Timeline for Choice of coding scheme/planned contrasts using race as a categorical variable
Current License: CC BY-SA 4.0
10 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Apr 23, 2023 at 16:08 | history | bounty ended | Bryan Krause | ||
| Apr 21, 2023 at 16:38 | comment | added | Noah | The contrasts I recommend are for inferential models and are used to test hypotheses. As I explain, the contrasts I recommend are exactly equivalent to regression coefficients in some cases (which are obviously used for inference). And polynomial models aren't the only models that yield uninterpretable coefficients; models with many interactions to better reflect the complex processes also would benefit from this method of presenting results. This method is simply, broadly applicable, and solves the problem you identify (and explained by the authors in the paper you cite). | |
| Apr 21, 2023 at 16:32 | comment | added | Bryan Krause | Yeah, as they say, "conditional predictions, is useful for a model’s key explanatory variable when one wants to clarify what results imply in terms of the expected value(s) of the outcome" - but, like my last comment, that's often not what readers are expecting or wanting to know. | |
| Apr 21, 2023 at 16:29 | comment | added | Noah | Reading the article more carefully, they recommend exactly what I do in the section "Alternatives to using a reference category". So I stand by recommendation. Simply omit categories that you don't have data to make valid comparisons on. Just because medical research currently uses poorly justified simple models, doesn't mean we shouldn't develop a method that both avoids the problem you identify in the question and allows straightforward interpretation of complicated models, which my recommended does. | |
| Apr 21, 2023 at 16:25 | comment | added | Bryan Krause | The solution to present all pairwise comparisons seems reasonable in a non-data-limited case, but often less common categories are undersampled without some more sophisticated attempt to target them. If you take as a sample the group of people walking through a US hospital's doors and plan to do pairwise contrasts between each race category, some of those contrasts are going to be pretty useless, based on tiny samples that can't possibly give good estimates of any actual differences. | |
| Apr 21, 2023 at 16:21 | comment | added | Bryan Krause | It seems standard in medical/health publishing to both use relatively simple, interpretable models when the goals are inference rather than prediction, and typically to present effects rather than predictions. The question is generally "does X make Y better or worse and how much", rather than "what's the expected result for A". For categorical predictors there's no need to think about polynomial relationships. | |
| Apr 21, 2023 at 16:17 | comment | added | Noah | It seemed to me that the problem was due to presenting regression coefficients, which, no matter how the model is parameterized, prioritize one group of others, and so my recommendation was to avoid showing regression coefficients as they often have no useful interpretation except in the simplest of models. My "tangent" was to explain how to present the results of a model that avoid regression coefficients. That article also focuses on presenting regression coefficients as though that is the only way to present the results of a regression model. | |
| Apr 21, 2023 at 16:15 | comment | added | Bryan Krause | journals.sagepub.com/doi/full/10.1177/0081175020982632 argues that a reference category is unnecessary and advocates for mean contrasts or binary contrasts. | |
| Apr 21, 2023 at 16:09 | comment | added | Bryan Krause | Sure, it's good to point out that ultimately these are all the same model, but I think it's missing the point a bit and ending up on a tangent about linear approximations. Probably largely due to how I presented the problem. In that case, though, the problem remains about how to present the results. Ultimately, that means presenting contrasts between groups, and the argument is that using the majority category as a reference exhibits bias. | |
| Apr 17, 2023 at 17:06 | history | answered | Noah | CC BY-SA 4.0 |