Suchergebnisse

ObjectiveTo change the common practice of eliminating independent variables from models because they produce multicollinearity in an independent variable of special interest.MethodsI supplement my presentation, which is based on the purposes of regression analysis, by using Venn diagrams, simple formulas, and two small simulations.ResultsIndependent variables that when removed from a model substantially change the statistics associated with the independent variable(s) of most interest are variables that should typically be kept in the model. Multicollinearity is not a sufficient reason to drop variables from a model.ConclusionI argue against the routine dropping of variables that cause multicollinearity in an independent variable of interest from regression models. A more important criterion to consider when contemplating dropping a variable from a model is "model influence."

If a researcher wants to estimate the individual age, period, and cohort coefficients in an age-period-cohort (APC) model, the method of choice is constrained regression, which includes the intrinsic estimator (IE) recently introduced by Yang and colleagues. To better understand these constrained models, the author shows algebraically how each constraint is associated with a specific generalized inverse that is associated with a particular solution vector that (when the model is just identified under the constraint) produces the least square solution to the APC model. The author then discusses the geometry of constrained estimators in terms of solutions being orthogonal to constraints, solutions to various constraints all lying on a line single line in multidimensional space, the distance on that line between various solutions, and the crucial role of the null vector. This provides insight into what characteristics all constrained estimators share and what is unique about the IE. The first part of the article focuses on constrained estimators in general (including the IE), and the latter part compares and contrasts the properties of traditionally constrained APC estimators and the IE. The author concludes with some cautions and suggestions for researchers using and interpreting constrained estimators.

According to the Uniform Crime Reports, violent crime rates increased dramatically over the past two decades. National Crime Victimization Survey data, on the other hand, indicate that the rates of violent crime remained relatively stable or dropped during this period. Which series provides a "correct" estimate of crime‐rate trends is of more than academic interest. Highly publicized statistics on crime trends influence the public's concerns about crime and the decisions of policymakers both directly through their own perceptions of crime trends and indirectly through demands by the general public to control crime. This article compares these two major series on trends in violent crime rates in the United States for the period 1973–1992, with the goal of assessing the extent to which they measure the same underlying phenomenon: fluctuations in violent crime rates. The series are related (but not strongly). My conclusion, with some reservation, is that changes in law enforcement agencies rather than changes in the rates of violent crime incidents have created the upward trend in UCR violent crime rates during the past two decades.

Social scientists often use the mean of individual-level characteristics to describe aggregates such as organizations, schools, or programs. They seldom, however, attempt to assess the reliability of these measures. This article indicates how the internal consistency reliability of such measures can be estimated for several commonly used research designs. The method is based on generalizability theory and requires only the data that would normally be used to answer the substantive questions of a study; for example retests or parallel forms are not necessary.

The compositional effects of relatively large young‐adult cohorts on the total rate of serious crimes is well established. The more subtle effect of relative cohort size on age‐specific crime rates, suggested by Richard Easterlin, is more controversial. The literature contains no adequate test of Easterlin's hypothesis as it relates to crime. To provide an adequate test of Easterlin's theory, this study includes age‐specific rates and measures of relative cohort size and controls for age and period in an age‐period‐relative‐cohort‐size model. Using arrest data from the Uniform Crime Reports (Part I crimes) for the years 1960, 1965, 1970, 1975, 1980, and 1985, the analysis provides support for Easterlin's theory for property crimes, that is, for robbery, burglary, and larceny (but not for motor vehicle theft). Though these relationships were small in comparison to those between age or period and age‐specific crime rates, they were generally statistically significant and were replicated with data from 1962, 1967, 1972, 1977, 1982, and 1987. The relationships between relative cohort size and assaultive crimes provided little consistent support for Easterlin's theory.

Data on criminal homicides (from the Uniform Crime Reports) and aggravated assaults and simple assaults (from the National Crime Surveys) are analyzed to determine the extent to which violent crimes occur within or between sexes. The routine activities approach is used to develop hypotheses, and those hypotheses are tested using models that estimate the proportion of ingroup and outgroup crimes "expected." With the exception of homicides, in which women murder men more often than expected, each of these violent crimes occurs within sexes more often than expected. There is a strong relationship between the type of violence (simple assault, aggravated assault, and homicide) and the extent to which the target of female aggression is a male.

Social scientists often measure variables that are best thought of as having continuous underlying distributions by using crude rank-category techniques (ie, measures that have few rank categories). The result is a large amount of grouping error. Computer simulations are used to investigate the effects of four factors that contribute to grouping error: (1) the underlying distribution of scores; (2) the marginal proportion of cases in each category; (3) the scoring systems used to score categories; & (4) the number of categories used. A distinction is made between two types of grouping error -- transformation & categorization. This distinction is used to clarify two distinct strategies that can be used to reduce grouping error. 4 Tables, 11 References. HA.

It has been suggested that researchers comparing causal models across populations employ path regression coefficients corrected for unreliability. This procedure, however, does not eliminate the problem of differences in the units of measure across populations. This paper outlines a technique designed to eliminate this problem by comparing "effect ratios," rather than effect coefficients, across populations. Both hypothetical and real data are used to demonstrate the technique.