Bianca De Stavola

Issues in the reporting of epidemiological studies: a survey of recent practice

Abstract Objectives To review current practice in the analysis and reporting of epidemiological research and to identify limitations. Design Examination of articles published in January 2001 that investigated associations between risk factors/exposure variables and disease events/measures in individuals. Setting Eligible English language journals including all major epidemiological journals, all major general medical journals, and the two leading journals in cardiovascular disease and cancer. Main outcome measure Each article was evaluated with a standard proforma. Results We found 73 articles in observational epidemiology; most were either cohort or case-control studies. Most studies looked at cancer and cardiovascular disease, even after we excluded specialty journals. Quantitative exposure variables predominated, which were mostly analysed as ordered categories but with little consistency or explanation regarding choice of categories. Sample selection, participant refusal, and data quality received insufficient attention in many articles. Statistical analyses commonly used odds ratios (38 articles) and hazard/rate ratios (23), with some inconsistent use of terminology. Confidence intervals were reported in most studies (68), though use of P values was less common (38). Few articles explained their choice of confounding variables; many performed subgroup analyses claiming an effect modifier, though interaction tests were rare. Several investigated multiple associations between exposure and outcome, increasing the likelihood of false positive claims. There was evidence of publication bias. Conclusions This survey raises concerns regarding inadequacies in the analysis and reporting of epidemiological publications in mainstream journals.

Quantification of the completeness of follow-up

Completeness of follow-up is important, especially in clinical trials, since unequal follow-up in the treatment groups can bias the analysis of results. In survival studies, information on participants who do not complete the study is often omitted because their data can be included up to the time at which they were lost to follow-up. We propose a simple measure of completeness that is the ratio of the total observed person-time and the potential person-time of follow-up in a study. Our measure is easy to calculate, can be illustrated pictorially, and can be used to identify subgroups with especially poor follow-up.

Predicting prognosis in stable angina—results from the Euro heart survey of stable angina: prospective observational study

Abstract Objectives To investigate the prognosis associated with stable angina in a contemporary population as seen in clinical practice, to identify the key prognostic features, and from this to construct a simple score to assist risk prediction. Design Prospective observational cohort study. Setting Pan-European survey in 156 outpatient cardiology clinics. Participants 3031 patients were included on the basis of a new clinical diagnosis by a cardiologist of stable angina with follow-up at one year. Main outcome measure Death or non-fatal myocardial infarction. Results The rate of death and non-fatal myocardial infarction in the first year was 2.3 per 100 patient years; the rate was 3.9 per 100 patient years in the subgroup (n = 994) with angiographic confirmation of coronary disease. The clinical and investigative factors most predictive of adverse outcome were comorbidity, diabetes, shorter duration of symptoms, increasing severity of symptoms, abnormal ventricular function, resting electrocardiogaphic changes, or not having any stress test done. Results of non-invasive stress tests did not significantly predict outcome in the population who had tests done. A score was constructed using the parameters predictive of outcome to estimate the probability of death or myocardial infarction within one year of presentation with stable angina. Conclusions A score based on the presence of simple, objective clinical and investigative variables makes it possible to discriminate effectively between very low risk and very high risk patients and to estimate the probability of death or non-fatal myocardial infarction over one year.

A structured approach to modelling the effects of binary exposure variables over the life course

Background There is growing interest in the relationship between time spent in adverse circumstances across life course and increased risk of chronic disease and early mortality. This accumulation hypothesis is usually tested by summing indicators of binary variables across the life span to form an overall score that is then used as the exposure in regression models for health outcomes. This article highlights potential issues in the interpretation of results obtained from such an approach. Methods We propose a model-building framework that can be used to formally compare alternative hypotheses on the effect of multiple binary exposure measurements collected across the life course. The saturated model where the order and value of the binary variable at each time point influence the outcome of interest is compared with nested alternative specifications corresponding to the critical period, cumulative risk or hypotheses about the effect of changes in environment. This framework is illustrated with data on adult body mass index and socioeconomic position measured once in childhood and twice in adulthood from the Medical Research Council National Survey of Health and Development, using a series of liner regression models. Results We demonstrate how analyses that only consider the association of a cumulative score with a later outcome may produce misleading results. Conclusion We recommend comparing a set of nested models—each corresponding to the accumulation, critical period and effect modification hypotheses—to an all-inclusive (saturated) model. This approach can provide a formal and clearer understanding of the relative merits of these alternative hypotheses.

Mammographic Features and Subsequent Risk of Breast Cancer: A Comparison of Qualitative and Quantitative Evaluations in the Guernsey Prospective Studies

Mammographic features are known to be associated with breast cancer but the magnitude of the effect differs markedly from study to study. Methods to assess mammographic features range from subjective qualitative classifications to computer-automated quantitative measures. We used data from the UK Guernsey prospective studies to examine the relative value of these methods in predicting breast cancer risk. In all, 3,211 women ages ≥35 years who had a mammogram taken in 1986 to 1989 were followed-up to the end of October 2003, with 111 developing breast cancer during this period. Mammograms were classified using the subjective qualitative Wolfe classification and several quantitative mammographic features measured using computer-based techniques. Breast cancer risk was positively associated with high-grade Wolfe classification, percent breast density and area of dense tissue, and negatively associated with area of lucent tissue, fractal dimension, and lacunarity. Inclusion of the quantitative measures in the same model identified area of dense tissue and lacunarity as the best predictors of breast cancer, with risk increasing by 59% [95% confidence interval (95% CI), 29-94%] per SD increase in total area of dense tissue but declining by 39% (95% CI, 53-22%) per SD increase in lacunarity, after adjusting for each other and for other confounders. Comparison of models that included both the qualitative Wolfe classification and these two quantitative measures to models that included either the qualitative or the two quantitative variables showed that they all made significant contributions to prediction of breast cancer risk. These findings indicate that breast cancer risk is affected not only by the amount of mammographic density but also by the degree of heterogeneity of the parenchymal pattern and, presumably, by other features captured by the Wolfe classification.

Birth Size and Breast Cancer Risk: Re-analysis of Individual Participant Data from 32 Studies

BACKGROUND Birth size, perhaps a proxy for prenatal environment, might be a correlate of subsequent breast cancer risk, but findings from epidemiological studies have been inconsistent. We re-analysed individual participant data from published and unpublished studies to obtain more precise estimates of the magnitude and shape of the birth size-breast cancer association. METHODS AND FINDINGS Studies were identified through computer-assisted and manual searches, and personal communication with investigators. Individual participant data from 32 studies, comprising 22,058 breast cancer cases, were obtained. Random effect models were used, if appropriate, to combine study-specific estimates of effect. Birth weight was positively associated with breast cancer risk in studies based on birth records (pooled relative risk [RR] per one standard deviation [SD] [= 0.5 kg] increment in birth weight: 1.06; 95% confidence interval [CI] 1.02-1.09) and parental recall when the participants were children (1.02; 95% CI 0.99-1.05), but not in those based on adult self-reports, or maternal recall during the womans adulthood (0.98; 95% CI 0.95-1.01) (p for heterogeneity between data sources = 0.003). Relative to women who weighed 3.000-3.499 kg, the risk was 0.96 (CI 0.80-1.16) in those who weighed < 2.500 kg, and 1.12 (95% CI 1.00-1.25) in those who weighed > or = 4.000 kg (p for linear trend = 0.001) in birth record data. Birth length and head circumference from birth records were also positively associated with breast cancer risk (pooled RR per one SD increment: 1.06 [95% CI 1.03-1.10] and 1.09 [95% CI 1.03-1.15], respectively). Simultaneous adjustment for these three birth size variables showed that length was the strongest independent predictor of risk. The birth size effects did not appear to be confounded or mediated by established breast cancer risk factors and were not modified by age or menopausal status. The cumulative incidence of breast cancer per 100 women by age 80 y in the study populations was estimated to be 10.0, 10.0, 10.4, and 11.5 in those who were, respectively, in the bottom, second, third, and top fourths of the birth length distribution. CONCLUSIONS This pooled analysis of individual participant data is consistent with birth size, and in particular birth length, being an independent correlate of breast cancer risk in adulthood.

Updating prognosis in primary biliary cirrhosis using a time-dependent Cox regression model

BACKGROUND The precision of current prognostic models in primary biliary cirrhosis (PBC) is rather low, partly because they are based on data from just one time during the course of the disease. The aim of this study was to design a new, more precise prognostic model by incorporating follow-up data in the development of the model. METHODS We have performed Cox regression analyses with time-dependent variables in 237 PBC patients followed up regularly for up to 11 years. The validity of the obtained models was tested by comparing predicted and observed survival in 147 independent PBC patients followed for up to 6 years. RESULTS In the obtained model the following time-dependent variables independently indicated a poor prognosis: high bilirubin, low albumin, ascites, gastrointestinal bleeding, and old age. When including histological variables, cirrhosis, central cholestasis, and low immunoglobulin (Ig)M also indicated a poor prognosis. The survival predicted by the models agreed well with the survival observed in the independent PBC patients. The time-dependent models predicted better than our previously published time-fixed model. CONCLUSIONS Using the time-dependent Cox models, one can estimate a more precise probability of surviving the next 1, 3, or 6 months for any given patient at any time during the course of the disease. This may improve monitoring of PBC patients.

The cognitive cost of being a twin: evidence from comparisons within families in the Aberdeen children of the 1950s cohort study

Abstract Objectives To determine whether twins have lower IQ scores in childhood than singletons in the same family and, if so, whether differences in fetal growth explain this deficit. Design Cohort study. Setting Scotland. Participants 9832 singletons and 236 twins born in Aberdeen between 1950 and 1956. Results At age 7, the mean IQ score of twins was 5.3 points lower (95% confidence interval 1.5 to 9.1) and at age 9, 6.0 points lower (1.7 to 10.2) than that of singletons in the same family. Adjustment for sex, mothers age, and number of older siblings had little effect on these differences. Further adjustment for birth weight and gestational age attenuated the IQ difference between twins and singletons: the difference in mean IQ was 2.6 points (−1.5 to 6.7) at age 7 and 4.1 points (-0.5 to 8.8) at age 9. Conclusions Twins have substantially lower IQ in childhood than singletons in the same family. This effect cannot be explained by confounding due to socioeconomic, maternal, or other family characteristics, or by recruitment bias. The reduced prenatal growth and shorter gestations of twins may explain an important part of their lower IQ in childhood.

Using causal diagrams to guide analysis in missing data problems.

Estimating causal effects from incomplete data requires additional and inherently untestable assumptions regarding the mechanism giving rise to the missing data. We show that using causal diagrams to represent these additional assumptions both complements and clarifies some of the central issues in missing data theory, such as Rubins classification of missingness mechanisms (as missing completely at random (MCAR), missing at random (MAR) or missing not at random (MNAR)) and the circumstances in which causal effects can be estimated without bias by analysing only the subjects with complete data. In doing so, we formally extend the back-door criterion of Pearl and others for use in incomplete data examples. These ideas are illustrated with an example drawn from an occupational cohort study of the effect of cosmic radiation on skin cancer incidence.

The association of height, weight, menstrual and reproductive events with breast cancer: results from two prospective studies on the island of Guernsey (United Kingdom)

The association with breast cancer of menstrual and reproductive events, family history of breast cancer, and body size have been studied on two cohorts of 6,706 volunteers on the island of Guernsey (United Kingdom), 168 of whom had breast cancer detected during follow-up. The median follow-up time of the non-cases was 21 years in the first study and 10 years in the second. A time-dependent Cox regression model was fitted to the data with age as the time-dependent variable in order to represent the effect of changing menopausal status. Other variables examined in the model were age at menarche, parity, age at first birth, family history of breast cancer, height, weight (both directly measured), relative weight (weight [kg]/height[m]), and Quetelets body mass index (weight[kg]/height[m]2). Interactions between age and all other covariates also were examined. Family history was found to be the most important risk factor for women aged less than 51 years (relative risk [RR]=3.5, 95 percent confidence interval [CI]=2.0–6.0), and intervals between menarche and first birth longer than 14 years were found to increase significantly the risk of breast cancer in women older than 61 years (RR=2.4, CI=1.3–4.4). Height was the only indicator of body size which was associated significantly with risk of breast cancer, the estimated regression coefficient indicating an increase in risk of about 70 percent for women on the 90th centile of height relative to those on the 10th centile. A survey of the literature showed that the association between risk of breast cancer and height was found in those studies which used direct measurements of height but not in others which used self-reported values.