Hacker News new | past | comments | ask | show | jobs | submit login

Maybe it depends on the literature. I almost never read the introduction, unless I am unclear on the motivation for the work from the abstract.

The fundamental problem with abstracts and introductions is that they're sales content. You can't trust them. The only thing you can do is decide if you wish to read further, and see if the paper says what they claim it says. For that, an abstract is sufficient about 99% of the time. Spending lots (or...any amount) of time on the intro before you read the data tables is basically asking to have time stolen from you.

For what it's worth, the proportion of papers where the body doesn't support the claims in the abstract is...high. Like way above 50%, in my experience (approaching nearly 100% for any random paper you find on HN or X.)

My algorithm is:

* read abstract. decide to continue.

* formulate my own version of the experiment in my head -- how would I go about answering this question?

* read all data tables, including supplemental materials. see if anything stands out as weird.

* look at all figures, including supplemental materials. see if anything stands out as weird.

* do I understand the data? if not, search for table/figure references from results, until I do (generally I end up reading most of the results, out-of-order).

* is this paper consistent with the way I'd have approached the problem? If not, is that important? what are the possible methodological flaws? what would they look like in the data?

* re-examine data to look for signs of flaws.

* does the data support the conclusion? if not, why not? what would I do differently or as a follow-up?

* finally, decide if the paper worth reading in detail.




In the literature I spend the most time with (archaeology/anthropology), the abstract is often focused on the broad applicability of the research to the field, which I either don't care about or already know (often both). The last line/paragraph of the introduction contains the specific contributions/findings of the paper. If and only if that's reasonable might it be worth investing the time to read further. I only rarely touch the data tables (assuming they even exist), because I don't want to spend an hour trying to understand whatever software nightmare the probably-not-computer-literate author(s) used and miss the details hidden outside the data in the explanatory text (e.g. samples were processed by throwing away all bits we couldn't identify!) that make the time spent useless.


Yeah, we come from very different fields. But I'll say this:

> the probably-not-computer-literate author(s) used and miss the details hidden outside the data in the explanatory text (e.g. samples were processed by throwing away all bits we couldn't identify!) that make the time spent useless.

This is super common in my own field, and it's top-five indication that the paper is bad. If I can't understand how they manipulated the data, unless I have some specific other interest, the entire paper goes in the trash. There's no excuse.

For example, it's pretty routine in biology and medicine to see high profile papers that invent some wild/complex "statistical pseudo-control", instead of actually doing a good experiment (which is hard, and probably why the question is unpublished). These papers are basically never worth reading.

See also: not using conventional statistical tests, not including uncertainty estimates, not using standard/vanilla models before jumping to crazy stuff. Not to be too glib, but "we entered the data into SAS/Excel/R/Matlab" is like a bright, flashing warning sign that you're about to see a clown show.


Unfortunately, most of the foundational papers in the field did exactly that, usually without explicitly stating. It was pretty much the standard of practice for all bone analysis prior to ~2010 or so when microstructure analysis became a practical thing. Similar issues in other types of samples exist and it's still common today, it was just what came to my mind as the first example.


Yeah it depends heavily on the field and sometimes on the subfield.

Also the approach I use depends on how much I know about the field. If I know less, I weight the introduction more. If I know a lot, I skip almost all the sales and stage setting content. E.g. if it's a psych paper, often I jump directly to the results and only read other content if the results look reasonable. And so on.

So, I guess my only real point is that it's hard to take advice for how best to read a paper from someone (like your advisor) who knows the field much better than you do. You should probably gather as many tricks from the advice as you can, but maybe don't treat it as gospel and do lots of experimenting.


It's certainly true that it depends on the subfield and somewhat on your level of experience in that field. But the approach I outlined works pretty well, generally, and only serves to underscore how bad many papers are.

For example, I'm not going to be able to pick up a paper from CERN and do most of the steps I outlined, but if I did do that, and found an obvious statistical anomaly in the first table, well...now I'd have something interesting to ask about and dig into.


oh yeah, I agree. I was agreeing and expanding on your comment, not disagreeing with it.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: