Against quantiles: categorization of continuous variables in epidemiologic research, and its discontents

A Frailty Index predicts 10-year fracture risk in adults age 25 years and older: results from the Canadian Multicentre Osteoporosis Study (CaMos)

We created a 30-item Frailty Index in the Canadian Multicentre Osteoporosis Study. A Frailty Index is a sensitive measure that can quantify fracture risk according to degree of frailty. Our results indicated that at any age, frailty was an important independent risk factor for fracture over 10 years.

INTRODUCTION

In later life, frailty has been linked to fractures. It is likely that the antecedents of fracture are seen across the life course, in ways not entirely captured by traditional osteoporosis risk factors. Using data collected from the prospective, population-based Canadian Multicentre Osteoporosis Study (CaMos), we created the 30-item CaMos Frailty Index and examined whether it was associated with incident fractures over 10 years.

METHODS

All CaMos participants aged 25 years and older (n = 9,423) were included in the analysis. To examine the relationship between baseline Frailty Index scores and incident fractures, a competing risk proportional sub-distribution hazards model was used with death considered a competing risk. Analyses were adjusted for age, sex, body mass index, education level, femoral neck T-score, and antiresorptive therapy.

RESULTS

At baseline, the mean age was 62.1 years [standard deviation (SD) 13.4], and 69.4 % were women. The mean Frailty Index score was 0.13 (SD 0.11), ranging from 0 to 0.66. For every 0.10 increase in Frailty Index scores (approximately one SD), the hazard ratio was 1.25 (p < 0.001) for all fractures, 1.18 (p = 0.043) for hip fractures, and 1.30 (p ≤ 0.001) for clinical vertebral fractures.

CONCLUSION

The CaMos Frailty Index quantified fracture risk according to degree of frailty. Irrespective of age and bone mineral density, the Frailty Index was associated with hip, vertebral, and all-type clinical fractures. Predicting late onset illnesses may have to consider overall health status and not just traditional risk factors.

Immune biomarkers predictive of respiratory viral infection in elderly nursing home residents

OBJECTIVE

To determine if immune phenotypes associated with immunosenescence predict risk of respiratory viral infection in elderly nursing home residents.

METHODS

Residents ≥ 65 years from 32 nursing homes in 4 Canadian cities were enrolled in Fall 2009, 2010 and 2011, and followed for one influenza season. Following influenza vaccination, peripheral blood mononuclear cells (PBMCs) were obtained and analysed by flow cytometry for T-regs, CD4+ and CD8+ T-cell subsets (CCR7+CD45RA+, CCR7-CD45RA+ and CD28-CD57+) and CMV-reactive CD4+ and CD8+ T-cells. Nasopharyngeal swabs were obtained and tested for viruses in symptomatic residents. A Cox proportional hazards model adjusted for age, sex and frailty, determined the relationship between immune phenotypes and time to viral infection.

RESULTS

1072 residents were enrolled; median age 86 years and 72% female. 269 swabs were obtained, 87 were positive for virus: influenza (24%), RSV (14%), coronavirus (32%), rhinovirus (17%), human metapneumovirus (9%) and parainfluenza (5%). In multivariable analysis, high T-reg% (HR 0.41, 95% CI 0.20-0.81) and high CMV-reactive CD4+ T-cell% (HR 1.69, 95% CI 1.03-2.78) were predictive of respiratory viral infection.

CONCLUSIONS

In elderly nursing home residents, high CMV-reactive CD4+ T-cells were associated with an increased risk and high T-regs were associated with a reduced risk of respiratory viral infection.

Critical appraisal and data extraction for systematic reviews of prediction modelling studies: the CHARMS checklist

Carl Moons and colleagues provide a checklist and background explanation for critically appraising and extracting data from systematic reviews of prognostic and diagnostic prediction modelling studies. Please see later in the article for the Editors' Summary.

Capturing the two dimensions of residential segregation at the neighborhood level for health research

Two conceptual and methodological foundations of segregation studies are that (i) segregation involves more than one group, and (ii) segregation measures need to quantify how different population groups are distributed across space. Therefore, percentage of population belonging to a group is not an appropriate measure of segregation because it does not describe how populations are spread across different areal units or neighborhoods. In principle, evenness and isolation are the two distinct dimensions of segregation that capture the spatial patterns of population groups. To portray people’s daily environment more accurately, segregation measures need to account for the spatial relationships between areal units and to reflect the situations at the neighborhood scale. For these reasons, the use of local spatial entropy-based diversity index (SH_i) and local spatial isolation index (S_i) to capture the evenness and isolation dimensions of segregation, respectively, are preferable. However, these two local spatial segregation indexes have rarely been incorporated into health research. Rather ineffective and insufficient segregation measures have been used in previous studies. Hence, this paper empirically demonstrates how the two measures can reflect the two distinct dimensions of segregation at the neighborhood level, and argues conceptually and set the stage for their future use to effectively and meaningfully examine the relationships between residential segregation and health.

Genetic ancestry is associated with colorectal adenomas and adenocarcinomas in Latino populations

Colorectal cancer rates in Latin American countries are less than half of those observed in the United States. Latin Americans are the resultant of generations of an admixture of Native American, European, and African individuals. The potential role of genetic admixture in colorectal carcinogenesis has not been examined. We evaluate the association of genetic ancestry with colorectal neoplasms in 190 adenocarcinomas, 113 sporadic adenomas and 243 age- and sex-matched controls enrolled in a multicentric case-control study in Colombia. Individual ancestral genetic fractions were estimated using the STRUCTURE software, based on allele frequencies and assuming three distinct population origins. We used the Illumina Cancer Panel to genotype 1,421 sparse single-nucleotide polymorphisms (SNPs), and Northern and Western European ancestry, LWJ and Han Chinese in Beijing, China populations from the HapMap project as references. A total of 678 autosomal SNPs overlapped with the HapMap data set SNPs and were used for ancestry estimations. African mean ancestry fraction was higher in adenomas (0.13, 95% confidence interval (95% CI)=0.11-0.15) and cancer cases (0.14, 95% CI=0.12-0.16) compared with controls (0.11, 95% CI=0.10-0.12). Conditional logistic regression analysis, controlling for known risk factors, showed a positive association of African ancestry per 10% increase with both colorectal adenoma (odds ratio (OR)=1.12, 95% CI=0.97-1.30) and adenocarcinoma (OR=1.19, 95% CI=1.05-1.35). In conclusion, increased African ancestry (or variants linked to it) contributes to the increased susceptibility of colorectal cancer in admixed Latin American population.

Sex and age differences in the relation of depressive symptoms with blood pressure

BACKGROUND

Longitudinal associations between depressive symptoms and blood pressure have been inconsistent. Most studies have examined incident hypertension as an outcome, and few have examined effect modification.

METHODS

This study examined moderating influences of sex and age on coincident trajectories of depressive symptoms and blood pressure among 2,087 participants from the Baltimore Longitudinal Study of Aging (aged 19-97 years; 53% men; 74% white). Participants underwent clinical blood pressure measurement and completed the Center for Epidemiological Studies-Depression (CES-D) scale on up to 14 occasions (mean = 3.8; SD = 2.6) over up to 29 years (mean = 7.8; SD = 6.4). CES-D was log-transformed (CES-D(log)) for analyses.

RESULTS

Mixed-effects regression revealed that prospective relations of CES-D(log) to diastolic blood pressure differed by age in women (b = 0.095; P = 0.001) but not men; greater CES-D(log) attenuated the expected age-related decline in diastolic blood pressure. Across all testing sessions, greater CES-D(log) was associated significantly with higher average systolic blood pressure for women (b = 2.238; P = 0.006) but not men. Age-stratified analyses showed that greater CES-D(log) was associated significantly with higher average systolic (b = 3.348; P = 0.02) and diastolic (b = 1.730; P < 0.03) blood pressure for older adults (≥58.8 years at first visit). In the younger age cohort, sex moderated the relation of CES-D(log) to systolic blood pressure (b = -3.563; P = 0.007); greater CES-D(log) in women, but lesser CES-D(log) in men, was associated with higher systolic blood pressure.

CONCLUSIONS

Results demonstrate sex and age differences in the relation between depressive symptoms and blood pressure. Findings suggest the potential importance of preventing, detecting, and lowering depressive symptoms to prevent hypertension among women and older adults.

A systematic review of the volume-outcome relationship for radical prostatectomy

CONTEXT

Due to the complexity and challenging nature of radical prostatectomy (RP), it is likely that both short- and long-term outcomes strongly depend on the cumulative number of cases performed by the surgeon as well as by the hospital.

OBJECTIVE

To review systematically the association between hospital and surgeon volume and perioperative, oncologic, and functional outcomes after RP.

EVIDENCE ACQUISITION

A systematic review of the literature was performed, searching PubMed, Embase, and Scopus databases for original and review articles between January 1, 1995, and December 31, 2011. Inclusion and exclusion criteria comprised RP, hospital and/or surgeon volume reported as a predictor variable, a measurable end point, and a description of multiple hospitals or surgeons.

EVIDENCE SYNTHESIS

Overall 45 publications fulfilled the inclusion criteria, where most data originated from retrospective institutional or population-based cohorts. Studies generally focused on hospital or surgeon volume separately. Although most of these analyses corroborated the impact of increasing volume with better outcomes, some failed to find any significant effect. Studies also differed with respect to the proposed volume cut-off for improved outcomes, as well as the statistical means of evaluating the volume-outcome relationship. Five studies simultaneously compared hospital and surgeon volume, where results suggest that the importance of either hospital or surgeon volume largely depends on the end point of interest.

CONCLUSIONS

Undeniable evidence suggests that increasing volume improves outcomes. Although it would seem reasonable to refer RP patients to high-volume centers, such regionalization may not be entirely practical. As such, the implications of such a shift in practice have yet to be fully determined and warrant further exploration.

Exploring the relationship between baseline physical activity levels and mortality reduction associated with increases in physical activity: a modelling study

BACKGROUND

Increasing physical activity (PA) levels among the general adult population of developed nations is important for reducing premature mortality and the burdens of preventable illness. Assessing how effective PA interventions are as health interventions often involves categorising participants as either 'active' or 'sedentary' after the interventions. A model was developed showing that doing this could significantly misestimate the health effect of PA interventions.

METHODS

A life table model was constructed combining evidence on baseline PA levels with evidence indicating the non-linear relationship between PA levels and all-cause mortality risks. PA intervention scenarios were modelled which had the same mean increase in PA but different levels of take-up by people who were more active or more sedentary to begin with.

RESULTS

The model simulations indicated that, compared with a scenario where already-active people did most of the additional PA, a scenario where the least active did the most additional PA was around a third more effective in preventing deaths between the ages of 50 and 60 years. The relationship between distribution of PA take-up and health effect was explored systematically and appeared non-linear.

CONCLUSIONS

As the health gains of a given PA increase are greatest among people who are most sedentary, smaller increases in PA in the least active may have the same health benefits as much larger PA increases in the most active. To help such health effects to be assessed, PA studies should report changes in the distribution of PA level between the start and end of the study.

Does raising type 1 error rate improve power to detect interactions in linear regression models? A simulation study

Introduction

Statistical interactions are a common component of data analysis across a broad range of scientific disciplines. However, the statistical power to detect interactions is often undesirably low. One solution is to elevate the Type 1 error rate so that important interactions are not missed in a low power situation. To date, no study has quantified the effects of this practice on power in a linear regression model.

Methods

A Monte Carlo simulation study was performed. A continuous dependent variable was specified, along with three types of interactions: continuous variable by continuous variable; continuous by dichotomous; and dichotomous by dichotomous. For each of the three scenarios, the interaction effect sizes, sample sizes, and Type 1 error rate were varied, resulting in a total of 240 unique simulations.

Results

In general, power to detect the interaction effect was either so low or so high at α = 0.05 that raising the Type 1 error rate only served to increase the probability of including a spurious interaction in the model. A small number of scenarios were identified in which an elevated Type 1 error rate may be justified.

Conclusions

Routinely elevating Type 1 error rate when testing interaction effects is not an advisable practice. Researchers are best served by positing interaction effects a priori and accounting for them when conducting sample size calculations.

Use of generalised additive models to categorise continuous variables in clinical prediction

BACKGROUND

In medical practice many, essentially continuous, clinical parameters tend to be categorised by physicians for ease of decision-making. Indeed, categorisation is a common practice both in medical research and in the development of clinical prediction rules, particularly where the ensuing models are to be applied in daily clinical practice to support clinicians in the decision-making process. Since the number of categories into which a continuous predictor must be categorised depends partly on the relationship between the predictor and the outcome, the need for more than two categories must be borne in mind.

METHODS

We propose a categorisation methodology for clinical-prediction models, using Generalised Additive Models (GAMs) with P-spline smoothers to determine the relationship between the continuous predictor and the outcome. The proposed method consists of creating at least one average-risk category along with high- and low-risk categories based on the GAM smooth function. We applied this methodology to a prospective cohort of patients with exacerbated chronic obstructive pulmonary disease. The predictors selected were respiratory rate and partial pressure of carbon dioxide in the blood (PCO2), and the response variable was poor evolution. An additive logistic regression model was used to show the relationship between the covariates and the dichotomous response variable. The proposed categorisation was compared to the continuous predictor as the best option, using the AIC and AUC evaluation parameters. The sample was divided into a derivation (60%) and validation (40%) samples. The first was used to obtain the cut points while the second was used to validate the proposed methodology.

RESULTS

The three-category proposal for the respiratory rate was ≤ 20;(20,24];> 24, for which the following values were obtained: AIC=314.5 and AUC=0.638. The respective values for the continuous predictor were AIC=317.1 and AUC=0.634, with no statistically significant differences being found between the two AUCs (p =0.079). The four-category proposal for PCO2 was ≤ 43;(43,52];(52,65];> 65, for which the following values were obtained: AIC=258.1 and AUC=0.81. No statistically significant differences were found between the AUC of the four-category option and that of the continuous predictor, which yielded an AIC of 250.3 and an AUC of 0.825 (p =0.115).

CONCLUSIONS

Our proposed method provides clinicians with the number and location of cut points for categorising variables, and performs as successfully as the original continuous predictor when it comes to developing clinical prediction rules.

Decay of impact after self-management education for people with chronic illnesses: changes in anxiety and depression over one year

BACKGROUND

In people with chronic illnesses, self-management education can reduce anxiety and depression. Those benefits, however, decay over time. Efforts have been made to prevent or minimize that "decay of impact", but they have not been based on information about the decay's characteristics, and they have failed. Here we show how the decay's basic characteristics (prevalence, timing, and magnitude) can be quantified. Regarding anxiety and depression, we also report the prevalence, timing, and magnitude of the decay.

METHODS

Adults with various chronic conditions participated in a self-management educational program (n = 369). Data were collected with the Hospital Anxiety and Depression Scale four times over one year. Using within-person effect sizes, we defined decay of impact as a decline of ≥0.5 standard deviations after improvement by at least the same amount. We also interpret the results using previously-set criteria for non-cases, possible cases, and probable cases.

RESULTS

Prevalence: On anxiety, decay occurred in 19% of the participants (70/369), and on depression it occurred in 24% (90/369). Timing: In about one third of those with decay, it began 3 months after the baseline measurement (6 weeks after the educational program ended). Magnitude: The median magnitudes of decay on anxiety and on depression were both 4 points, which was about 1 standard deviation. Early in the follow-up year, many participants with decay moved into less severe clinical categories (e.g., becoming non-cases). Later, many of them moved into more severe categories (e.g., becoming probable cases).

CONCLUSIONS

Decay of impact can be identified and quantified from within-person effect sizes. This decay occurs in about one fifth or more of this program's participants. It can start soon after the program ends, and it is large enough to be clinically important. These findings can be used to plan interventions aimed at preventing or minimizing the decay of impact.

Correction of the significance level when attempting multiple transformations of an explanatory variable in generalized linear models

BACKGROUND

In statistical modeling, finding the most favorable coding for an exploratory quantitative variable involves many tests. This process involves multiple testing problems and requires the correction of the significance level.

METHODS

For each coding, a test on the nullity of the coefficient associated with the new coded variable is computed. The selected coding corresponds to that associated with the largest statistical test (or equivalently the smallest pvalue). In the context of the Generalized Linear Model, Liquet and Commenges (Stat Probability Lett,71:33-38,2005) proposed an asymptotic correction of the significance level. This procedure, based on the score test, has been developed for dichotomous and Box-Cox transformations. In this paper, we suggest the use of resampling methods to estimate the significance level for categorical transformations with more than two levels and, by definition those that involve more than one parameter in the model. The categorical transformation is a more flexible way to explore the unknown shape of the effect between an explanatory and a dependent variable.

RESULTS

The simulations we ran in this study showed good performances of the proposed methods. These methods were illustrated using the data from a study of the relationship between cholesterol and dementia.

CONCLUSION

The algorithms were implemented using R, and the associated CPMCGLM R package is available on the CRAN.

The use of continuous data versus binary data in MTC models: a case study in rheumatoid arthritis

Background

Estimates of relative efficacy between alternative treatments are crucial for decision making in health care. When sufficient head to head evidence is not available Bayesian mixed treatment comparison models provide a powerful methodology to obtain such estimates. While models can be fit to a broad range of efficacy measures, this paper illustrates the advantages of using continuous outcome measures compared to binary outcome measures.

Methods

Using a case study in rheumatoid arthritis a Bayesian mixed treatment comparison model is fit to estimate the relative efficacy of five anti-TNF agents currently licensed in Europe. The model is fit for the continuous HAQ improvement outcome measure and a binary version thereof as well as for the binary ACR response measure and the underlying continuous effect. Results are compared regarding their power to detect differences between treatments.

Results

Sixteen randomized controlled trials were included for the analysis. For both analyses, based on the HAQ improvement as well as based on the ACR response, differences between treatments detected by the binary outcome measures are subsets of the differences detected by the underlying continuous effects.

Conclusions

The information lost when transforming continuous data into a binary response measure translates into a loss of power to detect differences between treatments in mixed treatment comparison models. Binary outcome measures are therefore less sensitive to change than continuous measures. Furthermore the choice of cut-off point to construct the binary measure also impacts the relative efficacy estimates.

Height and health-related quality of life: a nationwide population study

CONTEXT

Treatment for short stature in childhood has been recommended with the purpose of improving adult health-related quality of life (HRQoL). However, there are only limited data available concerning the consequences of body height for HRQoL in adulthood.

OBJECTIVE

Our objective was to investigate the relationship between body height and HRQoL.

DESIGN, SETTING, AND PARTICIPANTS

This national representative, cross-sectional household survey of the French general noninstitutionalized population included 8857 men and 9248 women, aged 18-50 yr, in 2003.

MAIN OUTCOME MEASURES

Scores on the eight subscales of the Medical Outcomes Study 36-item Short Form (SF-36) were the primary outcomes. Univariate and multivariate linear regression analyses were used to evaluate the effect of height on HRQoL while controlling for age and various socioeconomic variables and pathological conditions.

RESULTS

Height was found to be a very weak predictor of HRQoL both for men and women. Only heights lower than 149.2 and 136.0 cm and higher than 203.6 and 188.7 cm, in men and women, respectively, were associated with a clinically significant reduction in physical functioning. The effects of body height on other (mental and social) dimensions of HRQoL were negligible or undetectable.

CONCLUSIONS

Height appears to have minimal consequences for physical functioning and negligible effects on other dimensions of HRQoL. These results contrast with widely popularized stereotypes and common beliefs and should be carefully considered to avoid further stigmatization and unnecessary medical care of individuals who are at the lower end of the growth distribution.