Prognostic modeling with logistic regression analysis: in search of a sensible strategy in small data sets.

Clinical decision making often requires estimates of the likelihood of a dichotomous outcome in individual patients. When empirical data are available, these estimates may well be obtained from a logistic regression model. Several strategies may be followed in the development of such a model. In this study, the authors compare alternative strategies in 23 small subsamples from a large data set of patients with an acute myocardial infarction, where they developed predictive models for 30-day mortality. Evaluations were performed in an independent part of the data set. Specifically, the authors studied the effect of coding of covariables and stepwise selection on discriminative ability of the resulting model, and the effect of statistical "shrinkage" techniques on calibration. As expected, dichotomization of continuous covariables implied a loss of information. Remarkably, stepwise selection resulted in less discriminating models compared to full models including all available covariables, even when more than half of these were randomly associated with the outcome. Using qualitative information on the sign of the effect of predictors slightly improved the predictive ability. Calibration improved when shrinkage was applied on the standard maximum likelihood estimates of the regression coefficients. In conclusion, a sensible strategy in small data sets is to apply shrinkage methods in full models that include well-coded predictors that are selected based on external information.     

Modeling geographic and demographic variability in residential concentrations of environmental tobacco smoke using national data sets

Despite substantial attention toward environmental tobacco smoke (ETS) exposure, previous studies have not provided adequate information to apply broadly within community-scale risk assessments. We aim to estimate residential concentrations of particulate matter (PM) from ETS in sociodemographic and geographic subpopulations in the United States for the purpose of screening-level risk assessment. We developed regression models to characterize smoking using the 2006-7 Current Population Survey--Tobacco Use Supplement, and linked these with air exchange models using the 2007 American Housing Survey. Using repeated logistic and log-linear models (n = 1000), we investigated whether household variables from the 2000 United States census can predict exposure likelihood and ETS-PM concentration in exposed households. We estimated a mean ETS-PM concentration of 16 μg/m(3) among the 17% of homes with non-zero exposure (3 μg/m(3) overall), with substantial variability among homes. The highest exposure likelihood was in the South and Midwest regions, rural populations, and low-income households. Concentrations in exposed households were highest in the South and demonstrated a non-monotonic association with income, related to air exchange rate patterns. We provide estimates of ETS-PM concentration distributions for different subpopulations in the United States, providing a starting point for communities interested in characterizing aggregate and cumulative risks from indoor pollutants.     

Standard national perinatal data: a suggested minimum data set

Summary Recent interest and concern over perinatal practice and pregnancy outcome has highlighted the need for the collection of standard national perinatal data over and above that already provided by the civil registration system. A method is described for the derivation of a minimum data set based on a review of 38 perinatal enquiries and routine data collection systems. Only clearly defined parameters are included and it is suggested that other items should be added as and when definitional difficulties have been resolved by the relevant professional groups.     

Presentation and interpretation of food intake data: Factors affecting comparability across studies

Non-uniform, unclear, or incomplete presentation of food intake data limits interpretation, usefulness, and comparisons across studies. In this contribution, we discuss factors affecting uniform reporting of food intake across studies. The amount of food eaten can be reported as mean portion size, number of servings or total amount of food consumed per day; the absolute intake value for the specific study depends on the denominator used because food intake data can be presented as per capita intake or for consumers only. To identify the foods mostly consumed, foods are reported and ranked according to total number of times consumed, number of consumers, total intake, or nutrient contribution by individual foods or food groups. Presentation of food intake data primarily depends on a study's aim; reported data thus often are not comparable across studies. Food intake data further depend on the dietary assessment methodology used and foods in the database consulted; and are influenced by the inherent limitations of all dietary assessments. Intake data can be presented as either single foods or as clearly defined food groups. Mixed dishes, reported as such or in terms of ingredients and items added during food preparation remain challenging. Comparable presentation of food consumption data is not always possible; presenting sufficient information will assist valid interpretation and optimal use of the presented data. A checklist was developed to strengthen the reporting of food intake data in science communication.     

The use of comparability ratios to adjust hospital trend data

Hospital trend data may be affected by changes in diagnostic coding schemes. We studied the change from ICDA-8 (I-8) to ICD-9-CM (I-9) in a sample of roughly 13,600 double-coded Veterans Administration hospital diagnoses. Comparability ratios were computed and used to adjust trend data which overlap the time period when the shift from I-8 to I-9 occurred. With this adjustment for change in diagnostic coding scheme, apparent diagnostic trends are substantially altered.     

The international comparability of cancer mortality data. Results of an international death certificate study

In preparation for the 10th revision of the International Classification of Diseases (ICD-10), a two-part study was undertaken to assess the international comparability of the coding, by the 9th revision (ICD-9), of death certificates mentioning cancer, to see whether there had been improvement since the 8th revision (ICD-8). Part I repeated a 1978 study in which nine countries coded the same 1,234 United States death certificates mentioning cancer by ICD-9. The proportion of disagreements in coding the underlying cause of death fell about 35% between 1978 and the present study. This reduction was probably due to the new more detailed rules for coding cancer death certificates given in ICD-9. To combat the criticism of the possible bias associated with using United States death certificates only, in Part II of the study, each of seven countries submitted about 100 certificates translated into English which had posed problems in coding cancer. Discrepancies in assigning the underlying cause of death were found for 54% of these problem certificates. The major types of problems identified were coding when multiple cancer sites were mentioned on the death certificate, whether to select heart disease or cancer as the underlying cause of death, and the interpretation of the coding rules. Better rules for ICD-10 must be provided for both physicians and coders if international comparability of cancer mortality data is to be achieved.     

Minimum data sets in ambulatory care]

A sensible and rational approach is necessary in order to match the growing requirements of information on the ambulatory care sector characterised by a big deal of complexity and variability. This must be applied to any initiative attempting to obtain any degree of harmonization in ambulatory care data sets. This paper summarises a conceptual aspect of the work undertaken by the project Measurement Characterization and Control of Ambulatory Care in Europe (McACE) sponsored by the Commission of the European Communities under the exploratory phase of the Advanced Informatics in Medicine (AIM) programme. Our approach places in the first priority the efficiency when assessing the information needs of the different actors trying to make a balanced approach between the costs of collecting and processing a specific item and its use. For this we performed a conceptual validation of the term Minimum Basic Data Set and its potential application to the Ambulatory Care Sector. It is suggested its substitution by the term Minimum Data Set as a part of a practical tool for promoting the harmonization of the information in ambulatory care.     

Stepwise selection in small data sets: a simulation study of bias in logistic regression analysis.

Stepwise selection methods are widely applied to identify covariables for inclusion in regression models. One of the problems of stepwise selection is biased estimation of the regression coefficients. We illustrate this "selection bias" with logistic regression in the GUSTO-I trial (40,830 patients with an acute myocardial infarction). Random samples were drawn that included 3, 5, 10, 20, or 40 events per variable (EPV). Backward stepwise selection was applied in models containing 8 or 16 pre-specified predictors of 30-day mortality. We found a considerable overestimation of regression coefficients of selected covariables. The selection bias decreased with increasing EPV. For EPV 3, 10, or 40, the bias exceeded 25% for 7, 3, and 1 in the 8-predictor model respectively, when a conventional selection criterion was used (alpha = 0.05). For these EPV values, the bias was less than 20% for all covariables when no selection was applied. We conclude that stepwise selection may result in a substantial bias of estimated regression coefficients.