Quantitative Journal Club Past Meetings

2014

2013

2012

2010

2008

Abstracts

Time-dependent confounding

In this seminar, I'll start by describing time dependent confounding and the difficulty it poses for conventional analyses.I'll then describe three methods of analysis developed by Robins and his colleagues that allow for this sort of confounding, insofar as I am able to understand his papers. [This will be a short seminar.]These methods are: the parametric g formula; g-estimation and marginal structural modelling. I'll attempt to give some practical guidance on when these methods are needed and how to put them into practice.

Introduction to Discrete Choice Modelling: an eco-labelling application

There is increasing demand to study people's preferences or attitudes towards various issues that can be measured based on individual choices. In order to evaluate these choices, the discrete choice modelling (DCM) method can be utilised.

The DCM method requires a choice dataset that is obtained from a choice experiment; commonly a survey. In the experiment, a respondent is asked to indicate the preferred option between two or more choice alternatives that are described by attributes with various levels. As a result DCM enables the researcher to asses trade-offs and then quantify consumers' willingness to pay. The basic choice model is called the Multinomial logit (MNL) model, introduced by McFadden (1974), similar to the conditional logit model used in the field of biomedicine.

This talk will introduce the basics of choice modelling analysis in the context of the Eco-labelling research project being conducted in Agribusiness and Economics Research Unit (AERU) at Lincoln University.

Identifying risk factors for severe retinopathy of prematurity in preterm babies.

Retinopathy of prematurity (ROP) is an eye condition in preterm babies caused by anomalous development of blood vessels in the retina, which can result in bleeding, scarring, retinal detachment and even blindness. ROP is graded from stage 1 (mildest) to stage 5 (most severe). Traditionally, severe ROP corresponds to stage 3 or higher. Some risk factors that have been suggested include gestational age, birth weight, gender (Darlow, et al., 2005) and the level and/or variability of oxygen saturation (OS) in the blood (York, et al., 2004).

A subset of the NZ BOOST* II (Darlow, et al., 2004) was analysed to identify potential risk factors for severe ROP. The subset consists of those babies with complete data at the time of analysis and contains 201 babies in total - 23 babies with severe ROP (stage 3 or above or requiring laser treatment) and 178 babies with no ROP or stage 2 ROP and below. Six predictors were considered: gestational age, birth weight, gender and three measures of OS computed from the fist 4 weeks of oximeter recordings: the mean, standard deviation (SD) and coefficient of variation (CV = SD/mean) of OS.

Combining hierarchical Bayesian modelling and propensity score methods to compare hospital outcomes.

Two key issues in comparative studies of hospital outcomes are adequacy of case-mix adjustment and handling of random variation. Inadequate adjustment for case-mix or naïve approaches to dealing with random variation can lead to misclassification of hospitals as outliers and to erroneous estimates of variation in outcome risks. It is well known that hierarchical modelling can reduce the effect of random variation, however when there are numerous case-mix variables and study sizes are large, it can be impractical to control for all case-mix indicators by including them directly in a hierarchical model. In this talk I illustrate an alternative approach which uses the multiple category propensity score adjustment to adjust for case-mix variations, followed by hierarchical Bayesian modelling of case-mix adjusted summaries. As time permits, issues to be discussed will include: (i) the value of hierarchical Bayesian modelling; (ii) the logic of propensity score adjustment in the case of a multiple category exposure; (iii) specification and fitting of propensity score models in this case; (iv) checking covariate balance after propensity score adjustment.

Propensity scores: a beginner's guide

In 1983, Rosenbaum & Rubin published a landmark paper with the imposing title "The Central Role of the Propensity Score in Observational Studies for Causal Effects." Given that title and that Donald Rubin is no slouch as a statistician, I find it slightly worrying that - until this month - I've known absolutely zip about propensity scores. This talk is for any of you that might be in the same position. I'll cover what I've learnt so far: what a propensity score is, what it's for, how you create one and how you use it. Although I suspect you'll find that I don't have all the answers...

R with menus - R Commander software for teaching statistics to non-statisticians

R Commander is a package for R that allows users to create and run R code using menus that are similar to many point and click statistical packages. This provides a useful tool for non-statisticians in various fields to carry out statistical analyses.

At the Department of Conservation, we have only two statisticians for a staff of 1800 plus, including hundreds of graduates in science, technical and field roles. Many require significant statistical skills ranging from design to applying and interpreting statistical models. An effective way of improving their statistical skills is through training. Last year we modified our course teaching linear models and glms, changing from R scripts to R Commander.

We will look at the advantages of R Commander and demonstrate some of the material we teach, including importing data, creating graphs and carrying out statistical modelling. We will then briefly describe the results of a recent survey of DOC staff who attended last year's courses along with the information we have about the use of R Commander at local universities. We have found that R Commander works well for introducing our group of non-statisticians to using R.

Gender and the relationship between marital status and first onset of mental disorders: an example of using logistic regression on person year data to incorporate time dependent covariates

There has been debate about whether the impact of marriage on mental disorder differs for men and women. This topic is addressed by using data from 15 countries, and investigating marital status and the onset of mental disorders. After brief presentation of results, the use of cross-sectional surveys as if they were cohort studies will be discussed. The remainder of the presentation will focus on the use of discrete time survival, namely the creation of a person years data set and analysis of that by logistic regression. A very simple example will be presented and then there will be examination of the assumptions underlying this method for incorporating time dependent covariates in survival (onset) analysis

The paper on marital status is available through Psychological Medicine First View. Scott KM, Wells JE, et al. "Gender and the relationship between marital status and first onset of mood, anxiety and substance use disorders." Psychological Medicine, available through doi:10.1017/S0033291709991942

The initial paper on discrete time survival is: Efron, B. (1988). "Logistic regression, survival analysis, and the Kaplan-Meier curve." Journal of the American Statistical Association 83(402): 414-425.

A translational paper is: Willett, J. B. and J. D. Singer (1993). "Investigating onset, cessation, relapse, and recovery: why you should, and how you can, use discrete-time survival analysis to examine event occurrence." Journal of Consulting & Clinical Psychology 61(6): 952-65.

Partner relationships and mental health: findings from a 30-year longitudinal study

Marriage is known to be associated with improved mental health, with married individuals having lower rates of depression, anxiety disorders, and substance use disorders than unmarried individuals. However, few previous studies have examined the association between the duration of partner relationship and mental health.

This study examined the associations between relationship duration and mental health using data from the Christchurch Health and Development Study, a 30-year longitudinal study of a birth cohort of New Zealand-born individuals. Associations between relationship duration and mental health were examined using generalised estimating equation (GEE) models.

Adjustments were made for a wide range of potential confounding factors including measures of family background, childhood family functioning, child abuse, individual characteristics and behaviours, legal status of relationship (married/de-facto), and prior mental health problems.

Longer relationships were significantly (p<.05) associated with lower rates of depression, suicidal behaviour, alcohol and illicit drug abuse and dependence. Significant associations remained after adjustment for a wide range of covariate factors. For most outcomes, the protective effect of relationship duration was similar for males and females. These results suggest that relationship duration has a protective effect on a range of mental health outcomes, for both men and women, even after due allowance has been made for covariate factors.

Comparison of models for analysing aggregated categorical data accounting for missing item response: rates of mental health service use among Cook Islanders in New Zealand

Objective

To identify a method of establishing the level mental health service use by Cook Islanders compared with people from other ethnic groups in New Zealand accounting for missing ethnic group data.

Data

A 9 year extract from the Mental Health Information National Collection (MHINC). This is a national dataset that is reported to by mental health services throughout New Zealand.

Method

A comparison of imputation methods with Binomial and Poisson regression models are used to produce the number and rate from MHINC.

Results

Comparing the numbers of Pacific, in particular Cook Islanders, who have used mental health services in New Zealand is complicated by the around 4-5%, and as much as 10% in some years, of people with no known ethnicity.
In this talk we will compare a variety of imputation models that address this phenomenon as a missing data problem. A common method used in official analyses is presented as “naïve” imputation models. These are compared with a multiple imputation model and a hierarchical Bayes model.
Alternatively, the greatest increase is derived from looking at the data itself and noting that people with no ethnic group code had often in the past stated an ethnic group. Using this past value has reduced the number of missing ethnic groups by more than 50%.
By comparison the multiple imputation and Bayesian models seemed to yield similar results.

Conclusion

It seems that the numbers, and especially the rates, of Cook Islanders and other Pacific people who use mental health services are increased but also more reliably estimated having appropriately allocated the missing ethnicity data.

Empirical likelihood-based interval estimation in biomedical problem settings.

Nonparametric maximum likelihood estimation is now more than 50 years old, dating back to the Kaplan-Meier estimator and the related results of Thomas and Grunkemeier on pointwise interval estimation via the likelihood ratio statistic. However, Owen's pioneering 1988 paper in which he coined the term "empirical likelihood" spawned a fruitful new thread in statistical technology. In this talk I sketch the empirical likelihood solutions to several related biomedical problems, and identify their close connection to parametric estimation. As time permits, we will look at the estimation of diagnostic test likelihood ratios, both when the test result is a simple binary outcome and when the response measurement scale is continuous. We will also consider the problem of estimating the number needed to treat, a study outcome measure that is widely used to report the findings of clinical studies.