Can test statistics in covariance structure analysis be trusted?

Covariance structure analysis uses chi 2 goodness-of-fit test statistics whose adequacy is not known. Scientific conclusions based on models may be distorted when researchers violate sample size, variate independence, and distributional assumptions. The behavior of 6 test statistics is evaluated with a Monte Carlo confirmatory factor analysis study. The tests performed dramatically differently under 7 distributional conditions at 6 sample sizes. Two normal-theory tests worked well under some conditions but completely broke down under other conditions. A test that permits homogeneous nonzero kurtoses performed variably. A test that permits heterogeneous marginal kurtoses performed better. A distribution-free test performed spectacularly badly in all conditions at all but the largest sample sizes. The Satorra-Bentler scaled test statistic performed best overall.     

Using Lancaster's mid-P correction to the Fisher's exact test for adverse impact analyses

Adverse impact is often assessed by evaluating whether the success rates for 2 groups on a selection procedure are significantly different. Although various statistical methods have been used to analyze adverse impact data, Fisher's exact test (FET) has been widely adopted, especially when sample sizes are small. In recent years, however, the statistical field has expressed concern regarding the default use of the FET and has proposed several alternative tests. This article reviews Lancaster's mid-P (LMP) test (Lancaster, 1961), an adjustment to the FET that tends to have increased power while maintaining a Type I error rate close to the nominal level. On the basis of Monte Carlo simulation results, the LMP test was found to outperform the FET across a wide range of conditions typical of adverse impact analyses. The LMP test was also found to provide better control over Type I errors than the large-sample Z-test when sample size was very small, but it tended to have slightly lower power than the Z-test under some conditions.     

Detecting a lack of association: An equivalence testing approach

Researchers often test for a lack of association between variables. A lack of association is usually established by demonstrating a non‐significant relationship with a traditional test (e.g., Pearson's r). However, for logical as well as statistical reasons, such conclusions are problematic. In this paper, we discuss and compare the empirical Type I error and power rates of three lack of association tests. The results indicate that large, sometimes very large, sample sizes are required for the test statistics to be appropriate. What is especially problematic is that the required sample sizes may exceed what is practically feasible for the conditions that are expected to be common among researchers in psychology. This paper highlights the importance of using available lack of association tests, instead of traditional tests of association, for demonstrating the independence of variables, and qualifies the conditions under which these tests are appropriate.     

A Monte Carlo evaluation of tests for comparing dependent correlations

The authors conducted a Monte Carlo simulation of 8 statistical tests for comparing dependent zero-order correlations. In particular, they evaluated the Type I error rates and power of a number of test statistics for sample sizes (Ns) of 20, 50, 100, and 300 under 3 different population distributions (normal, uniform, and exponential). For the Type I error rate analyses, the authors evaluated 3 different magnitudes of the predictor-criterion correlations (rho(y,x1) = rho(y,x2) = .1, .4, and .7). For the power analyses, they examined 3 different effect sizes or magnitudes of discrepancy between rho(y,x1) and rho(y,x2) (values of .1, .3, and .6). They conducted all of the simulations at 3 different levels of predictor intercorrelation (rho(x1,x2) = .1, .3, and .6). The results indicated that both Type I error rate and power depend not only on sample size and population distribution, but also on (a) the predictor intercorrelation and (b) the effect size (for power) or the magnitude of the predictor-criterion correlations (for Type I error rate). When the authors considered Type I error rate and power simultaneously, the findings suggested that O. J. Dunn and V. A. Clark's (1969) z and E. J. Williams's (1959) t have the best overall statistical properties. The findings extend and refine previous simulation research and as such, should have greater utility for applied researchers.     

Testing two variances for superiority/non-inferiority and equivalence: Using the exhaustion algorithm for sample size allocation with cost

The equality of two group variances is frequently tested in experiments. However, criticisms of null hypothesis statistical testing on means have recently arisen and there is interest in other types of statistical tests of hypotheses, such as superiority/non-inferiority and equivalence. Although these tests have become more common in psychology and social sciences, the corresponding sample size estimation for these tests is rarely discussed, especially when the sampling unit costs are unequal or group sizes are unequal for two groups. Thus, for finding optimal sample size, the present study derived an initial allocation by approximating the percentiles of an F distribution with the percentiles of the standard normal distribution and used the exhaustion algorithm to select the best combination of group sizes, thereby ensuring the resulting power reaches the designated level and is maximal with a minimal total cost. In this manner, optimization of sample size planning is achieved. The proposed sample size determination has a wide range of applications and is efficient in terms of Type I errors and statistical power in simulations. Finally, an illustrative example from a report by the Health Survey for England, 1995–1997, is presented using hypertension data. For ease of application, four R Shiny apps are provided and benchmarks for setting equivalence margins are suggested.     

GPOWER: A general power analysis program

GPOWER is a completely interactive, menu-driven program for IBM-compatible and Apple Macintosh personal computers. It performs high-precision statistical power analyses for the most common statistical tests in behavioral research, that is,t tests,F tests, andχ ² tests. GPOWER computes (1) power values for given sample sizes, effect sizes andα levels (post hoc power analyses); (2) sample sizes for given effect sizes,α levels, and power values (a priori power analyses); and (3)α andβ values for given sample sizes, effect sizes, andβ/α ratios (compromise power analyses). The program may be used to display graphically the relation between any two of the relevant variables, and it offers the opportunity to compute the effect size measures from basic parameters defining the alternative hypothesis. This article delineates reasons for the development of GPOWER and describes the program’s capabilities and handling.     

Hypothesis tests for population heterogeneity in meta‐analysis

Choice of the appropriate model in meta‐analysis is often treated as an empirical question which is answered by examining the amount of variability in the effect sizes. When all of the observed variability in the effect sizes can be accounted for based on sampling error alone, a set of effect sizes is said to be homogeneous and a fixed‐effects model is typically adopted. Whether a set of effect sizes is homogeneous or not is usually tested with the so‐called Q test. In this paper, a variety of alternative homogeneity tests – the likelihood ratio, Wald and score tests – are compared with the Q test in terms of their Type I error rate and power for four different effect size measures. Monte Carlo simulations show that the Q test kept the tightest control of the Type I error rate, although the results emphasize the importance of large sample sizes within the set of studies. The results also suggest under what conditions the power of the tests can be considered adequate.     

Heterogeneous heterogeneity by default: Testing categorical moderators in mixed-effects meta-analysis

Categorical moderators are often included in mixed-effects meta-analysis to explain heterogeneity in effect sizes. An assumption in tests of categorical moderator effects is that of a constant between-study variance across all levels of the moderator. Although it rarely receives serious thought, there can be statistical ramifications to upholding this assumption. We propose that researchers should instead default to assuming unequal between-study variances when analysing categorical moderators. To achieve this, we suggest using a mixed-effects location-scale model (MELSM) to allow group-specific estimates for the between-study variance. In two extensive simulation studies, we show that in terms of Type I error and statistical power, little is lost by using the MELSM for moderator tests, but there can be serious costs when an equal variance mixed-effects model (MEM) is used. Most notably, in scenarios with balanced sample sizes or equal between-study variance, the Type I error and power rates are nearly identical between the MEM and the MELSM. On the other hand, with imbalanced sample sizes and unequal variances, the Type I error rate under the MEM can be grossly inflated or overly conservative, whereas the MELSM does comparatively well in controlling the Type I error across the majority of cases. A notable exception where the MELSM did not clearly outperform the MEM was in the case of few studies (e.g., 5). With respect to power, the MELSM had similar or higher power than the MEM in conditions where the latter produced non-inflated Type 1 error rates. Together, our results support the idea that assuming unequal between-study variances is preferred as a default strategy when testing categorical moderators.     

Power & Effect: A statistical utility for Macintosh and Windows systems

Significance tests are not the only step in statistics. Other considerations include effect sizes and adequate sample sizes for a respectable level of statistical power. However, many statistical packages are spotty in their offerings of effect size, complex, and lack a friendly interface. Textbooks may have limited coverage, and calculations entail several formulas and tables. Power & Effect offers a calculator- and formula-based metaphor to compute popular measures of effect size, simple significance tests between effect sizes, combining of effect sizes, simple significance tests based on known statistical values, and sample size determinations based on predicted results or effect size.     

Methodological issues in short-term motor memory research

Four methodological issues identified and studied by short-term motor memory researchers were examined in this paper: (a) the distributions of three commonly used measures of error (AE, CE, and VE); (b) statistical analysis of these scores as dependent measures in an experiment; (c) within-cell and experiment-wise estimates of the reliability of CE; and (d) the power of the statistical test with CE as the dependent measure. The first two sub-problems were explored theoretically, while the latter two were subjected to empirical examination using four data sets from previously published experiments. The distributions of AE and VE scores were described as non-normal under certain conditions, while the CE score distribution was shown to be normal. Therefore analyses of these three dependent measures may require different statistical procedures. In light of this information, the assumptions affecting estimates of reliability and power in an experimental framework were examined. Based on an analysis of the four data sets, a strategy was proposed in which a relatively small number of subjects could be utilized in an experiment without sacrificing desired experimental controls.     

Sample size determination for at test given at value from a previous study: A FORTRAN 77 program

When uncertain about the magnitude of an effect, researchers commonly substitute in the standard sample-size-determination formula an estimate of effect size derived from a previous experiment. A problem with this approach is that the traditional sample-size-determination formula was not designed to deal with the uncertainty inherent in an effect-size estimate. Consequently, estimate-substitution in the traditional sample-size-determination formula can lead to a substantial loss of power. A method of sample-size determination designed to handle uncertainty in effect-size estimates is described. The procedure uses thet value and sample size from a previous study, which might be a pilot study or a related study in the same area, to establish a distribution of probable effect sizes. The sample size to be employed in the new study is that which supplies an expected power of the desired amount over the distribution of probable effect sizes. A FORTRAN 77 program is presented that permits swift calculation of sample size for a variety oft tests, including independentt tests, relatedt tests,t tests of correlation coefficients, andt tests of multiple regressionb coefficients.     

Testing hypotheses involving Cronbach's alpha using marginal models

We discuss the statistical testing of three relevant hypotheses involving Cronbach's alpha: one where alpha equals a particular criterion; a second testing the equality of two alpha coefficients for independent samples; and a third testing the equality of two alpha coefficients for dependent samples. For each of these hypotheses, various statistical tests have been proposed. Over the years, these tests have depended on progressively fewer assumptions. We propose a new approach to testing the three hypotheses that relies on even fewer assumptions, is especially suited for discrete item scores, and can be applied easily to tests containing large numbers of items. The new approach uses marginal modelling. We compared the Type I error rate and the power of the marginal modelling approach to several of the available tests in a simulation study using realistic conditions. We found that the marginal modelling approach had the most accurate Type I error rates, whereas the power was similar across the statistical tests.     

Revisiting dispersion in count data item response theory models: The Conway–Maxwell–Poisson counts model

Count data naturally arise in several areas of cognitive ability testing, such as processing speed, memory, verbal fluency, and divergent thinking. Contemporary count data item response theory models, however, are not flexible enough, especially to account for over- and underdispersion at the same time. For example, the Rasch Poisson counts model (RPCM) assumes equidispersion (conditional mean and variance coincide) which is often violated in empirical data. This work introduces the Conway–Maxwell–Poisson counts model (CMPCM) that can handle underdispersion (variance lower than the mean), equidispersion, and overdispersion (variance larger than the mean) in general and specifically at the item level. A simulation study revealed satisfactory parameter recovery at moderate sample sizes and mostly unbiased standard errors for the proposed estimation approach. In addition, plausible empirical reliability estimates resulted, while those based on the RPCM were biased downwards (underdispersion) and biased upwards (overdispersion) when the simulation model deviated from equidispersion. Finally, verbal fluency data were analysed and the CMPCM with item-specific dispersion parameters fitted the data best. Dispersion parameter estimates indicated underdispersion for three out of four items. Overall, these findings indicate the feasibility and importance of the suggested flexible count data modelling approach.     

The power of statistical tests in meta-analysis

Calculations of the power of statistical tests are important in planning research studies (including meta-analyses) and in interpreting situations in which a result has not proven to be statistically significant. The authors describe procedures to compute statistical power of fixed- and random-effects tests of the mean effect size, tests for heterogeneity (or variation) of effect size parameters across studies, and tests for contrasts among effect sizes of different studies. Examples are given using 2 published meta-analyses. The examples illustrate that statistical power is not always high in meta-analysis.     

Comparison of a two-stage and three-stage interim-analysis procedure.

A statistical model for combining p values from multiple tests of significance is used to define rejection and acceptance regions for two-stage and three-stage sampling plans. Type I error rates, power, frequencies of early termination decisions, and expected sample sizes are compared. Both the two-stage and three-stage procedures provide appropriate protection against Type I errors. The two-stage sampling plan with its single interim analysis entails minimal loss in power and provides substantial reduction in expected sample size as compared with a conventional single end-of-study test of significance for which power is in the adequate range. The three-stage sampling plan with its two interim analyses introduces somewhat greater reduction in power, but it compensates with greater reduction in expected sample size. Either interim-analysis strategy is more efficient than a single end-of-study analysis in terms of power per unit of sample size.     

The relationship between mean square differences and standard error of measurement: comment on Barchard (2012)

In the discussion of mean square difference (MSD) and standard error of measurement (SEM), Barchard (2012) concluded that the MSD between 2 sets of test scores is greater than 2(SEM)2 and SEM underestimates the score difference between 2 tests when the 2 tests are not parallel. This conclusion has limitations for 2 reasons. First, strictly speaking, MSD should not be compared to SEM because they measure different things, have different assumptions, and capture different sources of errors. Second, the related proof and conclusions in Barchard hold only under the assumptions of equal reliabilities, homogeneous variances, and independent measurement errors. To address the limitations, we propose that MSD should be compared to the standard error of measurement of difference scores (SEMx-y) so that the comparison can be extended to the conditions when 2 tests have unequal reliabilities and score variances.     

Non‐parametric three‐way mixed ANOVA with aligned rank tests

Research problems that require a non‐parametric analysis of multifactor designs with repeated measures arise in the behavioural sciences. There is, however, a lack of available procedures in commonly used statistical packages. In the present study, a generalization of the aligned rank test for the two‐way interaction is proposed for the analysis of the typical sources of variation in a three‐way analysis of variance (ANOVA) with repeated measures. It can be implemented in the usual statistical packages. Its statistical properties are tested by using simulation methods with two sample sizes (n = 30 and n = 10) and three distributions (normal, exponential and double exponential). Results indicate substantial increases in power for non‐normal distributions in comparison with the usual parametric tests. Similar levels of Type I error for both parametric and aligned rank ANOVA were obtained with non‐normal distributions and large sample sizes. Degrees‐of‐freedom adjustments for Type I error control in small samples are proposed. The procedure is applied to a case study with 30 participants per group where it detects gender differences in linguistic abilities in blind children not shown previously by other methods.     

Hypothesis testing for coefficient alpha: An SEM approach

We show how to test hypotheses for coefficient alpha in three different situations: (1) hypothesis tests of whether coefficient alpha equals a prespecified value, (2) hypothesis tests involving two statistically independent sample alphas as may arise when testing the equality of coefficient alpha across groups, and (3) hypothesis tests involving two statistically dependent sample alphas as may arise when testing the equality of alpha across time or when testing the equality of alpha for two test scores within the same sample. We illustrate how these hypotheses may be tested in a structural equation-modeling framework under the assumption of normally distributed responses and also under asymptotically distribution free assumptions. The formulas for the hypothesis tests and computer code are given for four different applied examples. Supplemental materials for this article may be downloaded from http://brm.psychonomic-journals.org/content/supplemental.     

Assessing individual differences in categorical data

In cognitive modeling, data are often categorical observations taken over participants and items. Usually subsets of these observations are pooled and analyzed by a cognitive model assuming the category counts come from a multinomial distribution with the same model parameters underlying all observations. It is well known that if there are individual differences in participants and/or items, a model analysis of the pooled data may be quite misleading, and in such cases it may be appropriate to augment the cognitive model with parametric random effects assumptions. On the other hand, if random effects are incorporated into a cognitive model that is not needed, the resulting model may be more flexible than the multinomial model that assumes no heterogeneity, and this may lead to overfitting. This article presents Monte Carlo statistical tests for directly detecting individual participant and/or item heterogeneity that depend only on the data structure itself. These tests are based on the fact that heterogeneity in participants and/or items results in overdispersion of certain category count statistics. It is argued that the methods developed in the article should be applied to any set of participant 3 item categorical data prior to cognitive model-based analyses.     

A nonparametric approach for assessing latent trait unidimensionality

Assuming a nonparametric family of item response theory models, a theory-based procedure for testing the hypothesis of unidimensionality of the latent space is proposed. The asymptotic distribution of the test statistic is derived assuming unidimensionality, thereby establishing an asymptotically valid statistical test of the unidimensionality of the latent trait. Based upon a new notion of dimensionality, the test is shown to have asymptotic power 1. A 6300 trial Monte Carlo study using published item parameter estimates of widely used standardized tests indicates conservative adherence to the nominal level of significance and statistical power averaging 81 out of 100 rejections for examinee sample sizes and psychological test lengths often incurred in practice.The referees' comments were remarkably detailed and greatly enhanced the writeup and sensitized the author to certain pertinent issues. Discussions with Fritz Drasgow, Lloyd Humphreys, Dennis Jennings, Brian Junker, Robert Linn, Ratna Nandakumar, and Robin Shealy were also very useful.This research was supported by the Office of Naval Research under grant N00014-84-K-0186; NR 150-533, and by the National Science Foundation under grant DMS 85-03321.