Filtering and model-based analysis independently improve skin-conductance response measures in the fMRI environment: Validation in a sample of women with PTSD

Numerous methods exist for the pre-processing and analysis of skin-conductance response (SCR) data, but there is incomplete consensus on suitability and implementation, particularly with regard to signal filtering in conventional peak score (PS) analysis. This is particularly relevant when SCRs are measured during fMRI, which introduces additional noise and signal variability. Using SCR-fMRI data (n = 65 women) from a fear conditioning experiment, we compare the impact of three nested data processing methods on analysis using conventional PS as well as psychophysiological modeling. To evaluate the different methods, we quantify effect size to recover a benchmark contrast of interest, namely, discriminating SCR magnitude to a conditioned stimulus (CS+) relative to a CS not followed by reinforcement (CS-). Findings suggest that low-pass filtering reduces PS sensitivity (Δd = −20%), while band-pass filtering improves PS sensitivity (Δd = +27%). We also replicate previous findings that a psychophysiological modeling approach yields superior sensitivity to detect contrasts of interest than even the most sensitive PS method (Δd = +110%). Furthermore, we present preliminary evidence that filtering differences may account for a portion of exclusions made on commonly applied metrics, such as below zero discrimination. Despite some limitations of our sample and experimental design, it appears that SCR processing pipelines that include band-pass filtering, ideally with model-based SCR quantification, may increase the validity of SCR response measures, maximize research productivity, and decrease sampling bias by reducing data exclusion. •Skin-conductance response data filtering and processing methods are inconsistent.•We compare filter effects on peak-score and model-based response contrasts.•Low-pass filtering reduces and band-pass filtering increases contrast effect size.•Model-based analysis was more than twice as sensitive as peak-scoring.•Preliminary evidence suggests filtering reduces exclusion of “non-responders”.