Thyroid Research and Practice

: 2019  |  Volume : 16  |  Issue : 2  |  Page : 91--92

Variable selection, unnecessary adjustments, and multicollinearity in thyroid eye disease studies

Red Thaddeus Dela PeÑa Miguel1, Ronald Steven Seno Medalle2,  
1 Medical Researcher, Ottawa, ON, Canada
2 Department of Health - Eye Center, East Avenue Medical Center, Quezon City, Philippines

Correspondence Address:
Dr. Red Thaddeus Dela PeÑa Miguel
Ortigas Center, Pasig City 1605, Philippines

How to cite this article:
Miguel RT, Medalle RS. Variable selection, unnecessary adjustments, and multicollinearity in thyroid eye disease studies.Thyroid Res Pract 2019;16:91-92

How to cite this URL:
Miguel RT, Medalle RS. Variable selection, unnecessary adjustments, and multicollinearity in thyroid eye disease studies. Thyroid Res Pract [serial online] 2019 [cited 2022 Sep 25 ];16:91-92
Available from:

Full Text


We read with interest the study by Lavaju et al. regarding patterns of ocular manifestations in patients with thyroid eye disease (TED).[1] The study is similar to a study we completed last year in a tertiary hospital in the Philippines, which is now being prepared for publication. The study conducted in Nepal carried out a multiple logistic regression model to find out the independent association of variables, such as age and sex to TED. The model employed, however, raises few questions.

First, the study failed to discuss the reasons for selecting the variables included in the multiple regression model. The lack of a theoretical or statistical framework for deciding which covariates to control for in the model may lead readers to assume certain relationships exist. For example, in univariate analysis, the study failed to show a significant difference between groups with or without cardiovascular disease (CVD) in having TED (P = 0.164). We, however, note in [Table 5] that CVD is included in the model. This leads us to assume an underlying relationship with CVD and TED that needs to be accounted for, despite the absence of significant difference in earlier analysis conducted. Furthermore, the failure to report the reason for choosing the variables may leave readers to wonder if including all variables could have caused diluted true associations or spurious associations.[2]

Second, the study explains that TED was diagnosed clinically based on the presence of a thyroid disorder associated with one of the thyroid eye signs. However, Mourits clinical activity score (CAS) was also entered into the model as a risk factor for TED. Given that everyone with CAS has a thyroid disorder, and a higher CAS score will indicate a higher probability of having an eye sign (being a component of CAS), is it therefore not possible that the relationship of CAS and the outcome could have affected the precision of the other variables unnecessarily? Adjusting for both CAS and NOSPECS, given that their components define TED, could have increased the likelihood of accuracy gain or loss in the results of the model.[3] Although the effect is evident in the association of mean CAS score and TED as manifested with a high odds ratio (49.17, P < 0.0001), the effect of using CAS and NOSPECS in the model as confounding variables could have had even more detrimental effects on the association of the other variables with TED.

Finally, the model includes variables with multicollinearity and overlapping variables without adequate investigation. For example, components of CAS are already accounted for in NOSPECS and the model could be deemed to be overadjusting.[4] Further, there could be collinearity between age and duration of disease as well as age and the comorbidities included. Without proper consideration for the relationship between variables, the results could be misleading.[5]

In conclusion, we look forward to the reply of our colleagues from Nepal. Further, we feel that future studies on TED could account for the concerns raised herein.

Financial support and sponsorship


Conflicts of interest

There are no conflicts of interest.


1Lavaju P, Badhu BP, Maskey R. Pattern of ocular manifestations in patients with thyroid disease presenting in Eastern Nepal. Thyroid Res Pract 2019;16:20-5.
2Ranganathan P, Pramesh CS, Aggarwal R. Common pitfalls in statistical analysis: Logistic regression. Perspect Clin Res 2017;8:148-51.
3Schisterman EF, Cole SR, Platt RW. Overadjustment bias and unnecessary adjustment in epidemiologic studies. Epidemiology 2009;20:488-95.
4Leistikow B. Commentary: Questionable premises, overadjustment, and a smoking/suicide association in younger adult men. Int J Epidemiol 2003;32:1005-6.
5Vatcheva KP, Lee M, McCormick JB, Rahbar MH. Multicollinearity in regression analyses conducted in epidemiologic studies. Epidemiology (Sunnyvale) 2016;6. pii: 227.