Uncorrected estimates
Hypertension is defined within one individual if the mean of its BP
measurements (for diastolic or systolic BP) exceeds a threshold, or if
the patient takes an treatment for HT.
The crude estimation of the prevalence of hypertension consists in
counting hypertensive patients from the data. This can be done using
the correct_htn function:
correct_htn(form=htn ~ 1,
data_long = dt_nhanes,
correct = FALSE)
## # A tibble: 1 x 4
## htn se low up
## <dbl> <dbl> <dbl> <dbl>
## 1 0.263 0.00357 0.256 0.270
The sampling design of the NHANESIII survey may be taken into account
using the surv_des argument of the function:
## First, define the design
d.s <- svydesign(ids=~1,
data=dt_nhanes,
weights=~samp_weight)
## Then, use it
correct_htn(form=htn ~ 1,
data_long = dt_nhanes,
surv_des = d.s,
correct = FALSE)
## # A tibble: 1 x 4
## htn se low up
## <dbl> <dbl> <dbl> <dbl>
## 1 0.195 0.00446 0.187 0.204
It worth saying that in example above, the specification of the survey
design was loose, as applied to data given in long format, whereas it
should have been defined from data in a wide format instead. This
doesn’t matter actually, as the correct_htn function only keeps the
specification of the survey design, and applies it to data with only
one observation per individual.
This point is confirmed bellow:
## First, define the design
dt_nhanes_wide<-
dt_nhanes%>%
ungroup()%>%
mutate(vm=paste("bp",visit,meas,sep="_"))%>%
select(-visit,-meas)%>%
spread(vm,bp)
d.s.w <- svydesign(ids=~1,
data=dt_nhanes_wide,
weights=~samp_weight)
## Then, use it
correct_htn(form=htn ~ 1,
data_long = dt_nhanes,
surv_des = d.s.w,
correct = FALSE)
## # A tibble: 1 x 4
## htn se low up
## <dbl> <dbl> <dbl> <dbl>
## 1 0.195 0.00446 0.187 0.204
The correct_htn function also allows for estimations of HT
prevalence according to covariates, for example, according to body
mass index (variable bmi).
Note that this estimation can be done only in the subpopulation with
non missing value of BMI, so that we should restrict or estimates to
this subpop. This is done using the subpop argument of the correct_htn.
## Define boolen for the subpop with not missing bmi
dt_nhanes%<>%mutate(has_bmi=!is.na(bmi))
## Estimate the variation of HT according to bmi with a natural spline of 3 df
bmi_htn<-correct_htn(form=htn ~ ns(bmi,df=3),
data_long = dt_nhanes,
subpop=~has_bmi,
surv_des = d.s,
correct = FALSE)
## Plot the results
bmi_htn%>%
mutate_at(vars(htn,low,up),~.x*100)%>%
ggplot(data=.)+
aes(bmi,htn,ymin=low,ymax=up)+
geom_line()+
geom_ribbon(fill=NA,colour="black",linetype=5)+
xlab("Body mass index")+ylab("HT prevalence (in %)")
Corrected estimates
Taking into account the within-person blood pressure (BP) variability in the estimation of
hypertension is simply done passing the correct argument from the
correct_htn function to TRUE. The number of samples from the
posterior distribution used for the estimation is controlled by the
n_samp argument. As the estimation time increases with the number of
samples, we’ll keep it low (n_samp=100) in this illustrative application here.
To illustrate the consequences of neglecting within-person blood
pressure variability in the estimates, we estimated HT prevalence
using the two visits from NHANESIII, or only the first or the second.
dt_nhanes_v1<-dt_nhanes%>%filter(visit==1)
dt_nhanes_v2<-dt_nhanes%>%filter(visit==2)
htn_cor<-correct_htn(form=htn ~ 1,
data_long = dt_nhanes,
surv_des = d.s,
n_samp = 100,
correct = TRUE)
htn_cor_v1<-correct_htn(form=htn ~ 1,
data_long = dt_nhanes_v1,
surv_des = d.s,
n_samp = 100,
correct = TRUE)
htn_cor_v2<-correct_htn(form=htn ~ 1,
data_long = dt_nhanes_v2,
surv_des = d.s,
n_samp = 100,
correct = TRUE)
htn_uncor<-correct_htn(form=htn ~ 1,
data_long = dt_nhanes,
surv_des = d.s,
n_samp = 1,
correct = FALSE)
htn_uncor_v1<-correct_htn(form=htn ~ 1,
data_long = dt_nhanes_v1,
surv_des = d.s,
n_samp = 1,
correct = FALSE)
htn_uncor_v2<-correct_htn(form=htn ~ 1,
data_long = dt_nhanes_v2,
surv_des = d.s,
n_samp = 1,
correct = FALSE)
The results are reported in the table below:
Table 2.1: Corrected and un-corrected estimates of HT according to the visits used
|
Visit
|
Un-corrected HT
|
Corrected HT
|
|
All
|
19.53 [18.67 - 20.42]
|
17.80 [17.00 - 18.64]
|
|
Visit 1
|
22.84 [21.89 - 23.82]
|
19.34 [18.49 - 20.22]
|
|
Visit 2
|
19.94 [19.08 - 20.84]
|
16.91 [16.12 - 17.72]
|
The table shows the importance of accounting for BP variability when
estimating HT prevalence. If no correction is performed, HTN estimated
with visit 1 or visit 2 data only are larger than HTN estimated with
the two visits. This latter however also exhibits some bias, corrected
HTN (i.e. HT prevalence estimated with BP collected over an infinite
number visits and measures in the population) being 9% lower than
uncorrected HTN. As expected, the overestimation of HTN is higher
(~15%) when only one visit is used for the estimation.
Some differences however still remains between visit 1 and visit 2
estimates, probably due to lower BP reactivity to physicians at the
second visit (lower white coat effect) and to differences in
conditions under which BP was measured in the two visits. Without
additional information on the reasons for those differences however,
there is no indication about which visit to favor and no obvious
additional correction to apply.
Similarly, correct_htn can be used to estimating the level of
corrected HT prevalence at different BMI levels with the same syntax
as in the previous section, but specifying correct = TRUE in the
arguments.
The corrected curve (in blue) is superimposed with the previous uncorrected
curve (in red).
bmi_htn_correct<-correct_htn(form=htn ~ ns(bmi,df=3),
data_long = dt_nhanes,
subpop=~has_bmi,
surv_des = d.s,
n_samp = 100,
correct = TRUE)
