How Lesbian and Heterosexual Parents Convey Attitudes about Gender to their Children: The Role of Gendered Environments

Abstract

We studied associations among parents’ gender role attitudes, gender stereotyping in children’s environments, and children’s gender role attitudes and whether these associations were similar for families with lesbian and heterosexual parents. Fifty-seven 4- to 6-year-olds and 114 parents from the US participated. Parents completed self-report questionnaires and responded to interview questions. Researchers collected data regarding the child’s environment and attitudes about gender. Results revealed that children with lesbian mothers had less stereotyped environments and less traditional attitudes. Parental attitudes were associated with stereotyping in children’s environments and with children’s attitudes about gender. Both for lesbian and heterosexual parents, the impact of parents’ attitudes on children’s attitudes was partially mediated by the nature of children’s environments.

Appendix

We analyzed some of our data with Item Response Theory (IRT) models. IRT is a set of models and associated statistical techniques for analyzing questionnaires, tests, and other instruments containing multiple items with ordered categorical data. IRT is often considered the dominant form of contemporary psychological measurement of latent traits (Embretson and Reise 2000). Despite the long history of IRT (Rasch 1960/1980; Lord and Novick 1968), its heavy use in such fields as educational testing and rehabilitation medicine, and its many theoretical and statistical advantages (Embretson 1996), the use of IRT in psychological research has been limited, especially in developmental psychology (although see, e.g., Wellman and Liu 2004). Therefore, we provide a brief introduction to IRT. More complete introductions are available (Bond and Fox 2001; Embretson and Reise 2000; Hambleton et al. 1991).

The most common IRT models have several features in common. First, there is an object of measurement, usually a person, with an unknown level of a single trait of interest, such as an ability or a psychological attribute. The trait level is usually symbolized by θ. Second, there is an instrument of measurement, such as a questionnaire or raters, or both. The instrument of measurement is described by one or more unknown parameters that reflect aspects of the instrument, such as item difficulty or rater severity, that affect the observed response. Finally, there is a model, that is, a functional form, relating the unknown trait level and instrument parameters to the probabilities of observed response (Embretson and Reise 2000, chs. 4 and 5). In practice, the observed responses are used, in the framework of the model, to estimate the unknown parameters, yielding estimates of the level of the trait for each person, as well as information on the functioning of the measurement instrument.

Consider, for example, the Rasch (1960/1980) model, which is the simplest IRT model, appropriate for tests or questionnaires with dichotomous item responses (e.g., correct/incorrect or endorsement/non-endorsement). Under the Rasch model, each item on the test can be described by a single parameter, item difficulty. The model is of logistic form:

$$P\left( {X_{in} = 1} \right) = \frac{{\exp \left( {\theta _n - \beta _i } \right)}}{{1 + \exp \left( {\theta _n - \beta _i } \right)}}$$

where P(X _in  = 1) is the probability of a correct response or endorsement of an item by person n on item i, θ _n is the trait level of person n, and β _i is the item difficulty of item i (i.e., difficulty of a correct response or of endorsement). Data from a test or questionnaire can be analyzed using the Rasch model, using any of several commercially available software packages (e.g., FACETS, Linacre 2000; Winsteps, Linacre and Wright 2003; ConQuest, Wu et al. 1998) to yield estimates of the trait level of each person and estimates of the difficulty of each item, as well as such information as the appropriateness of the Rasch model to describe the response process (i.e. fit of the model to the data).

In this paper, we employed two extended versions of the Rasch model that incorporate more complicated measurement instruments. First, we employed the rating scale model (RSM; Andrich 1978) to analyze data from the PIQ. The RSM allows for rating scale instruments, instead of just items with dichotomous outcomes as with the Rasch model. Under the RSM, all rating scales are assumed to have the same parameters. The formula for the model is:

$$P\left( {X_{in} = x} \right) = \frac{{\exp \left( {\sum\limits_{k = 0}^x {\left[ {\theta _n - \beta _i - \tau _k } \right]} } \right)}}{{\sum\limits_{r = 0}^m {\exp \left( {\sum\limits_{k = 0}^r {\left[ {\theta _n - \beta _i - \tau _k } \right]} } \right)} }}$$

where$\sum\limits_{k = 0}^0 {\left[ {\theta _n - \beta _i - \tau _k } \right] = 0}$, $P{\left( {X_{{in}} = x} \right)}$ is the probability of a response in category x by person n on item i, the set of τ _k s are category threshold parameters (i.e., rating scale effects; similar to item difficulty, but indicative of the difficulty of being in the higher of two adjacent categories; see Andrich 1998), and all other parameters have the same interpretation as with the Rasch model. For the PIQ, θ _n is interpreted as the strength of parental traditional gender attitudes.

Second, because data for the stereotypicality of bedroom décor involved not just multiple items, but also multiple raters, we used the multifaceted Rasch model (MRM; Linacre 1989) incorporating the partial credit model (Masters 1982), which allows for multiple independent facets of measurement beyond person trait level and item difficulty. In our case, we employed three facets of measurement: person trait level, rating scale effects, and rater severity. The version of the MRM used in this article is:

$$P\left( {X_{ijn} = x} \right) = \frac{{\exp \left( {\sum\limits_{k = 0}^x {\theta _n - \gamma _j - \tau _{ik} } } \right)}}{{\sum\limits_{r = 0}^m {\exp \left( {\sum\limits_{k = 0}^r {\theta _n - \gamma _j - \tau _{ik} } } \right)} }}$$

where γ _j is the severity of rater j, and the β _i s and τ _k s are collapsed into a set of item-specific rating scale thresholds, τ _ik .

The fit of IRT models is an important concern, and can be assessed in many ways. With the RSM and MRM (as well as other models in the Rasch family, Rost 2001), model fit is generally assessed at the facet level. The fit of individual items, individual persons, and individual raters is examined, using two fit statistics, Infit and Outfit (Wright and Masters 1982). Both statistics are based on the residual scores, that is, the differences between the observed score and the predicted score. Outfit is relatively more sensitive to outliers, where the observed score is very different from the predicted score. Infit, on the other hand, is relatively more sensitive to patterns of minor misfit. Both statistics are reported two ways, as a mean-square and standardized. The expected value of the mean-square fit statistics is 1.0, whereas the standardized fit statistics have an expected value of 0. Higher values than expected indicate that the data are more random than predicted, whereas lower values indicate overfit, where the item, person, or rater is providing less information than expected. Overfit can be from any of a number of sources (Wright 1991) including multidimensionality (Masters 1988) and differential usage of the rating scale (Bowles 1999). The standardized fit statistics have associated statistical tests for significant misfit, based on a z-test. The mean-square fit statistics have associated rules of thumb for indicating when misfit or overfit is substantial enough to cause problems (Wright and Linacre 1994). When an item (or person or rater) is identified as having issues with fit, the data for that item are examined for anomalies, such as obvious miscoding, and either the anomalous data is removed, or the item is deleted. The analysis is then rerun to see if removal of the data or item affected the results.

The RSM and MRM have several important features that were useful here. First, the MRM, as well as most IRT models, is essentially unaffected by missing data. Estimates of trait levels are approximately the same with data missing as with complete data. Therefore, for example, raters need not rate every object of measurement. Second, because error is modeled explicitly, measurement reliability can be estimated directly, instead of the typical practice of estimating reliability indirectly with a statistic designed to assess a related concept (such as internal consistency with coefficient alpha). Finally, the MRM adjusts trait level estimates for rater severity. Typically, when raters are used to measure a psychological trait, rater effects are assessed with some form of interrater reliability, with the goal of maximizing the reliability, or equivalently, the interchangeability of the raters. With the MRM, raters can differ in severity, so that interchangeability is not a concern. Furthermore, other rater effects beyond variability in severity, such as differential interpretation of the rating scale, will appear as misfit and can therefore be identified. Thus, the many strengths of IRT models were especially valuable for use with the data presented here.