**Bad Hessian » R**, and kindly contributed to R-bloggers]. (You can report issue about the content on this page here)

Want to share your content on R-bloggers? click here if you have a blog, or here if you don't.

Last week, Alaska took it home with her *dangerous* performance, while Ivy Winters was sent home after going up against Alyssa Edwards. This is sad on many fronts. First, I love me some Ivy Winters. Second, Jinkx had revealed that she had a crush on Ivy, and the relationship that may have flourished between the two would have been too cute. But lastly, and possibly the worst, *both of my models from last week had Ivy on top*. Ugh.

What went wrong? Well, this certainly wasn’t Ivy’s challenge. But it’s high time that I started interrogating the models a little further.

## Testing the proportional hazards assumption

One of the critical assumptions of the Cox proportional hazard model is the proportional hazard (PH) assumption. That is, the hazard ratio across time is a fixed value.

where is the baseline hazard at time t, is the hazard for individual *i*, and *x* is the set of covariates.

Testing this assumption is somewhat straightforward in R: the `cox.zph`

function produces a quick way to test how this holds across covariates. If we can reject the null hypothesis, then the PH assumption cannot hold for that covariate. For the first model, using the raw lipsync count, yields this:

> t.zph <- cox.zph(t.cox2) > t.zph rho chisq p Age -0.2278 2.441 1.18e-01 PlusSize -0.2363 5.422 1.99e-02 PuertoRico -0.0464 0.216 6.42e-01 Wins -0.4532 30.637 3.11e-08 Highs -0.4192 25.826 3.74e-07 Lows -0.4500 30.175 3.95e-08 Lipsyncs -0.5025 32.763 1.04e-08 GLOBAL NA 38.287 2.67e-06

Looks like all the time-variant covariates fail this test in this model. Box-Steffensmeier and Jones recommend adding interactions with those variables that vary by time. In this case, I’ve constructed a new variable to approximate the number of competitors left in the competition, `CompLeft`

. I say “approximate” because it would decline monotonically if it wasn’t for edge cases, like this season’s double elimination and cases in which Ru lets both queens stay. I’ll touch on this later below.

For the moment, this produces the following summary:

> summary(t.cox2_ph) Call: coxph(formula = t.surv ~ (Age + PlusSize + PuertoRico + Wins + Highs + Lows + Lipsyncs + CompLeft + Wins * CompLeft + Highs * CompLeft + Lows * CompLeft + Lipsyncs * CompLeft) + cluster(ID), data = df) n= 314, number of events= 41 coef exp(coef) se(coef) robust se z Pr(>|z|) Age 8.427e-05 1.000e+00 4.501e-02 4.092e-02 0.002 0.998 PlusSize 2.120e-02 1.021e+00 5.188e-01 5.888e-01 0.036 0.971 PuertoRico -3.089e-01 7.342e-01 5.812e-01 6.138e-01 -0.503 0.615 Wins -7.285e-01 4.827e-01 6.502e-01 6.769e-01 -1.076 0.282 Highs 3.483e-01 1.417e+00 4.006e-01 3.874e-01 0.899 0.369 Lows -3.523e-01 7.031e-01 7.448e-01 5.337e-01 -0.660 0.509 Lipsyncs 5.860e-01 1.797e+00 5.022e-01 4.624e-01 1.267 0.205 CompLeft -7.443e-02 9.283e-01 1.502e-01 1.128e-01 -0.660 0.509 Wins:CompLeft 1.203e-01 1.128e+00 1.313e-01 1.346e-01 0.894 0.371 Highs:CompLeft -1.173e-02 9.883e-01 6.628e-02 5.747e-02 -0.204 0.838 Lows:CompLeft 1.200e-01 1.127e+00 1.115e-01 9.687e-02 1.239 0.215 Lipsyncs:CompLeft 1.856e-01 1.204e+00 8.462e-02 6.754e-02 2.748 0.006 ** --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 exp(coef) exp(-coef) lower .95 upper .95 Age 1.0001 0.9999 0.9230 1.084 PlusSize 1.0214 0.9790 0.3221 3.239 PuertoRico 0.7342 1.3620 0.2205 2.445 Wins 0.4827 2.0719 0.1281 1.819 Highs 1.4167 0.7059 0.6630 3.027 Lows 0.7031 1.4224 0.2470 2.001 Lipsyncs 1.7967 0.5566 0.7260 4.447 CompLeft 0.9283 1.0773 0.7441 1.158 Wins:CompLeft 1.1279 0.8866 0.8663 1.468 Highs:CompLeft 0.9883 1.0118 0.8831 1.106 Lows:CompLeft 1.1275 0.8869 0.9325 1.363 Lipsyncs:CompLeft 1.2039 0.8306 1.0546 1.374 Concordance= 0.899 (se = 0.058 ) Rsquare= 0.194 (max possible= 0.542 ) Likelihood ratio test= 67.77 on 12 df, p=8.35e-10 Wald test = 179.6 on 12 df, p=0 Score (logrank) test = 93.04 on 12 df, p=1.266e-14, Robust = 35.33 p=0.0004154 (Note: the likelihood ratio and score tests assume independence of observations within a cluster, the Wald and robust score tests do not).

And passes the proportional hazard’s test:

> t.zph <- cox.zph(t.cox2_ph) > t.zph rho chisq p Age -0.1583 1.5353 0.21532 PlusSize -0.1149 1.4122 0.23469 PuertoRico 0.0285 0.0708 0.79012 Wins -0.2363 4.3315 0.03741 Highs -0.3154 9.1745 0.00245 Lows -0.3241 5.6739 0.01722 Lipsyncs -0.1252 1.2698 0.25981 CompLeft 0.0259 0.0338 0.85409 Wins:CompLeft 0.0546 0.1811 0.67039 Highs:CompLeft 0.1623 1.2236 0.26866 Lows:CompLeft -0.1125 0.8778 0.34880 Lipsyncs:CompLeft -0.1160 0.8600 0.35373 GLOBAL NA 17.4112 0.13477

Testing the second model, the one with `LipsyncWithoutOut`

, we get the following:

> t.zph <- cox.zph(t.cox3) > t.zph rho chisq p Age -0.0746 0.311 0.577 PlusSize -0.0627 0.349 0.554 PuertoRico 0.1157 0.441 0.507 Wins -0.0586 0.132 0.716 Highs -0.1100 0.776 0.378 Lows -0.1021 0.709 0.400 LipsyncWithoutOut 0.1170 0.363 0.547 GLOBAL NA 1.894 0.965

Looks good on that front.

## Time-dependency

One thing that I’ve been puzzling over is that, given that we know the baseline hazard, we could directly put it in the model. This is the intuition behind the `CompLeft`

variable.

Since I’ve already incorporated this into the first model, let’s see how it fairs with the second model.

> summary(t.cox3s) Call: coxph(formula = t.surv ~ (Age + PlusSize + PuertoRico + Wins + Highs + Lows + LipsyncWithoutOut + CompLeft) + cluster(ID), data = df) n= 314, number of events= 41 coef exp(coef) se(coef) robust se z Pr(>|z|) Age -0.06092 0.94090 0.03999 0.03855 -1.580 0.1140 PlusSize -0.12822 0.87966 0.52217 0.57030 -0.225 0.8221 PuertoRico -0.18703 0.82942 0.49655 0.43475 -0.430 0.6671 Wins -1.28217 0.27743 0.30276 0.21728 -5.901 3.62e-09 *** Highs -0.78254 0.45724 0.22516 0.17009 -4.601 4.21e-06 *** Lows -1.51944 0.21884 0.39846 0.35582 -4.270 1.95e-05 *** LipsyncWithoutOut -0.56267 0.56969 0.32437 0.22211 -2.533 0.0113 * CompLeft 0.15362 1.16604 0.09737 0.08953 1.716 0.0862 . --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 exp(coef) exp(-coef) lower .95 upper .95 Age 0.9409 1.0628 0.8724 1.0147 PlusSize 0.8797 1.1368 0.2877 2.6900 PuertoRico 0.8294 1.2057 0.3538 1.9446 Wins 0.2774 3.6045 0.1812 0.4247 Highs 0.4572 2.1870 0.3276 0.6382 Lows 0.2188 4.5696 0.1090 0.4395 LipsyncWithoutOut 0.5697 1.7554 0.3686 0.8804 CompLeft 1.1660 0.8576 0.9784 1.3897 Concordance= 0.762 (se = 0.058 ) Rsquare= 0.11 (max possible= 0.542 ) Likelihood ratio test= 36.56 on 8 df, p=1.389e-05 Wald test = 67.04 on 8 df, p=1.903e-11 Score (logrank) test = 37.35 on 8 df, p=9.941e-06, Robust = 28.44 p=0.0003982 (Note: the likelihood ratio and score tests assume independence of observations within a cluster, the Wald and robust score tests do not).

So attempting to estimate the baseline hazard directly, there’s a slightly statistically significant effect. I read this as a small increase in the risk of elimination for each additional competitor that is left.

## Assessing models

I promised last week that I was going to do some cross-validation. Particularly I was thinking of doing something like a Leave-One-Out Cross Validation (LOOCV) where the “one” would be a season for each of the four completed seasons. When it came time to do this, though, I mostly puzzled in how I would assess the adequacy of my model. Given that this is a survival model, would I say the model is good if it is able to predict the next queen who is eliminated before each episode? Would I say it is good if it predicted the correct order of placement? I’m at somewhat of an impasse, here. It seems like I need a test that uses the most information to assess goodness-of-fit. I’d appreciate comments on this.

Until then, I decided to look at deviance residuals across queens from the past four seasons to judge model fit. `c2`

is the original model, `c2ph`

is the model with interacted terms to account for the PH assumption, `c3`

is the second model, and `c3s`

is the second model with the `CompLeft`

variable included. The vertical dotted lines denote separation between seasons.

It looks like, in general, the model is above zero, which means it underestimates the probability of elimination. Curiously, the `c3`

models seem to *really* underestimate failure in season 4. Let’s look at the mean squared residuals.

c2 - 0.9984813 c2ph - 0.9882616 c3 - 1.186669 c3s - 1.128196

There doesn’t seem to be much movement here — small decreases in the `c2ph`

and `c3s`

models. Not terribly surprising.

## So who’s going to sashay away?

Enough of all that. Let’s run the numbers for this week, using the latter two models. Using `c2ph`

, Jinkx-y comes out on top, while Alyssa is in imminent danger of elimination.

1 Jinkx Monsoon 1.1237273 1.1355589 2 Alaska 1.3448615 1.3524366 3 Detox 2.0229849 0.5172689 4 Roxxxy Andrews 7.3588579 3.8368226 5 Coco Montrese 8.3889735 1.5895090 6 Alyssa Edwards 12.1112424 2.3408098

All I have to say about this is…

## Thank you Jesus.

In model `c3s`

, we see a somewhat different story.

1 Roxxxy Andrews 0.01433954 0.09192587 2 Alaska 0.02843240 0.11964074 3 Jinkx Monsoon 0.14901925 0.23457349 4 Alyssa Edwards 0.32526816 0.24033382 5 Coco Montrese 1.09607385 0.50933051 6 Detox 1.61778023 0.36059397

Roxxxy takes lead honors, but Detox is on bottom. However, the relative risks here are rather close to each other, compared to the previous model.

One of these days, I’m going to get a prediction right. Is it this week? We’ll find out soon enough… Until next time, squirrel friends.

P.S. All new code for this week can be found in this gist. It should be used in conjunction with the code from the original one, which has been slightly edited to fix some data errors.

**leave a comment**for the author, please follow the link and comment on their blog:

**Bad Hessian » R**.

R-bloggers.com offers

**daily e-mail updates**about R news and tutorials about learning R and many other topics. Click here if you're looking to post or find an R/data-science job.

Want to share your content on R-bloggers? click here if you have a blog, or here if you don't.