iiinova — where your digital self keeps evolving

Gabe Loewinger

@gabeloewinger.bsky.social

101 followers 166 following 24 posts

Machine learning research scientist @ NIMH interested in statistics, optimization, ML, neuroscience, Brazilian jiu-jitsu, cats.

Posts Replies Media

Gabe Loewinger @gabeloewinger.bsky.social · 18/03/2025

FLMM finds effects obscured by standard analyses! For example, FLMM reveals effects that “wash out” when analyzed with AUCs. In published work, Cue Period AUC finds no effects because it averages over time-windows (1) and (2) that have opposing effects. 12/13

100

Gabe Loewinger @gabeloewinger.bsky.social · 18/03/2025

FLMM can disentangle components with distinct temporal dynamics. It can also be used to run analogues of standard hypothesis tests (e.g., ANOVAs, correlations) at each trial timepoint. Below is an example akin to the FLMM version of a paired t-test. 11/13

100

Gabe Loewinger @gabeloewinger.bsky.social · 18/03/2025

Informally, functional random-effects allow one to model variability across animals in the signal “shape.” 10/13

100

Gabe Loewinger @gabeloewinger.bsky.social · 18/03/2025

Functional random-effects allow one to model how the dynamics of signal-covariate associations vary across animals. 9/13

100

Gabe Loewinger @gabeloewinger.bsky.social · 18/03/2025

FLMM plots can be conceptualized as pooling signal values (dF/F) at a given trial time-point (e.g., 1.7 sec) across animals and trials, correlating it with covariate(s) (e.g., Latency-to-press) and plotting the slope of the correlation. 8/13

100

Gabe Loewinger @gabeloewinger.bsky.social · 18/03/2025

FLMM outputs a coefficient estimate plot that shows how the signal– covariate association evolves across trial timepoints. 7/13

100

Gabe Loewinger @gabeloewinger.bsky.social · 18/03/2025

FLMMs exploit autocorrelation to construct *joint* 95% CIs (light grey) that show time windows where effects are statistically significant (any intervals that do not contain 0). All you need to do is visually inspect! 6/13

100

Gabe Loewinger @gabeloewinger.bsky.social · 18/03/2025

FLMM combines the benefits of 1) Mixed Models to account for between-animal heterogeneity, and 2) Functional Regression to model effects at each trial timepoint. 5/13

100

Gabe Loewinger @gabeloewinger.bsky.social · 18/03/2025

Solution: We propose an analysis framework based on Functional Linear Mixed Models (FLMM) that allows one to analyze signal–covariate associations at every trial timepoint. 4/13

100

Gabe Loewinger @gabeloewinger.bsky.social · 18/03/2025

Problem: Photometry is often applied in nested longitudinal experiments with multiple trials per session and sessions per animal. This induces correlation, missing data, etc., that can obscure effects if not accounted for statistically. 3/13

100

Gabe Loewinger @gabeloewinger.bsky.social · 18/03/2025

Problem: Common photometry analysis methods reduce detection of effects because, among other things, they average across trials and use summary statistics (e.g., AUC, peak amplitude). 2/13

100

Gabe Loewinger @gabeloewinger.bsky.social · 18/03/2025

Test effects of behavior/events at every trial timepoint in photometry analyses! Paper with Erjia Cui, Dave Lovinger, Francisco Pereira. “A Statistical Framework for Analysis of Trial-Level Temporal Dynamics in Fiber Photometry Experiments.” Python+R packages! elifesciences.org/articles/95802. 1/13

4288