This chapter examined various approaches and methodologies that have been utilized in measuring neighbourhood e↵ects. A fundamental issue in capturing neighbourhood e↵ects is to define neighbourhood boundaries to capture both the dimension of the ge- ographical entity and the dimension of the social space where interactions take place. The majority of data available for neighbourhood research such as electoral districts and educational authorities administratively define neighbourhoods and capture only spatial boundaries (Sloggett and Joshi, 1998; Garner and Raudenbush, 1991). Commu- nity surveys include social aspects of neighbourhoods but are difficult to implement and involve inconsistent boundaries. Taking into consideration the incongruence between spatial and social characteristics in delineating neighbourhood boundaries, special at- tention was paid in selecting a measure to define neighbourhoods in the current research. The study employs Lower Super Output Areas (LSOAs) to operationalize neighbour- hoods. LSOAs have consistent geographic boundaries (32,482 small areas in England and Wales of around 1,500 people) and were developed to include areas that share similar social characteristics and thus capture both spatial and social neighbourhood characteristics.
After operationalizing neighbourhoods, it is important to decide whether an experi- mental or an observational methodological approach will be employed. Experimental designs involve random assignment of individuals to neighbourhoods and therefore permit the investigation of how a change in neighbourhood context influences young people’s outcomes. Additionally, due to random assignment to treatment, experiments
allow for causality of neighbourhood characteristics to be tested. Compared to the rest of the approaches, they provide a better estimate of true neighbourhood e↵ects by minimizing selection bias as a problem. However, they are difficult to implement because of practical and ethical concerns. The next best solution is quasiexperimen- tal designs which involve comparable groups of similar individuals or families. Quasi experimental designs allow selection biases to be reduced and causal relationships to be established. They are more easily implemented than randomised designs, however unmeasured di↵erences may still a↵ect the results. Observational data on the other hand, include longitudinal and cross-sectional studies. Longitudinal studies involve a large range of socio-economic status and income characteristics for families and neigh- bourhoods. Therefore, they allow causal relationships to be tested and selection bias to be reduced. Longitudinal studies permit the researcher to study changes in neighbour- hood characteristics over time. Cross-sectional studies are the least preferred approach. Cross-sectional studies involve observations and examine correlations between charac- teristics of neighbourhoods, families and young people at one specific point in time. They reflect associations at the time the census was taken and therefore they do not permit causal relationships to be investigated.
Analysis based on observational data is the most common approach, however careful statistical modeling needs to be employed to address the selection bias issue. The most commonly used statistical approaches are regression models controlling for many vari- ables considered to a↵ect selection of neighbourhood and instrumental variables such as for example the sibling fixed-e↵ects models. Regression models need to include a wide range of individual and family variables in the analysis to avoid omitting unmea- sured family characteristics that a↵ect both neighbourhood choice and young people’s outcomes and could lead to omitted variables or selection bias problems which will subsequently lead to over or under estimates of neighbourhood e↵ects. In the instru- mental variables approach, an instrument is used to produce a consistent estimator of a parameter when the explanatory variables are correlated with the error terms. The
instrumental variables technique is subject to large standard errors and IV estimators only capture the e↵ect of the treatment on the subset of the sample that is on the margin. Siblings fixed-e↵ects models allow the researcher to di↵erence out the unob- served heterogeneity in the family fixed e↵ects, such as parental ability however, the sibling fixed-e↵ects models often have large standard errors and do not control for un- observed family characteristics that vary over time and are di↵erent between siblings. A relatively new approach in the social science literature is propensity score matching and sensitivity analysis. Propensity score matching approximates a randomised trial by comparing outcomes among units that received a treatment versus those that did not and aims to control for overt bias, that is, bias that can be seen and controlled for. Sensitivity analysis is employed to indicate the magnitude of hidden bias in propensity score analysis. Hidden bias refers to bias that is caused by unobserved characteris- tics that are not included in the analysis such as for example individual motivation or ability.
Dataset Description
5.1
Introduction
This chapter introduces the dataset which will form the basis of the current analysis on neighbourhood e↵ects on young people’s outcomes. It is well established that a number of longitudinal studies o↵er rich and high quality data in the UK to inform and assess policy. Section 5.2 reviews the key goals and areas of interest covered by administrative datasets that investigate young people in their early teens in the UK. Four datasets will be examined and compared with the LSYPE study to explain the choice of dataset. The key criterion that will drive the final decision is to use a longitudinal dataset that will allow the research hypothesis of this study to be tested and causal associations to be investigated. In particular, a dataset is required with information on young people and their transitions as they move from compulsory education to further education or to economic activities, a dataset that will provide neighbourhood deprivation data and rich data to control for family, individual, school and peer group characteristics. For the purposes of the current research, a large scale dataset was selected, the Lon- gitudinal Study of Young People in England (LSYPE). The LSYPE follows the tran- sitions of a representative cohort of young people in England into adulthood o↵ering
one of the most detailed and in-depth data sources on young people in the UK today. Section 5.3 will introduce the LSYPE and describe the uniqueness of the dataset in relation to the data provided and the goals of this thesis. Section 5.3.1 will investigate how the LSYPE content will provide an understanding of the trajectories of individual life histories and of the dynamic processes that a↵ect young people in their area of residence. Sections 5.3.2 to 5.3.4 provide a description of the sampling approach, the data collection methods, the achieved sample of the study, the weights used to accu- rately represent the population of young people and their families, and the temporary or permanent loss of sample members due to attrition. Section 5.4 briefly introduces the two administrative datasets linked to the LSYPE (the National Pupil Database (NPD) and school level data) and describes in detail the geographical indicator that will be employed in the analysis, the IMD and its seven decomposed indices, that were linked to the Wave 1 of the study. Section 5.5 describes the analytic sample employed
in the analysis to study the main activity of young people at 18 19 and the missing
value analysis that will be conducted as a method to investigate the patterns of miss- ing data. Section 5.5.2 describes the statistical modeling approach that will be carried out to reduce the selection bias that plagues neighbourhood e↵ects studies which will involve logistic regression analysis, propensity score matching and sensitivity analysis. Section 5.5.3 introduces the variables that will be used in the analysis employing the LSYPE and the decomposed IMD.