labels14 1 "No", modify
label def labels14 2 "Yes", modify
label values Bx_Steatohepatitis labels19
label def labels19 0 "Minimal", modify
label def labels19 1 "Mild", modify
label def labels19 2 "Moderate", modify

---

This is my code:

stepwise, pr(0.3) : ologit Bx_Steatohepatitis i.Age_group i.BMI_cat Gender Married Employment University_degree Smoke Alcohol T2DM Hypo_Thyroid HLP HTN CVD Sweeteater Volumeeater Emotionaleater Snackernibbling Cancer_F HTN_F T2DM_F i.Sono_Grade SGOT SGPT ALK_p, or

---

And this is my result:

Bx_Steatohepatitis	Odds ratio	Std. err.	z	P>z	[95% conf.	interval]
Age_group
31-40	3.997381	4.217089	1.31	0.189	.5055795	31.60542
>50	50.54159	87.94102	2.25	0.024	1.669479	1530.09
2.Sono_Grade	2.85757	2.601264	1.15	0.249	.4798913	17.01574
CVD	71.35374	172.0147	1.77	0.077	.6329919	8043.32
ALK_p	.9761676	.0107919	-2.18	0.029	.9552434	.99755
Gender	16.21916	29.71174	1.52	0.128	.4474235	587.9463
Married	.0912319	.1329141	-1.64	0.100	.0052486	1.585796
Employment	.0104316	.019311	-2.46	0.014	.0002771	.3927363
University_degree	.0146786	.0255267	-2.43	0.015	.0004858	.4435643
Smoke	.0063704	.0115658	-2.78	0.005	.0001814	.2236534
SGPT	1.036996	.0231966	1.62	0.104	.9925136	1.083472
T2DM	.134202	.2045136	-1.32	0.188	.0067701	2.660268
HTN_F	16.57867	19.52767	2.38	0.017	1.647922	166.7871
HLP	.039629	.0843666	-1.52	0.129	.0006108	2.571254
HTN	30.4704	40.1952	2.59	0.010	2.296207	404.3388

[Clyde Schechter]

First, I recommend you read https://www.stata.com/support/faqs/s...sion-problems/ for a compelling exposition of why stepwise regression is a bogus procedure that should not be used.

That said, even in a single regression, you are taking a modest sample of 150 and chopping it up into a large number of groups: 4 outcome levels * 4 age groups * 3 (or more?) BMI categories * 2 sexes * ... Even in the optimistic scenario where everything spreads into these bins as evenly as possible, they are going to be very small. In that situation, logistic regression coefficients are strongly biased upwards (in magnitude: large negative coefficients are also seen). You are simply asking too much of the available data.

For more reasonable results, use a simpler model with fewer predictors. Choose them thoughtfully based on prior research and reasonable causal models of the data generating process, rather than trying to throw in "the kitchen sink." Or, get a much larger data set that can accommodate fine-grained subdivision.