iiinova — where your digital self keeps evolving

David Baranger @davidbaranger.bsky.social · 06/10/2026

Wow, looks like exactly the right special issue for me!

020

David Baranger @davidbaranger.bsky.social · 08/09/2026

Data visualization is storytelling. After Elspeth Kirkman: www.instagram.com/reel/DcSlOws...

Plot of four charts, each showing time vs age, in years. Each plot shows a different aspect - actual (linear), reported (staircase), development (log age), and perception of time passage (log time).

032

David Baranger @davidbaranger.bsky.social · 24/03/2026

And if you're worried that we possibly missed an important regional association, we did some post-hoc analyses to show that more than 90% of all brain-wide associations with substance use (thickness, surface area, and volume) are explained by global thickness!

ortical Thickness Accounts for the Majority of Substance Use Associations.
A) The p values of associations between all available drug variables and brain measures before controlling for global brain thickness are pictured. Most of the significant drug variable associations are also strongly associated with global brain thickness, including mAUDIT-C and lifetime marijuana use. B) Analyses pictured in Panel A were repeated with global cortical thickness as an additional covariate.

Colors indicate drug type. Shades of each color indicate a dimension of use (e.g., dependency, lifetime use, drug test). Individual points represent the p value of the association between a drug variable and a brain measure. Points outlined in black indicate survival of multiple test correction.

010

David Baranger @davidbaranger.bsky.social · 24/03/2026

Comparing within- and between-families, we see evidence for predispositional risk with marijuana use (significant between family effect and genetic correlation), as well as evidence for exposure effects (possibly bi-directional) for both alcohol and marijuana (significant within-family effects).

Drug Use Associations With Brain Structure Reflect a Combination of Predispositional Risk and Exposure Effects.
A) Regression estimates of within- and between-family drug use variables fit to predict whole brain thickness are pictured. Y-axis variables are split by mAUDIT-C and Marijuana Use. X-axis data are split by within- (left) and between- (right) family estimates. Color denotes the sample included in the analysis. B) Regression estimates of within- and between-family whole brain thickness fit to predict substance use variables are pictured. Y-axis variables are split by mAUDIT-C and Marijuana Use, the two drugs evidencing unique-effects on whole brain thickness. X-axis data are split by within- (left) and between- (right) family whole brain estimates. Color denotes the sample included in the analysis. Bold indicates significance. C) Standardized regression estimates of environmental variance (green) and heritability (pink) fit to predict mAUDIT-C, Marijuana Use, and Brain Thickness are pictured. D) Variance component correlations of additive genetics (blue) and non-shared environment (red) between mAUDIT-C and Marijuana Use are pictured. Points reflect estimates, lines indicate 95% confidence intervals. MJ Use = Marijuana Use. Bold indicates significance.

100

David Baranger @davidbaranger.bsky.social · 24/03/2026

Looking across a several dimensions of substance use, age of onset of alcohol use, as well as lifetime marijuana and tobacco use, are also associated with global thickness. But only marijuana shows a unique effect over and above alcohol use (and vice versa)!

Shared- and Unique-Drug Associations with Attenuated Brain Thickness.
Standardized regression estimates of alcohol, marijuana, tobacco, and illicit drug use variables predicting whole brain cortical thickness. Lines reflect the 95% confidence interval of the estimate. Note that these dimensions were not standardized across drug type, as collection of these data varied. Bold indicates survival of multiple test correction and evidence for shared-effects on global brain thickness. Starred indicates drug-specific, unique effects on global brain thickness.

100

David Baranger @davidbaranger.bsky.social · 24/03/2026

In the HCP, which includes many people with moderate-severe alcohol use and polysubstance use, we replicate associations of hazardous alcohol use with many brain structures (onlinelibrary.wiley.com/doi/full/10....). But, these effects are largely attributable the association with global thickness!

A) Composite mAUDIT-C scores ranged from zero to 12 and were categorized into four levels of past-year hazardous risk: low (green; 0 to 3), moderate (yellow; 4 to 5), high (blue; 6 to 7), and severe (pink; 8 to 12). B) Color indicates the number of drugs used across the lifetime. Participant endorsement of having ever used any of the four major substance types are pictured. 65% (n = 726) of the sample endorsed a pattern of polysubstance use throughout the lifetime (i.e., 2+ substances).

Global Brain Thickness Explains Regional Alcohol Effects on Brain Structure.
Standardized regression estimate associations between brain ROI and mAUDIT-C are pictured. Brain structures on the y-axis are split by modality, i.e., volume, thickness, and global measures. X-axis panels are split before (left, blue) and after (right, red) controlling for global brain thickness (grey). Bold indicates significance following multiple test correction.

Hazardous Alcohol Use Predicts Attenuated Global Brain Thickness.
Association of hazardous alcohol use (mAUDIT-C) with global cortical thickness, residualized for covariates and standardized (i.e., SD=1). Covariates included age, age2, sex, socioeconomic status, level of educational attainment, intracranial volume, and sibling status. Shading reflects the density of overlapping points. mAUDIT-C was uniquely associated with attenuated global brain thickness (β = -0.12, p < 0.001).

100

David Baranger @davidbaranger.bsky.social · 25/11/2025

𝐘𝐨𝐮𝐭𝐡 𝐂𝐨𝐫𝐫𝐞𝐥𝐚𝐭𝐞𝐬 𝐨𝐟 𝐆𝐞𝐧𝐞𝐭𝐢𝐜 𝐋𝐢𝐚𝐛𝐢𝐥𝐢𝐭𝐲 𝐭𝐨 𝐒𝐮𝐛𝐬𝐭𝐚𝐧𝐜𝐞 𝐔𝐬𝐞 𝐃𝐢𝐬𝐨𝐫𝐝𝐞𝐫𝐬. New from us, led by @sarahepaul.bsky.social. PheWAS in ABCD identifies many potentially modifiable substance use risk factors! www.medrxiv.org/content/10.1...

081

David Baranger @davidbaranger.bsky.social · 14/11/2025

Great work led by Andrew Castillo extending sample size stability analyses to interactions! We've also added a function implementing these analyses to the InteractionPoweR R package: dbaranger.github.io/InteractionP... #rstats

073

David Baranger @davidbaranger.bsky.social · 11/11/2025

Reliability and sample size have a non-linear relationship. Linear increases in reliability yield diminishing reductions in the required sample size.

040

David Baranger @davidbaranger.bsky.social · 27/10/2025

Shout out to BEAR Lab research tech Daniella Fernandez (who only joined 4 months ago!) whose poster abstract was selected for a nanosymposium at the annual meeting of the Upper Midwest Chapter of SFN this weekend! #neuroskyence #MRI 🧠🍺

060

David Baranger @davidbaranger.bsky.social · 28/04/2025

Reliability of a difference score vs. reliability of an interaction term.

272

David Baranger @davidbaranger.bsky.social · 02/02/2025

I am reminded of similar concerns regarding candidate genes (again, see my early work) and similarly wonder how much signal vs. noise would be thrown out. The approach I describe is a global analysis, and I think our results support the hypothesis that that is a reasonable approach for this task...

Spatial map correlations with task performance and cognition. Spatial map correlations between individual difference measure associations with task activation and the 2-back>0-back main effect of task, standardized regression estimates. Pearson correlation coefficient (r) and the robust p-value are shown below each plot.

Spatial map correlations with psychopathology. Spatial map correlations between individual difference measure associations with task activation and the 2-back>0-back main effect of task, standardized regression estimates. Pearson correlation coefficient (r) and the robust p-value are shown below each plot.

120

David Baranger @davidbaranger.bsky.social · 01/02/2025

But what if you don't have thousands and thousands of participants? This is where neural signatures really shine. We find that you only need a couple hundred participants to train a model that achieves near-peak performance!

131

David Baranger @davidbaranger.bsky.social · 01/02/2025

How do signatures compare with a standard machine learning analysis? Using the full sample, we compared to an elastic-net model, finding that the signature was slightly worse at predicting cognition, but better at predicting psychopathology! The two approaches also capture non-overlapping variance.

The neural signature achieves similar performance to machine learning models trained on each outcome. A) The absolute value of the standardized regression estimate from models predicting each outcome using the neural signature or a linear elastic net regression machine learning (ML) algorithm. The dashed line indicates the diagonal. Colors reflect the type of individual difference measure. Bolded points indicate the model fit statistics for the neural signature and ML models significantly differ from each other after false discovery rate correction for multiple comparisons at p<0.05. B) Standardized regression estimates from models that included both the neural signature and predictions from a linear elastic net regression machine learning (ML) algorithm. Lines reflect the 95% confidence intervals. Bolded points survive false discovery rate correction for multiple comparisons at p<0.05. Note: ML model associations are largely positive, as models were separately trained to predict each outcome. Directionality of neural signature associations (e..g, positive for cognition and negative for psychopathology) are consistent with associations when ML model predictions are not included in regressions. These are examined in the section Neural signature associations reflect correspondence with the main effect of task.

110

David Baranger @davidbaranger.bsky.social · 01/02/2025

But what do neural signature associations mean? Intriguingly, they correspond to the correlation between the main effect of task map and the individual difference (BWAS) map. So they tell you whether regions that change with the task are positively or negatively correlated with your trait!

Neural signature associations reflect spatial alignment with the main effect of task. A) Main effect of task, 2-back>0-back, colored by the standardized regression estimate. B) An example regional association map for attention problems (association of the 2-back>0-back regional contrast with the CBCL attention problems score), finding that attention problems are associated with less activation of key fronto-parietal working memory regions. The map is colored by the standardized regression estimate. C) Example spatial map correlation, between attention problems regional associations and the 2-back>0-back main effect of task, standardized regression estimates (r=-0.86, probust=1.7x10-4; plots for all outcomes are shown in Supplemental Figures 7&8). Attention problems are associated with less activation in regions that are activated by the 2-back condition (relative to 0-back), and conversely, more activation in regions that are deactivated by the 2-back condition. D) Neural signatures associations (regression estimates from Figure 3) and C (the spatial map correlation) for attention problems. E) Rank-ordered neural signature associations and rank-ordered map correlations. F) Neural signatures associations and the percent of individual regional associations that survive false discovery rate (fdr) multiple test correction. D-F) Colors reflect the type of individual difference measure (orange=Cognition, green=N-back performance, purple=Psychopathology).

131

David Baranger @davidbaranger.bsky.social · 01/02/2025

What's more, the neural signature predictions were *more* reliable than activation in individual regions and *more* sensitive to individual differences across 33 measures!

Reliability of the Neural Signature and Individual Regional Activation. The split-half reliability (comparing runs within one scan) and longitudinal stability (comparing two scans) was examined for neural signature predictions during the 0-back condition, 2-back condition, and their difference (2-back > 0-back). Red points are the neural signatures, with 95% confidence intervals indicated by horizontal lines. ICC=intraclass correlation coefficient. Shaded blue regions represent the distribution of reliabilities across all regions (n=167) for each of the corresponding conditions and contrasts. Analyses were restricted to unrelated participants. A) Split-half reliability in the baseline session, N=6,551. B) Split-half reliability in the year 2 follow up, N=5,386. C) Longitudinal stability across the two waves, N=4,111.

The neural signature is associated with individual differences in behavior, cognition, and psychopathology. The neural signature: positively correlates with A) N-back task performance, B) positively correlates with NIH toolbox task performance and composite measures, and C) negatively correlates with psychopathology. Shaded yellow represents the distribution of effects across all regions. Blue points reflect neural signature association estimates with 95% Confidence intervals (horizontal lines). Estimates reflect standardized regression coefficients.

111

David Baranger @davidbaranger.bsky.social · 01/02/2025

But does this preliminary evidence generalize? Is this an approach that anyone with task fMRI data could use? To find out, we trained a working memory signature in the ABCD study (N=9,024). The classifier performed well and captured the known neurobiology of working memory.

Plots showing the performance of a classifier trained on working memory fMRI data. A) Main-effect of the 2-back > 0-back contrast in the N-back task, using 13,477 observations from 9,024 participants. Colors reflect the t-statistic of the test. B&C) An elastic-net classifier distinguishing high and low working memory conditions was trained. Model performance - area under the receiver operating characteristic curve (AUC) - was comparable across balanced split-halves in both the training splits (B; Split 1 AUC=0.884; Split 2 AUC=0.872) and testing splits (C; Split 1 AUC=0.877; Split 2 AUC= 0.867). D) A Haufe-transformation of variable (brain region) weights identified the contribution of individual regions to model performance. Plot shows the correlation of the haufe-transformed variable importance with the main-effect of the task in both splits. Points are individual regions. E) Visualization of the haufe-transformed variable importance of each brain region, averaged across splits. Variables with larger absolute weights contribute more to the model’s predictions. F) Histogram of model predictions for each MRI scan for each of the two conditions (purple: 0-back; orange; 2-back), combined across data splits. Values reflect the predicted probability that a scan is from the 2-back condition, given a forced choice between 2-back and 0-back. G) Model predictions for the two conditions for each individual scan (points), colored by density (number of overlapping points). The white line reflects the correlation between predictions for the two conditions. H) Histogram of the difference between the model predictions for the two conditions. Values reflect how separable the two conditions are for each scan. Positive values indicate that the 2-back condition had a higher 2-back probability than the 0-back condition .

100

David Baranger @davidbaranger.bsky.social · 01/02/2025

Task fMRI is very useful for localization and understanding neural computations, but less so for studies of differences *between* people, largely because it is less reliable than we thought (Elliott 2020 is the canonical citation pubmed.ncbi.nlm.nih.gov/32489141/)

Example simulated data, showing that there can be large replicable within-subject effects, but these do not necessarily yield reliable difference scores. Plots are of box plots, with individual points shown.

130

David Baranger @davidbaranger.bsky.social · 11/12/2024

Bookmark for all future grants: Use of Promotional Language in Grant Applications and Grant Success: "the percentage of promotional words was positively associated with the probability of receiving funding (NIH grants: odds ratio, 1.51 [95% CI, 1.10-2.11])." jamanetwork.com/journals/jam...

0309

David Baranger @davidbaranger.bsky.social · 05/11/2024

Distracting myself from current events by spending too long on this figure. 😅

A colorful raster plot showing the linear relationship between two variables.

120

David Baranger @davidbaranger.bsky.social · 08/07/2024

Our article exploring the neurobiological correlates of prenatal cannabis exposure is now out! "Prenatal cannabis exposure, the brain, and psychopathology during early adolescence" rdcu.be/dMJqH

142

David Baranger @davidbaranger.bsky.social · 22/05/2024

Stoked for my first APS #APS24SF! I'll be presenting on our InteractionPoweR #rstats package in an invited session on sample size planning!

000

David Baranger @davidbaranger.bsky.social · 13/03/2024

@Alex_PMiller And at the regional level, we see expected decreases (i.e., frontal cortex thickness), but otherwise lots of regions also show *increases* relative to participants with no substance use initiation!

100

David Baranger @davidbaranger.bsky.social · 13/03/2024

@Alex_PMiller A bit more about why I really like this paper: this isn't a drugs = smaller brain story. In fact, we mostly see that substance use is associated with *larger* brain metrics!

100

David Baranger @davidbaranger.bsky.social · 18/01/2024

From the supplement of the paper I'm writing - 2-year longitudinal stability of Neuroimaging IDPs from the ABCD study.

000

David Baranger @davidbaranger.bsky.social · 07/12/2023

A quick data simulation/visualization about being cautious when using difference scores.

A series of charts illustrating that replicable within-subject differences may not yield reliable difference scores, because the variables themselves are not necessarily very reliable to begin with.

1135

David Baranger @davidbaranger.bsky.social · 20/09/2023

Interestingly, a few of these measures are correlated with adolescent psychopathology (externalizing), including two of the ones that survived more stringent correction, and longitudinal mediation analyses were significant for attention/ADHD

100

David Baranger @davidbaranger.bsky.social · 20/09/2023

These were mostly measures from diffusion imaging, but included not only white matter tracts, but also measures of both cortical grey and white matter.

100

David Baranger @davidbaranger.bsky.social · 20/09/2023

Only 3 associations survived correction for multiple tests, and 14 more survived a less stringent correction.

100

David Baranger @davidbaranger.bsky.social · 21/08/2023

It's official!!🎉I got my NOA for my K99/R00 from @NIAAAnews . Really looking forward to starting this project. Also, seems like a good time to share that I'll be on the job market this year! I study alcohol and substance use across development with neuroimaging and genetics.

100

David Baranger @davidbaranger.bsky.social · 05/05/2023

The unreliability of difference scores seems to be a recent theme for me. This figure underscores why.

100

David Baranger @davidbaranger.bsky.social · 19/04/2023

Inspired by a fun conversation with @ThomasIsPsyched, the InteractionPoweR Shiny app for analytic power analyses now includes the option to simulate and visualize single data sets. Available at david-baranger.shinyapps.io/Interac…

000

David Baranger @davidbaranger.bsky.social · 27/09/2022

There's now an even easier way to run a power analysis for an interaction! We've launched a new Shiny app for the {InteractionPoweR} package functions for analytic power. The app is reactive and helps you to visualize your model as you run the analysis:...

100

David Baranger @davidbaranger.bsky.social · 04/08/2022

and variables can be continuous, binary, or likert. You can even simulate what happens to power when variables are artificially discretized.

100

David Baranger @davidbaranger.bsky.social · 04/08/2022

The skew of all three variables...

100

David Baranger @davidbaranger.bsky.social · 04/08/2022

The reliability of all three variables....

100

David Baranger @davidbaranger.bsky.social · 04/08/2022

The package also makes it very easy for users to get more accurate power estimates, as they can specify aspects of the data including the correlation between the interacting variables...

100

David Baranger @davidbaranger.bsky.social · 04/08/2022

Computing power in an interaction is as easy as a simple command...

100

David Baranger @davidbaranger.bsky.social · 23/06/2022

{InteractionPoweR} is now updated with a function for analytic power - leagues faster if all variables are continuous/normal. dbaranger.github.io/InteractionPowe… Vignette now includes my favorite figure, highlighting why correlations are important when computing...

100

David Baranger @davidbaranger.bsky.social · 14/04/2022

We were interested to see that positional-mapping techniques - fastBAT, MAGMA, and H-MAGMA - were the best at gene prioritization, both individually, and were the most useful in the machine learning analyses. See the paper for the secondary/control analyses. 8/11

100

David Baranger @davidbaranger.bsky.social · 14/04/2022

We then took a look at a more common use of multi-omics - gene prioritization - identifying which significant genes are most likely to remain significant in a larger GWAS. Here multi-omics performs fairly well - though far from perfectly. 7/11

100

David Baranger @davidbaranger.bsky.social · 14/04/2022

They don't. No single method, or combination of methods, identifies a majority of true-positive novel genes (PPV < 0.2 = False positive rate > 80%). At best, machine learning in SCZ2 identified 8 novel genes that were subsequently found in SCZ3. 6/11

100

David Baranger @davidbaranger.bsky.social · 26/01/2022

Here's my take on this weeks #TidyTuesday! It was fun to have one of my ideas picked. I took a stab at reducing all the data on game mechanics to just 2 principal components.

000

David Baranger @davidbaranger.bsky.social · 04/06/2021

Really loving Solmi et al. www.nature.com/articles/s41380-021-… This is going to be a regular citation for me! Did a quick plot of Table 2 in the style of every developmental researcher's favorite disorder-onset figure :)

000

David Baranger @davidbaranger.bsky.social · 07/08/2020

What predicts the stability of activation? We looked at activation strength, activation variability, and the size of the ROI. Out of the three, only the variability of activation was replicably associated with stability in both samples. 4/9

100

David Baranger @davidbaranger.bsky.social · 07/08/2020

Lots of prior work has shown that the same reward regions reliably show up in group-level reward contrast maps of reward tasks, so we thought that they would be some of the most-stable regions - instead they were among the least stable. 3/9

100

David Baranger @davidbaranger.bsky.social · 07/08/2020

Win>Baseline had moderate stability (!), and in both samples was more stable than Win>Loss. We think this is probably because task activation is a difference score, and difference scores of correlated variables are unreliable. 2/9

100