(C) PLOS One
This story was originally published by PLOS One and is unaltered.
. . . . . . . . . .
The effect of population-based blood pressure screening on long-term cardiometabolic morbidity and mortality in Germany: A regression discontinuity analysis [1]
['Sara Pedron', 'Professorship Of Public Health', 'Prevention', 'Technical University Of Munich', 'Munich', 'Institute Of Health Economics', 'Health Care Management', 'Helmholtz Zentrum München', 'German Research Center For Environmental Health', 'Gmbh']
Date: 2023-01
We imputed missing values in the outcomes and covariates using a comprehensive model with a predictive mean matching approach with 20 replications using the mice package in R [ 23 ]. Further details on the imputation model and diagnostics are available in S2 Appendix . In a sensitivity analysis, we assessed the robustness of estimation results using the original sample without imputed values.
In each MONICA/KORA study, participants underwent several medical examinations, blood tests, a structured interview, and a self-administered questionnaire. For the present analysis, we excluded individuals who lacked data for all relevant variables or who had a history of CVD events at baseline (stroke or myocardial infarction). Additionally, we excluded those who were already taking BP-lowering drugs at baseline, measured by a computer-assisted drug recording procedure (see S1 Appendix for details). Additionally, for the secondary outcomes sample, we excluded individuals who lacked follow-up data and individuals who lacked data on one or more of the analyzed outcomes.
Specifically, we pooled data from the three cross-sectional WHO MONICA surveys S1 (1984/1985; n = 4,022), S2 (1989/1990; n = 4,940), S3 (1994/1995; n = 4,856), and the subsequent KORA survey S4 (1999 to 2001; n = 4,261). For the secondary outcomes analysis, we used the follow-up studies to S3 and S4, namely, KORA-F3 (2004 to 2005, n = 2,586) and KORA-F4 (2006 to 2008, n = 2,544), respectively. Furthermore, we used the CVD mortality and morbidity follow-up data that were collected longitudinally for all participants of the four cohorts. Recruitment and data collection in all MONICA and KORA studies followed a very similar, nearly identical, protocol. Details of the KORA study can be found in Holle and colleagues’ paper [ 22 ]. All KORA studies were carried out in accordance with ethical regulations at the time the studies were initiated. The KORA studies F3, S4, and F4 were approved by the Ethics Committee of the Bavarian Medical Association (Ethics number: F3 03097, S4 99186, F4 06068). Written informed consent was obtained by all study participants.
We used data from the population-based KORA platform (Cooperative Health Research in the Augsburg Region). The base for the KORA study consists of three World Health Organization (WHO) MONICA (Monitoring Trends and Determinants in Cardiovascular Disease) studies (S1, S2, and S3) that were started during the 1980s to 1990s as a general effort of WHO to generate population-based epidemiological data in several areas in Europe [ 22 ]. The study conducted in the city of Augsburg was then carried on and further developed within the KORA study, which added morbidity registries and follow-up studies and included new cohorts.
The exposure of interest was whether the participant’s BP result was above the designated threshold for hypertension (diastolic BP ≥ 90 mm Hg or systolic BP ≥ 140 mm Hg). Whereas all participants were screened for hypertension and received communication on their results, only those with BP above the threshold received an additional prompt with their hypertension diagnosis and encouragement to seek further care (see S3 Appendix for details). We compare participants diagnosed with hypertension with those not diagnosed.
BP was measured in all MONICA/KORA surveys following a standard procedure [ 25 ], which involved three sequential measurements, two of which were performed after at least 30 minutes rest. We extracted data on the value of the participants systolic BP and diastolic BP at the screening. In accordance with the value communicated to participants, we extracted data on the third BP measurement for participants of the S1 to S3 studies, and we computed the average of the second and third BP measurement for participants of the S4 study.
As such, the assessed intervention is an information treatment targeted at high-risk individuals, following a BP screening within a population-based epidemiological study [ 24 ]. The intervention represents a one-time occurrence outside of the usual healthcare context and is therefore different than regular screening interventions carried out as part of primary care.
All participants in the KORA study received clinical screening for BP, body mass index (BMI), cholesterol, and other parameters ( S3 Appendix ). Approximately 2 weeks after the survey, all participants received a letter with the results of their clinical screening and encouragement to seek a doctor if values were above recommended thresholds. For this study, we focus on communication of the results of the BP screening, which we consider a “light touch” intervention since individuals were simply informed of their diagnosis and encouraged to consult a physician through their own initiative.
As secondary outcomes, we analyzed systolic and diastolic BP, hypertension awareness, physical activity, smoking, and BMI. These outcomes were assumed to lie on the causal pathway between information treatment and CVD outcomes (intermediate outcomes) and were measured in the 7-year follow-up studies for the S3 and S4 studies (i.e., F3 and F4, respectively). No follow-up data were available for the studies S1 and S2. Previous hypertension diagnosis was measured by asking the question “Have you ever been diagnosed with hypertension?” (“Yes,” No,” “I don’t know”). The original question in German was: “Ist bei Ihnen jemals hoher Blutdruck festgestellt worden?”, which translates as “Have you ever been diagnosed with hypertension?”, but also includes the possibility of a diagnosis from someone other than the physician (oneself, a relative, a pharmacist, etc.) because the agent is not specified. Both current smoking status and physical activity levels were self-reported. Frequency of physical activity was collected in four categories (“No physical activity,” “Irregularly about 1 hour per week,” “Regularly about 1 hour per week,” “Regularly more than 2 hours per week”). We dichotomized this variable, considering individuals performing physical activity regularly at least 1 hour per week as “high physical activity,” and “low physical activity” otherwise. BMI was computed based on objectively measured height and weight.
Furthermore, we investigated the occurrence of any CVD event, including both information on mortality (fatal CVD event) and morbidity (nonfatal CVD event) in one outcome variable. The nonfatal incident CVD events were assessed through the population-based Augsburg Coronary Event Registry or by postal follow-up questionnaires and were validated using data from participants’ hospital records and their attending physician [ 26 – 28 ].
The primary outcome we considered was the incidence of any fatal CVD event (stroke, myocardial infarction, coronary heart disease and other CVD events–ICD-9 codes: 390 to 459, 798) within up to 16.9 years of follow-up (i.e., the longest common follow-up across surveys). Death certificates were received from local health authorities to confirm these events.
Study design and statistical analysis
We evaluated the impact of a hypertension diagnosis and encouragement letter on health outcomes using a regression discontinuity (RD) design. RD enables causal inference when an intervention is assigned based on a threshold rule. We compared KORA study participants who had baseline BP readings just below the threshold for hypertension (and therefore did not receive an information treatment comprising hypertension diagnosis and encouragement to seek care) (control group) with study participants who had baseline BP readings just above the threshold (and therefore received hypertension diagnosis and an encouragement prompt) (intervention group). Due to random variability in BP measurements, participants just above/below the threshold, apart from receiving the information treatment, can be assumed to be similar on both observed and unobserved baseline characteristics, as in a randomized trial [29,30]. This means that differences outside of the study context, such as differing access and standards of care, should also be balanced across intervention and control groups.
In the analysis, individuals who are just above/below the cutoff were selected using a bandwidth, which determines the span of the distribution around the cutoff in terms of the assignment variable to be considered in the analysis (e.g., a bandwidth of 5 mm Hg would imply considering individuals between 135 mm Hg and 145 mm Hg for systolic BP). The optimal bandwidth was computed by optimizing the trade-off between being in the closest proximity of the cutoff to improve comparability of observations, while at the same time ensuring that enough observations are considered in the analysis [29,30].
Hypertension is diagnosed based on a compound threshold rule including both systolic and diastolic BP. Participants with a diastolic BP ≥ 90 mm Hg or a systolic BP ≥ 140 mm Hg received the information that their BP was too high and advice that they should seek medical attention. Because a participant could be diagnosed due to elevated levels of either diastolic BP (dBP) or systolic BP (sBP), we combined the two values into a single assignment variable with a single threshold rule or binding score (Blood Pressure Score—BPS) [31]. First, we created the single running variable by centering the two distinct diastolic and systolic BP variables at the respective cutoffs and divided the values by the respective standard deviation so that diastolic BP and systolic BP would have a similar density for observations in a close proximity around the threshold (i.e., within a given bandwidth around the threshold). Second, to ensure that each individual contributed only one value, we selected the maximum value for each individual from the new running variable, while disregarding the lower value:
Based on this running variable, treatment was assigned to all individuals with a BPS ≥ 0. Individuals with a BPS < 0 were assigned no treatment and represent thus the control group.
This reformulation to a single running variable has a major advantage compared to the scenario with two running variables, one for diastolic BP and one for systolic BP. It allows estimating the effect of treatment in the context of a sharp RD design, where no individual below the cutoff received the intervention and all individuals above the cutoff received the intervention. The sharp design allows the estimation of the intention-to-treat effect (ITT) at the threshold, i.e., for those individuals in a close proximity to the threshold [29]. Considering the two assignment variables separately would have led to the case that some individuals were above the respective threshold for one BP dimension and below the threshold for the other—i.e., such an individual would have been assigned to receive the intervention according to the first BP dimension, yet assigned to the control for the latter. This situation would have required the estimation of two separate fuzzy RD designs, yielding the estimation of a local average treatment effect (LATE), only valid for compliers and thus not generalizable to the whole population [29,30].
One potential threat to validity of the RD designs is the possibility that the assignment variable (in our case, BP as measured in the study center) might be manipulated, e.g., to gain access to treatment [29,30]. This would occur if, for example, the KORA staff member conducting the physical examinations falsely registered systolic or diastolic BP measurements that were just below the threshold. This is unlikely as the study-based clinical exam was not part of routine patient care, so there was little motivation to manipulate screening values. We tested this assumption by graphically inspecting the density plot of the assignment variable and its components for each sample utilized in the analysis.
Furthermore, RD designs yield causal inferences because participants are comparable on either side of the threshold [29,30]. Similar to a balance table in a randomized controlled trial (RCT), we can show similarity on baseline observables. We estimate RD models replacing the outcome variable with baseline characteristics in order to assess differences at the threshold. Furthermore, we inspected a graphical depiction of covariates around the threshold.
Another issue regards the fact that the threshold that was adopted to warn individuals about their hypertension (diastolic/systolic BP ≥ 90/140 mm Hg) is also the threshold that is widely adopted by physicians to define and treat hypertension in the everyday clinical practice. However, given a substantial random element in the individual BP measurements, which can fluctuate by several units within hours, this does not represent a violation of the exclusion restriction and therefore does not threaten the internal validity of our estimator [32]. We used a standard linear model to estimate the effect of treatment exploiting the sharp RD design that takes on the following form: where y is the outcome and x is a continuous assignment variable, in our case the BPS. The variable cutoff is a dummy that takes on the value one if the assignment variable is above the deterministic treatment threshold, or zero otherwise. Therefore, β 0 represents the intercept, β 1 the effect of the assignment variable on the outcomes before the cutoff, and β 2 the change of this effect above the cutoff. The coefficient τ represents the treatment effect of interest, estimated within the optimal bandwidth around the threshold. X j represents a vector of j covariates (sex, age, age squared, survey wave), and the vector β j represents the respective coefficients. ε is the idiosyncratic error.
We estimated the effect of treatment on the primary outcomes using survival analysis in a Cox proportional hazard model. For the analysis of fatal CVD events, participants were censored at the first date of record of death due to CVD causes, death for other causes, or after 16.9 years, whichever came first. For the analysis of any CVD events, participants were censored at the first date of record of any (fatal or nonfatal) CVD event, death for other causes, or after 16.9 years, whichever came first. To verify the proportional hazards assumption, we plotted Schoenfeld residuals. For the secondary outcomes, we used the same model specification, using linear estimation for continuous outcomes and a logit specification for dichotomous outcomes.
We computed the optimal bandwidth around the treatment threshold using the Imbens–Kalyanaraman method, with a rectangular kernel [33]. Although this method was developed for linear models, no better methodology exists to compute such bandwidth for Cox or logit models. Therefore, we employed this method for all models. Furthermore, we tested sensitivity of our results to different bandwidth choices.
In our main analysis, the sample used for the long-term mortality and morbidity analysis differs from the sample used for the follow-up behavioral intermediate outcomes. Loss to follow-up (i.e., individuals who were discarded from the analysis because of missing values at follow-up) may have influenced the results for intermediate outcomes if individuals who died or who were not healthy enough to participate were more likely to refuse participation. For mortality and morbidity outcomes, we imputed missing data in the cause of death and CVD events, so no loss to follow-up occurred. This might have caused a mismatch for the two results. However, we chose to analyze these samples as such, because it allowed us to assess the largest number of participants for all outcomes.
The information treatment might have had different effects on individuals depending on their awareness of their hypertension status (self-reported previously diagnosed and untreated hypertension). Individuals who had already been diagnosed with hypertension before might react differently from individuals with no previous hypertension. Therefore, in a series of additional analyses, we stratified the analyses of primary outcomes by the self-reported previously diagnosed hypertension at baseline, again determined by the question “Have you ever been diagnosed with hypertension?” (“Yes,” “No,” “I don’t know”).
Finally, we tested the sensitivity of results to different bandwidths. All analyses were carried out using R (Version 4.0.3). A full variable list and the analysis code are available on the Open Science Framework (osf.io/pwgfh). For transparent reporting, we followed and compiled the TREND checklist (see S1 Checklist) [34].
[END]
---
[1] Url:
https://journals.plos.org/plosmedicine/article?id=10.1371/journal.pmed.1004151
Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.
via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/
