# Interpreting prognostic factor importance from univariate vs multivariable analysis

**URL:** <https://discourse.datamethods.org/t/interpreting-prognostic-factor-importance-from-univariate-vs-multivariable-analysis/2075>\
**Category:** research methods\
**Created:** [August 31, 2019, 8:55pm UTC](https://discourse.datamethods.org/t/interpreting-prognostic-factor-importance-from-univariate-vs-multivariable-analysis/2075 "2019-08-31T20:55:10Z")\
**Posts on this page:** 9\
**Page:** 1

<div class="post-metadata">

**Author:** ![tomasbencomo](https://discourse.datamethods.org/user_avatar/discourse.datamethods.org/tomasbencomo/32/878_2.png) [@tomasbencomo](https://discourse.datamethods.org/u/tomasbencomo)\
**Post date:** [August 31, 2019, 8:55pm UTC](https://discourse.datamethods.org/t/interpreting-prognostic-factor-importance-from-univariate-vs-multivariable-analysis/2075/1 "2019-08-31T20:55:10Z")

</div>

Some prognostic factor studies show results for both univariate and multivariable models. An example of this practice can be found in [Bill-Axelson et al](https://www.nejm.org/doi/full/10.1056/NEJMoa1807801) where they evaluate prognostic factors for prostate cancer. Table 3 shows results from univariate Cox models and an adjusted model with multiple factors

 ![](https://discourse.datamethods.org/uploads/default/original/1X/12a45816865a3c6700ac8d7948968c76639a6f57.jpeg)

What is the usefulness of the univariate Cox results? It’s my understanding that univariate results don’t provide meaningful information due to confounding from other prognostic factors. Without analyzing the prognostic factors together in a full, adjusted model, we can’t really make any conclusions from the univariate analysis.

Assuming my understanding is correct, why do studies even include univariate regression results? Is any information learned from viewing the univariate results?

I’m also aware that to truly prove the importance of a new prognostic factor we’d need to show improvements in model performance through methods like the Likelihood Ratio Test, improved R squared, Harrell’s adequacy index, etc - I’m more curious to learn if I’ve been underestimating the importance of univariate results.

---

<div class="post-metadata">

**Author:** ![f2harrell](https://discourse.datamethods.org/user_avatar/discourse.datamethods.org/f2harrell/32/165_2.png) [@f2harrell](https://discourse.datamethods.org/u/f2harrell)\
**Post date:** [August 31, 2019, 9:39pm UTC](https://discourse.datamethods.org/t/interpreting-prognostic-factor-importance-from-univariate-vs-multivariable-analysis/2075/2 "2019-08-31T21:39:24Z")

</div>

Note that the paper mis-analyzed Gleason score.

I believe that the univariate analyses not only are not helpful. They actually have negative information. You can’t interpret these marginal estimates without knowing the distribution of important covariates that they are **not** adjusted for. The univariate and multivariable estimates _should_ differ, and the univariate estimates do not help. You’ll also find that the proportional hazards assumption is violated more with unadjusted estimates than with adjusted ones.

---

<div class="post-metadata">

**Author:** ![tomasbencomo](https://discourse.datamethods.org/user_avatar/discourse.datamethods.org/tomasbencomo/32/878_2.png) [@tomasbencomo](https://discourse.datamethods.org/u/tomasbencomo)\
**Post date:** [September 1, 2019, 6:33pm UTC](https://discourse.datamethods.org/t/interpreting-prognostic-factor-importance-from-univariate-vs-multivariable-analysis/2075/3 "2019-09-01T18:33:05Z")

</div>

Thanks for the answer! Are there any good papers or texts you’d recommend to help convince colleagues that univariate prognostic factor results are uninformative?

---

<div class="post-metadata">

**Author:** ![f2harrell](https://discourse.datamethods.org/user_avatar/discourse.datamethods.org/f2harrell/32/165_2.png) [@f2harrell](https://discourse.datamethods.org/u/f2harrell)\
**Post date:** [September 2, 2019, 11:18am UTC](https://discourse.datamethods.org/t/interpreting-prognostic-factor-importance-from-univariate-vs-multivariable-analysis/2075/4 "2019-09-02T11:18:46Z")

</div>

I’m sure that such a paper exists, and hope that others can point us to it. One general comment: In randomized studies, even with perfect covariate balance, unadjusted odds and hazard ratios are hurt by the non-collapsibility of effect ratios, making them not estimate the same quantity as conditional (adjusted) ratios. When the outcome is continuous and the study is randomized, unadjusted differences in means at least estimate the same quantity as adjusted differences, so there is slightly more of a reason to show raw means in that case (but still not advised). Adjustment in the normal continuous Y case gains significant precision (narrows confidence intervals) but does not systematically shift the difference in means as happens with odds and hazard ratios.

---

<div class="post-metadata">

**Author:** ![Pavlos\_Msaouel](https://discourse.datamethods.org/user_avatar/discourse.datamethods.org/pavlos_msaouel/32/739_2.png) [@Pavlos\_Msaouel](https://discourse.datamethods.org/u/Pavlos_Msaouel)\
**Post date:** [September 2, 2019, 3:32pm UTC](https://discourse.datamethods.org/t/interpreting-prognostic-factor-importance-from-univariate-vs-multivariable-analysis/2075/5 "2019-09-02T15:32:58Z")

</div>

[This is a good article](https://www.sciencedirect.com/science/article/pii/089543569600025X) for univariate vs multivariable analyses.

[Here](https://stats.stackexchange.com/questions/129123/how-can-manova-report-a-significant-difference-when-none-of-the-univariate-anova) is also a simple example of how MANOVA is more informative than univariate ANOVAs.

And [this is another good one](https://journalofbigdata.springeropen.com/articles/10.1186/s40537-018-0143-6) for stepwise procedures in general.

---

<div class="post-metadata">

**Author:** ![f2harrell](https://discourse.datamethods.org/user_avatar/discourse.datamethods.org/f2harrell/32/165_2.png) [@f2harrell](https://discourse.datamethods.org/u/f2harrell)\
**Post date:** [September 2, 2019, 4:37pm UTC](https://discourse.datamethods.org/t/interpreting-prognostic-factor-importance-from-univariate-vs-multivariable-analysis/2075/6 "2019-09-02T16:37:06Z")

</div>

Those are really good articles. It would be good to also find an article discussing this when variable selection is not being done.

---

<div class="post-metadata">

**Author:** ![tomasbencomo](https://discourse.datamethods.org/user_avatar/discourse.datamethods.org/tomasbencomo/32/878_2.png) [@tomasbencomo](https://discourse.datamethods.org/u/tomasbencomo)\
**Post date:** [September 2, 2019, 7:22pm UTC](https://discourse.datamethods.org/t/interpreting-prognostic-factor-importance-from-univariate-vs-multivariable-analysis/2075/7 "2019-09-02T19:22:27Z")

</div>

The MANOVA example was especially helpful. [This discussion](https://stats.stackexchange.com/questions/254254/why-do-we-need-multivariate-regression-as-opposed-to-a-bunch-of-univariate-regr) cites your MANOVA example and has more examples showing why we need multivariable models to account for covariate dependence.  
It’s proving more difficult to find papers about univariate analyses that aren’t in a variable selection context.

---

<div class="post-metadata">

**Author:** ![lbautista](https://discourse.datamethods.org/letter_avatar_proxy/v4/letter/l/ea5d25/32.png) [@lbautista](https://discourse.datamethods.org/u/lbautista)\
**Post date:** [September 4, 2019, 2:32pm UTC](https://discourse.datamethods.org/t/interpreting-prognostic-factor-importance-from-univariate-vs-multivariable-analysis/2075/8 "2019-09-04T14:32:38Z")

</div>

> [@tomasbencomo](#):
>
> Assuming my understanding is correct, why do studies even include univariate regression results? Is any information learned from viewing the univariate results?

The answer to this question depends on what is the goal of the study. If the study is aimed to estimate causal effects, then you are right: crude (univariate) estimates are very likely biased (confounded) and, therefore, are not useful for causal inferences. On the other hand, if you have two patients with the same age, selected at random from the population in this study, one of a Gleason score of 4+3 and the other one with a Gleason score of 3-6, you would expect the risk in the one with the highest score to be 12 times higher than the one with the lowest score. This may lead you to further evaluate or treat the patient with the higher score, particularly if you don’t have information on other risk factors. Thus, the crude estimate in not completely uninformative.

---

<div class="post-metadata">

**Author:** ![scboone](https://discourse.datamethods.org/letter_avatar_proxy/v4/letter/s/edb3f5/32.png) [@scboone](https://discourse.datamethods.org/u/scboone)\
**Post date:** [September 4, 2019, 6:12pm UTC](https://discourse.datamethods.org/t/interpreting-prognostic-factor-importance-from-univariate-vs-multivariable-analysis/2075/9 "2019-09-04T18:12:23Z")

</div>

Although not really the same problem as we are dealing with here, maybe these papers are of interest to you as well. They discuss another issue of interpreting multiple/all coefficients in one multivariable model (which is mostly an issue when we deal with etiological/causal research):

> **[The Table 2 Fallacy: Presenting and Interpreting Confounder and Modifier...](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3626058/)**
>
> It is common to present multiple adjusted effect estimates from a single model in a single table. For example, a table might show odds ratios for one or more exposures and also for several confounders from a single logistic regression. This can lead...

> <https://www.ncbi.nlm.nih.gov/pubmed/29782045>
>
> A "Table Fallacy," as coined by Westreich and Greenland, reports multiple adjusted effect estimates from a single model. This practice, which remains common in published literature, can be problematic when different types of effect estimates are presented together in a single table. The purpose of this paper is to quantitatively illustrate this potential for misinterpretation with an example estimating the effects of preeclampsia on preterm birth.We analysed a retrospective population-based cohort of 2 963 888 singleton births in California between 2007 and 2012. We performed a modified Poisson regression to calculate the total effect of preeclampsia on the risk of PTB, adjusting for previous preterm birth. pregnancy alcohol abuse, maternal education, and maternal socio-demographic factors (Model 1). In subsequent models, we report the total effects of previous preterm birth, alcohol abuse, and education on the risk of PTB, comparing and contrasting the controlled direct effects, total effects, and confounded effect estimates, resulting from Model 1.The effect estimate for previous preterm birth (a controlled direct effect in Model 1) increased 10% when estimated as a total effect. The risk ratio for alcohol abuse, biased due to an uncontrolled confounder in Model 1, was reduced by 23% when adjusted for drug abuse. The risk ratio for maternal education, solely a predictor of the outcome, was essentially unchanged.Reporting multiple effect estimates from a single model may lead to misinterpretation and lack of reproducibility. This example highlights the need for careful consideration of the types of effects estimated in statistical models.
