The Cox proportional hazards model (implemented in R as
coxph() in the
survival package or as
rms package) is one of the most frequently used estimators in duration (survival) analysis. Because it is estimated using only the observed durations' rank ordering, typical quantities of interest used to communicate results of the Cox model come from the hazard function (e.g., hazard ratios or percentage changes in the hazard rate). These quantities are substantively vague and difficult for many audiences of research to understand. The
coxed package introduces a suite of methods to address these problems. The package allows researchers to calculate duration-based quantities from Cox model results, such as the expected duration (or survival time) given covariate values and marginal changes in duration for a specified change in a covariate. These duration-based quantities often match better with researchers' substantive interests and are easily understood by most readers.
This document is a walkthrough of the examples included in the help documentation for the
Before we begin, we load the
and packages from the
tidyverse for managing and plotting data as we go:
library(dplyr) library(tidyr) library(ggplot2)
The following quote from Kropko and Harden (2018) sets up our first working example:
Martin and Vanberg (2003) examine the determinants of negotiation time among political parties forming a coalition government. . . . The dependent variable in Martin and Vanberg’s analysis is the number of days between the beginning and end of the bargaining period. Martin and Vanberg model this variable as a function of the Range of government, which is a measure of the ideological distance between the extreme members of the coalition, the Number of government parties in the coalition, and several other variables. They interact Number of government parties with the natural log of time because that variable violates the proportional hazards assumption. Their hypotheses predict negative coefficients on the variables of interest, indicating that increases in the ideological distance between the parties and in the number of parties correspond with a decrease in the risk of government formation, or a longer negotiation time.
The authors demonstrate support for their hypotheses by computing changes in the hazard rate based on changes to these independent variables. However, to assess what the estimated effects of Range of government and Number of government parties mean in substantive terms, we use
coxed() to predict how long is each case predicted to last. We will also find answers to the following questions about duration:
How much longer will negotiations take for an ideologically polarized coaltion as compared to an ideologically homogeneous one?
How much longer will negotiations take for a multiparty coalition government than for a single-party government?
First we replicate the Cox model from Martin and Vanberg (2003):
mv.surv <- Surv(martinvanberg$formdur, event = rep(1, nrow(martinvanberg))) mv.cox <- coxph(mv.surv ~ postel + prevdef + cont + ident + rgovm + pgovno + tpgovno + minority, method = "breslow", data = martinvanberg) summary(mv.cox)
Next we will use the both versions of
coxed() to examine expected durations and marginal changes in duration.
The first version of
coxed() is the non-parametric step function (NPSF) approach. To use this version, specify
model="npsf" in the call to
coxed(). By default, quantities are estimated without standard errors, but to estimate SEs and confidence intervals specify
To see predicted durations from the Cox model, place the Cox model output as the first argument of
ed1 <- coxed(mv.cox, method="npsf")
There are a number of uses of the
coxed() output. First, the predicted durations for each individual observation are stored in the
summary() function, when applied to
coxed, reports either the mean or median duration in the estimation sample, depending on the option specified with
summary(ed1, stat="mean") summary(ed1, stat="median")
The predicted mean duration of government negotiations is 25.18 days, and the predicted median duration is 19.12 days.
In addition to reporting the mean and median duration, the NPSF version of
coxed() provides estimates of the cumulative baseline hazard function and the baseline survivor function in the data. These functions are stored as a data frame in the
We can plot these baseline functions with
baseline <- gather(ed1$baseline.functions, cbh, survivor, key="survivefunction", value="value") ggplot(baseline, aes(x=time, y=value)) + geom_line() + xlab("Time") + ylab("Function") + facet_wrap( ~ survivefunction, scales = "free")
We can calculate standard errors and confidence intervals for any of these quantities with the
bootstrap=TRUE option. By default the bootstrapping procedure uses 200 iterations (to set this value to a different number, use the
B argument). Here we use 30 iterations simply to ease the computational burden of compiling this vignette. For more reliable results, set
B to a higher value:
ed1 <- coxed(mv.cox, method="npsf", bootstrap = TRUE, B=30)
Now every predicted duration has a standard error and a 95% confidence interval.
The mean and median also have standard errors and confidence intervals.
summary(ed1, stat="mean") summary(ed1, stat="median")
To change the confidence interval to a different level, use the
ed1 <- coxed(mv.cox, method="npsf", bootstrap = TRUE, B=30, level=.8) summary(ed1, stat="mean") summary(ed1, stat="median")
There are different methods for calculating a bootstrapped confidence interval. The default method used by
coxed() (setting the argument
confidence="studentized") adds and subtracts
qnorm(level - (1 - level)/2) times the bootstrapped standard error to the point estimate. The alternative approach is to take the
level + (1-level)/2 quantiles of the bootstrapped draws, which can be done by specifying
confidence="empirical" (we recommend a higher number of bootstrap iterations for empirical confidence intervals):
ed1 <- coxed(mv.cox, method="npsf", bootstrap = TRUE, B=30, confidence="empirical") summary(ed1, stat="mean") summary(ed1, stat="median")
coxed() can be used to provide duration predictions for observations outside of the estimation sample. Suppose that we observe three new cases and place them inside a new data frame:
new.coalitions <- data.frame(postel = c(1,0,0), prevdef = c(1,0,1), cont = c(1,0,0), ident = c(1,3,2), rgovm = c(0.81, 0.62, 1.18), pgovno = c(2,3,4), tpgovno = c(3.58, 5.17, 10.2), minority = c(0,0,1)) new.coalitions
To forecast durations for these cases along with standard errors and confidence intervals, we use the
coxed() function and place
new.coalitions into the
forecast <- coxed(mv.cox, newdata=new.coalitions, method="npsf", bootstrap=TRUE, B=30) forecast$exp.dur
Here we use
coxed() to provide answers to the two duration-based questions we posed in the introduction. First consider "How much longer will negotiations take for an ideologically polarized coalition as compared to an ideologically homogeneous one?" To answer this question, we call
coxed() and specify two new datasets, one in which
rgovm=0 indicating that all political parties in the governing coalition have the same ideological position, and one in which
rgovm=1.24, indicating that the parties have very different ideological positions. We use
mutate() from the
dplyr library to quickly create new data frames in which
rgovm equals 0 or 1.24 for all cases, and set these two data frames as
me <- coxed(mv.cox, method = "gam", bootstrap = TRUE, B = 30, newdata = dplyr::mutate(martinvanberg, rgovm = 0), newdata2 = dplyr::mutate(martinvanberg, rgovm = 1.24))
coxed() calculates expected durations for all cases under each new data frame and subtracts the durations for each case. As an overall result, we can see either the mean or the median of these differences.
summary(me, stat="mean") summary(me, stat="median")
A coalition in which the parties have ideological differences so that
rgovm=1.24 will take 3.08 more days on average (with a median of 2.6 days) to conclude negotiations than a coalition in which all parties have the same position.
Next we consider "How much longer will negotiations take for a multiparty coalition government than for a single-party government?" In this case we compare coalitions with one party to coalitions with 6 parties by setting the
pgovno variable to 1 and 6 and setting these two data frames as the
newdata2 arguments of
me <- coxed(mv.cox, method="npsf", bootstrap = TRUE, B=30, newdata = dplyr::mutate(martinvanberg, pgovno=1), newdata2 = dplyr::mutate(martinvanberg, pgovno=6)) summary(me, stat="mean") summary(me, stat="median")
A coalition of 6 parties will take 50.5 more days on average (with a median of 28.5 days) to conclude negotiations than a coalition with one party.
We can use the GAM method to for all of the same uses for which we used the NPSF method above, except for estimating the baseline functions. We can however view and plot the output from the GAM model that maps predicted ranks to duration. While the
bootstrap=TRUE argument works when
method="gam", these functions take somewhat longer to run. We therefore run the following examples without bootstrapping.
As before, to see predicted durations from the Cox model, place the Cox model output as the first argument of
coxed(). The predicted durations for each individual observation are stored in the
ed2 <- coxed(mv.cox, method="gam") head(ed2$exp.dur)
summary() reports either the mean or median expected duration:
summary(ed2, stat="mean") summary(ed2, stat="median")
The GAM method can also forecast durations for new data along with standard errors and confidence intervals. Here we use the
coxed() function with
method="gam" and place the
new.coalitions we created above into the
forecast <- coxed(mv.cox, newdata=new.coalitions, method="gam") forecast$exp.dur
Here we again calculate the two marginal effects to better understand the substantive meaning of the Cox model. This time we employ the GAM method instead of the NPSF method. The GAM method may provide a warning that some observations have linear predictors that are greater than or less than all of the observed cases in the estimation sample. Some observations falling outside the range of the original linear predictors is to be expected when applying new data, but if it happens with too many of the new observations NPSF may be a better option for estimating these quantities.
me <- coxed(mv.cox, method="gam", newdata = dplyr::mutate(martinvanberg, rgovm=0), newdata2 = dplyr::mutate(martinvanberg, rgovm=1.24)) summary(me, stat="mean") summary(me, stat="median")
me <- coxed(mv.cox, method="gam", newdata = dplyr::mutate(martinvanberg, pgovno=1), newdata2 = dplyr::mutate(martinvanberg, pgovno=6)) summary(me, stat="mean") summary(me, stat="median")
The data used by
coxed() to map rankings to durations are stored in the
gam.data attribute, and the output from the GAM is stored in
gam.data can be used to visualize the fit of the GAM:
ggplot(ed2$gam.data, aes(x=rank.xb, y=y)) + geom_point() + geom_line(aes(x=rank.xb, y=gam_fit)) + geom_ribbon(aes(ymin=gam_fit_95lb, ymax=gam_fit_95ub), alpha=.5) + xlab("Cox model LP rank (smallest to largest)") + ylab("Duration")
coxed() contains two alternative methods for generating expected durations, it is possible to compare the estimates. Both correlate positively the observed durations, and the GAM and NPSF durations correlate even more strongly with each other.
tester <- data.frame(y=martinvanberg$formdur, npsf=ed1$exp.dur$exp.dur, gam=ed2$exp.dur$exp.dur) cor(tester)
Scatterplots visualize these correlations:
To illustrate the use of
coxed() with time-varying covariates, we use another working example. To set up this example, we quote from the online appendix to Kropko and Harden (2018):
Box-Steffensmeier (1996) examines whether U.S. House incumbents’ ability to raise campaign funds can effectively deter quality challengers from entering the race. The theoretical expectation is that as incumbents raise more money, challengers further delay their decision to run for the incumbent’s seat. She employs data on 397 House races in the 1989–1990 election cycle to test this hypothesis. The dependent variable in this analysis is the number of weeks after January 1, 1989 when a challenger entered the race. Races in which no challenger entered are coded as the number of weeks after January 1 when the state’s primary filing deadline occurred, and are treated as censored. The key independent variable is the incumbent’s War chest, or the amount of money in millions of dollars that the incumbent has in reserve at a given time. Importantly, this measure updates over the course of five Federal Election Commission (FEC) reporting periods, so it is a time-varying covariate (TVC). The theory predicts a negative coefficient on this variable, which would indicate that as the incumbent raises more money, the hazard of challenger entry declines (and the time until entry increases).
Box-Steffensmeier's model is replicated below. Note that the
Surv() function which sets up the dependent variable has two time arguments, representing the start and end of discrete intervals, which allows a covariate to take on different values across different intervals for the same observation.
bs.surv <- Surv(time = boxsteffensmeier$start, time2 = boxsteffensmeier$te, event = boxsteffensmeier$cut_hi) bs.cox <- coxph(bs.surv ~ ec + dem + south + iv, data = boxsteffensmeier, method = "breslow") summary(bs.cox)
coxed() function automatically detects whether time-varying covariates are used in the model and it takes steps to account for this structure in predicting expected durations and in estimating marginal effects. The only additional step that the user needs to take is to specify the ID variable in the
id argument, so that the function knows which intervals refer to which observations. The
id variable must be the ID from the data that was used to estimate the Cox PH model, and NOT the ID variable in any new data frame.
ed1 <- coxed(bs.cox, method="npsf", id=boxsteffensmeier$caseid) summary(ed1, stat="mean")
Here we look directly at the effect of the war chest on the length of time until a high quality challenger enters the race. We compare the 25th and 75th percentiles in war chest variable:
me <- coxed(bs.cox, method="npsf", newdata = mutate(boxsteffensmeier, ec=quantile(ec, .25)), newdata2 = mutate(boxsteffensmeier, ec=quantile(ec, .75)), id=boxsteffensmeier$caseid) summary(me, stat="mean") summary(me, stat="median")
An incumbent whose war chest is at the 75th percentile in the data delays the entry of a high quality challenger by 2.7 weeks (or 2.4 weeks when evaluated at the medians), on average, compared to an incumbent whose war chest is at the 25th percentile.
Box-Steffensmeier, J. M. (1996) "A Dynamic Analysis of The Role of War Chests in Campaign Strategy." American Journal of Political Science 40: 352-371.
Kropko, J. and Harden, J. J. (2018) "Beyond the Hazard Ratio: Generating Expected Durations from the Cox Proportional Hazards Model." British Journal of Political Science https://doi.org/10.1017/S000712341700045X
Martin, L. W and Vanberg, G. (2003) "Wasting Time? The Impact of Ideology and Size on Delay in Coalition Formation." British Journal of Political Science 33 323-344 https://doi.org/10.1017/S0007123403000140
Any scripts or data that you put into this service are public.
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.