I have questions about multivariable cox regression analysis including non-binary categorical variables.
My data consists of several variables, and some of them are binary (like sex, and age over 70, etc..)
whereas the rest of them are not (for example, ECOG)
I tried both analyse_multivariate function and coxph function, but it seems that I can only get overall hazard ratios regarding non-categorical variables, but I'd like to know both overall hazard ratios for the variable and individual hazard ratios for the subcategories in the variable (like hazard ratios for ECOG 0, ECOG 1, ECOG 2, and for overall ECOG)
What I tried in the process is like this:
(1)
ECOG = as.factor(df$ECOG)
analyse_multivariate(data=df,
time_status = vars(df$OS, df$survival_status==1),
covariates = vars(df$age70, df$sex, ECOG),
reference_level_dict = c(ECOG==0))
and the result is like this:
Hazard Ratios:
factor.id factor.name factor.value HR Lower_CI Upper_CI Inv_HR Inv_Lower_CI Inv_Upper_CI
df$age70 df$age70 <continuous> 1.07 0.82 1.41 0.93 0.71 1.22
ECOG:4 ECOG 4 1.13 0.16 8.19 0.89 0.12 6.43
df$sex df$sex <continuous> 1.87 0.96 3.66 0.53 0.27 1.04
ECOG:1 ECOG 1 2.14 1.63 2.81 0.47 0.36 0.61
ECOG:3 ECOG 3 12.12 7.83 18.76 0.08 0.05 0.13
ECOG:2 ECOG 2 13.72 4.92 38.26 0.07 0.03 0.2
(2)
analyse_multivariate(data=df,
time_status = vars(df$OS, df$survival_status==1),
covariates = vars(df$age70, df$sex, df$ECOG),
reference_level_dict = c(ECOG==0))
and the result is:
Hazard Ratios:
factor.id factor.name factor.value HR Lower_CI Upper_CI Inv_HR Inv_Lower_CI Inv_Upper_CI
df$age70 df$age70 <continuous> 0.89 0.68 1.16 1.13 0.86 1.47
df$sex df$sex <continuous> 1.87 0.96 3.65 0.53 0.27 1.04
df$ECOG df$ECOG <continuous> 1.9 1.69 2.15 0.53 0.47 0.59
Does it make sense if I use a p-value for ECOG in total from (2) and consider ECOG as a significant variable if its p-value is <0.05, and combine individual hazard ratios for individual ECOG status from (1)?
like for generating a table like followings:
p-value 0.01
ECOG 1 Reference
ECOG 2 13.72 (4.92-38.26)
ECOG 3 12.12 (7.83-18.76)
ECOG 4 1.13 (0.16-8.19)
I believe there are better solutions but couldn't find one.
Any comments would be appreciated!
Thank you in advance.
Short answer is no. In (2), it is a continuous response, meaning you expect the log odds ratio of survival to have a linear relationship with ECOG, whereas in (1) you expect every level (1 to 4) to have a different effect on survival. To test the variable ECOG collective, you can do an anova:
library(survivalAnalysis)
data = survival::lung
data$ECOG = factor(data$ph.ecog)
data$sex = factor(data$sex)
fit1 = data %>%
analyse_multivariate(vars(time, status),
covariates = vars(age, sex, ECOG, wt.loss))
anova(fit1$coxph)
Analysis of Deviance Table
Cox model: response is Surv(time, status)
Terms added sequentially (first to last)
loglik Chisq Df Pr(>|Chi|)
NULL -675.02
age -672.36 5.3325 1 0.020931 *
sex -667.82 9.0851 1 0.002577 **
ECOG -660.26 15.1127 3 0.001723 **
wt.loss -659.31 1.9036 1 0.167680
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Related
Whenever I use manova(), then summary.aov(), I only get df, Sum sq, and Mean Sq, with no p value.
My data frame looks like: (sorry I'm not sure if there's a better way to display this!)
subtype lymphocytosis anemia thrombocytopenia eosinophilia hypercalcemia hyperglobulinemia
1 MBC 0.60 0.18 0.17 0.02 0.01 0.04
2 SBC 0.25 0.18 0.14 0.03 0.02 0.12
3 BCLL 1.00 0.29 0.18 0.08 0.03 0.21
neutrophilia neutropenia lymphadenopathy_peripheral lymphadenopathy_visceral splenomegaly
1 0.23 0.02 1.00 0.65 0.60
2 0.22 0.04 0.99 0.62 0.49
3 0.23 0.04 0.40 0.25 0.49
hepatomegaly pleural_effusion peritoneal_effusion intestinal_mass mediastinal_mass pulmonary_mass
1 0.41 0.02 0.05 0.10 0.09 0.22
2 0.37 0.03 0.05 0.17 0.12 0.22
3 0.27 0.01 0.04 0.25 0.03 0.25
The values in the data frame represent the mean number of cases from each subtype for each clinical sign. I am a little worried that, for manova() to work, I should have each individual case and their clinical signs inputted so that manova can do its own math? Which would be a huge pain for me to assemble, hence why I've done it this way. Either way, I still think I should bet getting P values, they just might be wrong if my data frame is wrong?
The code I am using is:
cs_comp_try <- manova(cbind(lymphocytosis, anemia, thrombocytopenia, eosinophilia, hypercalcemia,
hyperglobulinemia, neutrophilia, neutropenia, lymphadenopathy_peripheral, lymphadenopathy_visceral,
splenomegaly, hepatomegaly, pleural_effusion, peritoneal_effusion, intestinal_mass, mediastinal_mass, pulmonary_mass) ~ subtype, data = cs_comp)
summary(cs_comp_try)
summary.aov(cs_comp_try)
The result I get for summary.aov() is:
Response peritoneal_effusion :
Df Sum Sq Mean Sq
subtype 2 6.6667e-05 3.3333e-05
Response intestinal_mass :
Df Sum Sq Mean Sq
subtype 2 0.011267 0.0056333
Response mediastinal_mass :
Df Sum Sq Mean Sq
subtype 2 0.0042 0.0021
Response pulmonary_mass :
Df Sum Sq Mean Sq
subtype 2 6e-04 3e-04
I think I've replicated all the examples I've seen on the internet, so I'm not sure why I'm not getting an F statistic and p value when I run this code.
You can just use the summary function to get the p-values like this (I use iris data as an example):
fit <- manova(cbind(Sepal.Length, Petal.Length) ~ Species, data = iris)
summary(fit)
#> Df Pillai approx F num Df den Df Pr(>F)
#> Species 2 0.9885 71.829 4 294 < 2.2e-16 ***
#> Residuals 147
#> ---
#> Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Created on 2022-07-15 by the reprex package (v2.0.1)
If you want to extract the actual p-values, you can use the following code:
fit <- manova(cbind(Sepal.Length, Petal.Length) ~ Species, data = iris)
summary(fit)$stats[1, "Pr(>F)"]
#> [1] 2.216888e-42
Created on 2022-07-15 by the reprex package (v2.0.1)
I have plotted the conditional density distribution of my variables by using cdplot (R). My independent variable and my dependent variable are not independent. Independent variable is discrete (it takes only certain values between 0 and 3) and dependent variable is also discrete (11 levels from 0 to 1 in steps of 0.1).
Some data:
dat <- read.table( text="y x
3.00 0.0
2.75 0.0
2.75 0.1
2.75 0.1
2.75 0.2
2.25 0.2
3 0.3
2 0.3
2.25 0.4
1.75 0.4
1.75 0.5
2 0.5
1.75 0.6
1.75 0.6
1.75 0.7
1 0.7
0.54 0.8
0 0.8
0.54 0.9
0 0.9
0 1.0
0 1.0", header=TRUE, colClasses="factor")
I wonder if my variables are appropriate to run this kind of analysis.
Also, I'd like to know how to report this results in an elegant way with academic and statistical sense.
This is a run using the rms-packages `lrm function which is typically used for binary outcomes but also handles ordered categorical variables:
library(rms) # also loads Hmisc
# first get data in the form you described
dat[] <- lapply(dat, ordered) # makes both columns ordered factor variables
?lrm
#read help page ... Also look at the supporting book and citations on that page
lrm( y ~ x, data=dat)
# --- output------
Logistic Regression Model
lrm(formula = y ~ x, data = dat)
Frequencies of Responses
0 0.54 1 1.75 2 2.25 2.75 3 3.00
4 2 1 5 2 2 4 1 1
Model Likelihood Discrimination Rank Discrim.
Ratio Test Indexes Indexes
Obs 22 LR chi2 51.66 R2 0.920 C 0.869
max |deriv| 0.0004 d.f. 10 g 20.742 Dxy 0.738
Pr(> chi2) <0.0001 gr 1019053402.761 gamma 0.916
gp 0.500 tau-a 0.658
Brier 0.048
Coef S.E. Wald Z Pr(>|Z|)
y>=0.54 41.6140 108.3624 0.38 0.7010
y>=1 31.9345 88.0084 0.36 0.7167
y>=1.75 23.5277 74.2031 0.32 0.7512
y>=2 6.3002 2.2886 2.75 0.0059
y>=2.25 4.6790 2.0494 2.28 0.0224
y>=2.75 3.2223 1.8577 1.73 0.0828
y>=3 0.5919 1.4855 0.40 0.6903
y>=3.00 -0.4283 1.5004 -0.29 0.7753
x -19.0710 19.8718 -0.96 0.3372
x=0.2 0.7630 3.1058 0.25 0.8059
x=0.3 3.0129 5.2589 0.57 0.5667
x=0.4 1.9526 6.9051 0.28 0.7773
x=0.5 2.9703 8.8464 0.34 0.7370
x=0.6 -3.4705 53.5272 -0.06 0.9483
x=0.7 -10.1780 75.2585 -0.14 0.8924
x=0.8 -26.3573 109.3298 -0.24 0.8095
x=0.9 -24.4502 109.6118 -0.22 0.8235
x=1 -35.5679 488.7155 -0.07 0.9420
There is also the MASS::polr function, but I find Harrell's version more approachable. This could also be approached with rank regression. The quantreg package is pretty standard if that were the route you chose. Looking at your other question, I wondered if you had tried a logistic transform as a method of linearizing that relationship. Of course, the illustrated use of lrm with an ordered variable is a logistic transformation "under the hood".
JAGS
I have an intercept-only logistic model in JAGS, defined as follows:
model{
for(i in 1:Ny){
y[i] ~ dbern(mu[s[i]])
}
for(j in 1:Ns){
mu[j] <- ilogit(b0[j])
b0[j] ~ dnorm(0, sigma)
}
sigma ~ dunif(0, 100)
}
When I plot the posterior distribution of b0 collapsing across all subjects (i.e., all b0[j]), my 95% HDI includes 0: -0.55 to 2.13. The Effective Sample Size is way above 10,000 for every b0 (around 18,000 on average). Diagnostics look good.
glmer()
Now, this is the equivalent glmer() model:
glmer(response ~ 1 + (1|subject), data = myData, family = "binomial")
The result of this model, however, is as follows:
Random effects:
Groups Name Variance Std.Dev.
speaker (Intercept) 0.3317 0.576
Number of obs: 1544, groups: subject, 27
Fixed effects:
Estimate Std. Error z value Pr(>|z|)
(Intercept) 0.7401 0.1247 5.935 2.94e-09 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
So here it says my estimate is significantly above 0.
What the data look like
Here are the proportions of 0s and 1s by subject. You can see that, for the vast majority of subjects, the proportion of 1 is above 50%.
Any ideas why JAGS and glmer() are so different here?
0 1
1 0.47 0.53
2 0.36 0.64
3 0.29 0.71
4 0.42 0.58
5 0.12 0.88
6 0.22 0.78
7 0.54 0.46
8 0.39 0.61
9 0.30 0.70
10 0.32 0.68
11 0.36 0.64
12 0.66 0.34
13 0.38 0.62
14 0.49 0.51
15 0.35 0.65
16 0.32 0.68
17 0.12 0.88
18 0.45 0.55
19 0.36 0.64
20 0.36 0.64
21 0.28 0.72
22 0.40 0.60
23 0.41 0.59
24 0.19 0.81
25 0.27 0.73
26 0.08 0.92
27 0.12 0.88
You forgot to include a mean value, so your intercept parameter is fixed to zero. Something like this should work:
model{
for(i in 1:Ny){
y[i] ~ dbern(mu[s[i]])
}
for(j in 1:Ns){
mu[j] <- ilogit(b0[j])
b0[j] ~ dnorm(mu0, sigma)
}
mu0 ~ dnorm(0,0.001)
sigma ~ dunif(0, 100)
}
Now the posterior density of mu0 should match the sampling distribution of the intercept parameter from glmer reasonably well.
Alternatively, if you use response ~ -1 + (1|subject) as your glmer formula, you should get results that match your current JAGS model.
According to the documentation of the mice package, if we want to impute data when we're interested in interaction terms we need to use passive imputation. Which is done the following way.
library(mice)
nhanes2.ext <- cbind(nhanes2, bmi.chl = NA)
ini <- mice(nhanes2.ext, max = 0, print = FALSE)
meth <- ini$meth
meth["bmi.chl"] <- "~I((bmi-25)*(chl-200))"
pred <- ini$pred
pred[c("bmi", "chl"), "bmi.chl"] <- 0
imp <- mice(nhanes2.ext, meth = meth, pred = pred, seed = 51600, print = FALSE)
It is said that
Imputations created in this way preserve the interaction of bmi with chl
Here, a new variable called bmi.chl is created in the original dataset. The meth step tells how this variable needs to be imputed from the existing ones. The pred step says we don't want to predict bmi and chl from bmi.chl. But now, if we want to apply a model, how do we proceed? Is the product defined by "~I((bmi-25)*(chl-200))" is just a way to control for the imputed values of the main effects, i.e. bmi and chl?
If the model we want to fit is glm(hyp~chl*bmi, family="binomial"), what is the correct way to specify this model from the imputed data? fit1 or fit2?
fit1 <- with(data=imp, glm(hyp~chl*bmi, family="binomial"))
summary(pool(fit1))
Or do we have to use somehow the imputed values of the new variable created, i.e. bmi.chl?
fit2 <- with(data=imp, glm(hyp~chl+bmi+bmi.chl, family="binomial"))
summary(pool(fit2))
With passive imputation, it does not matter if you use the passively imputed variable, or if you re-calculate the product term in your call to glm.
The reason that fit1 and fit2 yield different results in your example is because are not just doing passive imputation for the product term.
Instead you are transforming the two variables befor multiplying (i.e., you calculate bmi-25 and chl-100). As a result, the passively imputed variable bmi.chl does not represent the product term bmi*chl but rather (bmi-25)*(chl-200).
If you just calculate the product term, then fit1 and fit2 yield the same results like they should:
library(mice)
nhanes2.ext <- cbind(nhanes2, bmi.chl = NA)
ini <- mice(nhanes2.ext, max = 0, print = FALSE)
meth <- ini$meth
meth["bmi.chl"] <- "~I(bmi*chl)"
pred <- ini$pred
pred[c("bmi", "chl"), "bmi.chl"] <- 0
pred[c("hyp"), "bmi.chl"] <- 1
imp <- mice(nhanes2.ext, meth = meth, pred = pred, seed = 51600, print = FALSE)
fit1 <- with(data=imp, glm(hyp~chl*bmi, family="binomial"))
summary(pool(fit1))
# > round(summary(pool(fit1)),2)
# est se t df Pr(>|t|) lo 95 hi 95 nmis fmi lambda
# (Intercept) -23.94 38.03 -0.63 10.23 0.54 -108.43 60.54 NA 0.41 0.30
# chl 0.10 0.18 0.58 9.71 0.58 -0.30 0.51 10 0.43 0.32
# bmi 0.70 1.41 0.49 10.25 0.63 -2.44 3.83 9 0.41 0.30
# chl:bmi 0.00 0.01 -0.47 9.67 0.65 -0.02 0.01 NA 0.43 0.33
fit2 <- with(data=imp, glm(hyp~chl+bmi+bmi.chl, family="binomial"))
summary(pool(fit2))
# > round(summary(pool(fit2)),2)
# est se t df Pr(>|t|) lo 95 hi 95 nmis fmi lambda
# (Intercept) -23.94 38.03 -0.63 10.23 0.54 -108.43 60.54 NA 0.41 0.30
# chl 0.10 0.18 0.58 9.71 0.58 -0.30 0.51 10 0.43 0.32
# bmi 0.70 1.41 0.49 10.25 0.63 -2.44 3.83 9 0.41 0.30
# bmi.chl 0.00 0.01 -0.47 9.67 0.65 -0.02 0.01 25 0.43 0.33
This is not surprising because the ~I(bmi*chl) in mice and the bmi*chl in glm do the exact same thing. They merely calculate the product of the two variables.
Remark:
Note that I added a line saying that bmi.chl should be used as a predictor when imputing hyp. Without this step, passive imputation has no purpose because the imputation model would neglect the product term, thus being incongruent with the analysis model.
I am doing a factor analysis via the psych-package, which generates quite some output fa8<-fa(corMat, nfactors=8, ...)
The output includes some matrices and some text information. However, I did not find a good way of saving the matrices of the output to a file. So far, I was able to dump the data via sink("foo.txt"); f8; sink() as complete output. Neither write(fa8) nor write.csv(fa8) work, because the class of the output is a vector - it does not contain the matrix data itself, though.
Any suggestions on how I can get the fa-matrix itself for further analysis and saving it to a file?
update #1:
An examplatory output of fa(corMat, nfactors=2, ...)would be
Factor Analysis using method = pa
Call: fa(r = corMat, nfactors = 2, rotate = "oblimin", fm = "pa")
Standardized loadings based upon correlation matrix
PA1 PA2 h2 u2
BIO 0.86 0.02 0.75 0.255
GEO 0.78 0.05 0.63 0.369
CHEM 0.87 -0.05 0.75 0.253
ALG -0.04 0.81 0.65 0.354
CALC 0.01 0.96 0.92 0.081
STAT 0.13 0.50 0.29 0.709
PA1 PA2
SS loadings 2.14 1.84
Proportion Var 0.36 0.31
Cumulative Var 0.36 0.66
With factor correlations of
PA1 PA2
PA1 1.00 0.21
PA2 0.21 1.00
Test of the hypothesis that 2 factors are sufficient.
The degrees of freedom for the null model are 15 and the objective function was 2.87
The degrees of freedom for the model are 4 and the objective function was 0.01
The root mean square of the residuals is 0.01
The df corrected root mean square of the residuals is 0.02
Fit based upon off diagonal values = 1
Measures of factor score adequacy
PA1 PA2
Correlation of scores with factors 0.94 0.96
Multiple R square of scores with factors 0.88 0.93
Minimum correlation of possible factor scores 0.77 0.86
Source: http://rtutorialseries.blogspot.de/2011/10/r-tutorial-series-exploratory-factor.html
The question is: How do I get the standardized loadings matrix in the output for further analysis?