Noah Greifer

x[c(2,5,9),] identifies the submatrix

00
AD
00
00
BE
00
00
00
CF
00

R fills in matrices by column (alphabetically above), recycling the vector 1:2 as needed (in this case 3 times to fill 6 spots.

So we get A = 1, B = 2, C = 1, D = 2, E = 1, F = 2.

I TA'd for John at ICPSR in 2017. He was very supportive of me early on in my career and so inspiring as someone bridging statistics education, methodological research, and statistical computing. He is certainly a giant whose shoulders I am honored to stand on.

sprintf() is pretty close

Do you mean "immortal time bias"? Definitely one of the coolest named biases

"The probability that there is no effect is the same as the probability of flipping 7 heads in a row"

See, I can misinterpret S-values, too

Any #rstats advice for writing mathematical expressions in ggplot2 axis labels? Dying not to have to use the insanity that is bquote()

I could have bootstrapped the whole thing and it still would have been quicker!

What do you mean my Bayesian logistic regression on a moderate sample took 3 minutes to run

Hopefully this causes a paradigm shift and stops these papers from getting through.

Next we need clones of this paper for every social science discipline.

This is excellent, and I'm so glad this paper was finally written, and so clearly as well. I basically write an equivalent every time I am consulting with someone proposing this type of study, and I'm so glad I can save my effort and just send them this instead!

A "methods primer" article in the journal "BMJ Medicine", titled "Factors associated with: problems of using exploratory multivariable regression to identify causal risk factors"

We wrote an article explaining why you shouldn't put several variables into a regression model and report which are statistically significant - even as exploratory research. bmjmedicine.bmj.com/content/4/1/.... How did we do?

This sounds suspiciously like you want to interpret the coefficients from this model… Fit the best model for the data (probably logistic) and use marginaleffects to compute the quantity of interest.

Only true for the ATT, but also equivalent to the just identified CBPS!

Looks like observations are being dropped due to missingness. Did you encode the censoring event as NA if it didn’t occur? I think it’s supposed to be encoded as logical. Would be helpful to see the function call, not just the output.

🙃

Confidence Regions for Multiple Outcomes, Effect Modifiers, and Other Multiple Comparisons In epidemiology, some have argued that multiple comparison corrections are not necessary as there is rarely interest in the universal null hypothesis. From a parameter estimation perspective, epidemio...

For years I had trouble following some of the discussion about confidence bands, but at ACIC this year @noahgreifer.bsky.social pointed me to a helpful paper

So you don't have to be as perplexed as I once was, we have a new pre-print introducing the key ideas
arxiv.org/abs/2510.07076

My hot take is that "fixed effects" has a single, clear meaning that is equivalent across all subdisciplines of statistics.

bsky.app/profile/noah...

These methods are new and bespoke, but I'd love to hear about if they inspire or help you in your own research! Feel free to let me know if you have any questions about the methodology. How would you have solved these problems?

All this can be done using my {MatchingFrontier} #Rstats package, which isn't yet on CRAN. This is part of my growing body of "cool new methods I've programmed but am too lazy to write a paper about". Please get in touch if you want to collab on some.

We used g-computation with a cluster-robust SE for pair membership. Because we dropped units from both groups, this analysis targeted the ATO, which is the best we could do given the lack of overlap.

I won't speak about the results (IMO they are less cool than the methods 😉)

After selecting our subset, we did a 2:1 pair match on the scaled Euclidean distance to assign each control unit a treated unit, which supplied its initiation date. Finally, we regressed the 90-day outcomes on the treatment and selected covariates in the matched sample and estimated the effect.

The lowest energy distance was achieved with too small a sample to make inferences, given our hypothesized effect sizes. Instead, we chose the largest sample along the 2:1 energy distance frontier that had all SMDs below .1, and adjusted for remaining imbalance with regression.

At each step, we dropped the unit that, when dropped, yielded the largest decrease in the energy distance, subject to the constraint that the ratio of control units to treated units was 2:1 so we could perform 2:1 matching in the final selected subset. This process is greedy, not optimal, but worked

We dropped units to minimize the energy distance between the groups. The energy distance was described for balancing by Huling & Mak (2014); it is a scalar, multivariate measure of the difference between two *joint* distributions. In this case, those are the confounder distributions in the groups.

We dropped units from the sample one at a time so that the units that remained formed a balanced sample. How did we decide which unit to drop at each step and when to stop dropping units? The framework for this is described by King et al (2017) as the "matching frontier".

In traditional matching, pairing is used to identify a matched subset of the original sample that is balanced. Though we want pairs and and we want balance, it turns out we can do the subset selection first and the pairing second and get better results. Cho et al (2013) describe subset selection.

References for the above, since these papers are not well known, but should be:

King et al (2017): doi.org/10.1111/ajps...
Huling & Mak (2024): doi.org/10.1515/jci-...
Cho et al (2013): doi.org/10.1111/stan...

To solve this, we combined several existing methods to develop our own. This was a combination of the matching frontier developed by King et al (2017), energy balancing developed by Huling & Mak (2024), and subset selection by Cho et al (2013).

Traditional and even advanced matching methods failed. PS matching, Mahalanobis distance matching, and cardinality matching all failed to achieve adequate balance or maintain sufficient sample size, even after trying many variations (calipers, etc.). We needed a more specialized approach.