Independent Component Analysis

We applied an ICA questionnaire data from the UKB. ICA is a statistical decomposition method that separates multivariate data into a predetermined number of statistically independent components. For the ICA we used data from each participants’ answer to a self-report mental health questionnaire included in the UKB data collection. This mental health questionnaire covers questions reflecting various topics relevant to mental health such as depressive symptoms, anxiety, psychological trauma, and questions about substance use. We excluded individuals with a diagnosed psychiatric or neurological disorder (F or G ICD-10 diagnosis except for those with a nerve, nerve root, and plexus disorders, categories G50 to G59) and removed individuals with more than 10% missing answers. This resulted in a sample of 136,678 individuals.

In the feature domain, we used only questionnaire items that were answered by all subjects in the dataset, which excluded follow-up questions on the original items to ensure a robust dataset with minimal missing data. We also excluded questions that referred to symptoms only in the past two weeks prior to assessment to avoid unwanted temporal effects in the analyses. Finally, we excluded items with more than 10% missing answers. This reduced the number of questions in the dataset from about 140 to 43. We imputed missing responses through KNN imputation with k = 3. We subsequently z-score standardized the data.

This dataset of 136,678 individuals and 43 items was then used as input to the ICA. Since ICA requires an a priori decision of the number of clusters, we used a combination of expert knowledge and a data-driven approach implemented in the icasso toolbox^1,2. This provided us with evidence that 13 was the ideal number of independent components. These components each capture a separate domain of mental health contained within the original dataset. The individual loadings from each participant on the 13 independent components would later serve as the input vector for the GWAS described below.

1.

Himberg J, Hyvärinen A, Esposito F. Validating the independent components of neuroimaging time series via clustering and visualization. Neuroimage. 2004 Jul;22(3):1214–22.

2.

J. Himberg, A. Hyvarinen. Icasso: Software for investigating the reliability of ICA estimates by clustering and visualization. In: 2003 IEEE XIII Workshop on Neural Networks for Signal Processing (IEEE Cat No03TH8718). 17AD–2003. p. 259–68.