Estimation and testing in the random effects probit model

Guilkey, David K.; Murphy, James L.

doi:10.1016/0304-4076(93)90028-4

Cited by 163 publications

(78 citation statements)

References 9 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…For example, in the 2009 cross section there were 8.9 respondents, on average, per cluster. In a variety of analyses with these data we find empirically that correcting for the cluster design rarely matters and this is consistent with Monte Carlo simulations on the effect of correcting for cluster design (Guilkey and Murphy 1993). 20 In Appendix Table 1 Table 4.)…”

Section: Life Course Events and Panel Retentionsupporting

confidence: 70%

Do low survey response rates bias results? Evidence from Japan

Rindfuss¹,

Choe²,

Tsuya³

et al. 2015

DemRes

164

114

View full text Add to dashboard Cite

BACKGROUNDIn developed countries, response rates have dropped to such low levels that many in the population field question whether the data can provide unbiased results. OBJECTIVEThe paper uses three Japanese surveys conducted in the 2000s to ask whether low survey response rates bias results. A secondary objective is to bring results reported in the survey response literature to the attention of the demographic research community. METHODSUsing a longitudinal survey as well as paradata from a cross-sectional survey, a variety of statistical techniques (chi square, analysis of variance (ANOVA), logistic regression, ordered probit or ordinary least squares regression (OLS), as appropriate) are used to examine response-rate bias. RESULTSEvidence of response-rate bias is found for the univariate distributions of some demographic characteristics, behaviors, and attitudinal items. But when examining relationships between variables in a multivariate analysis, controlling for a variety of background variables, for most dependent variables we do not find evidence of bias from low response rates. COMMENTSThe results have two implications. First, demographers should not presume the presence or absence of low response-rate bias; rather they should test for it in the context of a specific substantive analysis. Second, demographers should lobby data gatherers to collect as much paradata as possible so that rigorous tests for low response-rate bias are possible. IntroductionSample surveys (cross-sectional and longitudinal) have become the dominant data source used by population researchers. Response rates, both the initial response rate and attrition rates in longitudinal studies, have historically been an important rough-andready yardstick to judge data quality. Response rates 6 have been declining in urbanized, high-income countries to the point where many cross-sectional surveys now have response rates below 50% (Atrostic et al. 2001;Brick and Williams 2013;de Leeuw and de Heer 2002;Groves 2011;Singer 2006). The survey literature has long recognized that low response rates only indicate potential bias (e.g., Lessler and Kalsbeck 1992), yet the almost automatic response among most in the population field has been to equate low response rates with poor data quality. Low response rates produce bias only to the extent that there are differences between responders and non-responders on the estimate(s) of interest, and then only if such differences cannot be eliminated or controlled for through the use of observable and available characteristics of responders and non-responders. The difficulty in evaluating non-response bias is that measures of the variable(s) of interest and characteristics of non-responders are generally not observed, and hence the population field has tended to rely on the presumption that low response rates necessarily mean low data quality.In this paper, using surveys designed by demographers and variables that have often been used as dependent variables in demographic research, we examine this 6 ...

show abstract

Section: Life Course Events and Panel Retentionsupporting

confidence: 70%

Do low survey response rates bias results? Evidence from Japan

Rindfuss¹,

Choe²,

Tsuya³

et al. 2015

DemRes

164

114

View full text Add to dashboard Cite

show abstract

“…A comparison of restricted`exact' and simulated ML for this error decomposition can be found in Guilkey and Murphy (1993). The error structure can be made more exible by allowing to vary over time and by the possible introduction of more than one factor (cf.…”

Section: Maximum Likelihood Estimationmentioning

confidence: 99%

“…W e use the algorithm suggested by Butler and Mo tt (1982). The number of evaluation points V is set to 5 as a compromise between computational speed and numerical accuracy (see Guilkey and Murphy, 1993, for more Monte Carlo results). When the assumed error structure is the true one this estimator is consistent and asymptotically ecient for ( ).…”

Section: Estimatorsmentioning

confidence: 99%

Convenient estimators for the panel probit model

Bertschek

Lechner

1998

Journal of Econometrics

View full text Add to dashboard Cite

The paper shows that several estimators for the panel probit model suggested in the literature belong to a common class of GMM estimators. They are relatively easy to compute because they are based on conditional moment restrictions involving univariate moments of the binary dependent v ariable only. Applying nonparametric methods we discuss an estimator that is optimal in this class. A Monte Carlo study shows that a particular variant of this estimator has good small sample properties and that the e ciency loss compared to maximumlikelihood is small. An application to the product innovation decisions of German rms reveals the expected e ciency gains.KEY WORDS: panel probit model, GMM, conditional moment restrictions, k-nearest neighbor estimation.

show abstract

“…The most e¢ cient method of computation that leads to the so called 'random e¤ects probit estimator'uses the Hermite integration formula (Butler and Mo¢ t, 1982). See also the paper by Guilkey and Murphy (1993) for more details on this model and estimator as well as for more discussion about the numerical algorithm.…”

Section: Maximum Likelihood Estimationmentioning

confidence: 99%