1.3.1 Missing data
The previous section of this chapter reviewed the literature on gender differences in quality of life and depression among people with CHD. Most of the studies found in the literature used samples from selected community hospitals and only two studies (Ford et al., 2008; Bjerkeset et al., 2005) used population-based samples. Samples from community hospitals are relatively small and non-response and attrition introduce bias and contribute to lower the sample size. Population-based samples on the other hand have higher sample sizes but, they are also affected by non-response and attrition. None of the studies reviewed have dealt with or acknowledged the problem of missing data in their results.
Researchers in epidemiology and other disciplines are often faced with the problem of incomplete data sets, particularly when the study aims at collecting a large number of characteristics for each individual. In longitudinal studies, missing data often occur because subjects do not respond to certain questions. This situation is often referred in the literature as item non-response. Other situations common in longitudinal studies are those in which subjects do not respond to a particular wave or drop-out of the study (because of moving out or death). This situation is known as unit non-response (in contrast to initial unit non-response it is prerequisite that the respondent has participated in at least one wave). The terms attrition, drop-out, loss to follow-up and withdrawal are used interchangeably in the literature to refer to this latter form of missingness.
Therefore, in longitudinal studies the problem could be severe since we face several
32 types of missing data, such as item non-response, unit non-response and drop-out (Clarke and Hardy, 2007).
Missing data can pose serious problems for researchers because missingness can affect properties of estimators and inferences. Ignoring missing data also affects the accuracy and precision of parameter estimation. The seriousness of the problem depends in part on how much data are missing. There is no clear rule regarding how much is too much missing data. This is because potential bias is inherent whenever observations are missing (Kline, 1998). Also the number or proportion of missing observations alone is not sufficient to indicate whether missing data are an issue or not. Rather the impact of missingness is determined by the research question, the information in the observed data, and the reason for the missing data.
The implication of missingness for the analysis depends on the missingness mechanism, which is usually unknown. In handling missing data it is important to differentiate among three missing data mechanisms (Little & Rubin, 2002; Rubin, 1976): missing completely at random (MCAR), missing at random (MAR), missing not at random (MNAR)’ (Rubin, 1976, Kline, 1998, Schumacker & Lomax, 1996). MCAR refers to data where the missingness mechanism does not depend on the variable of interest or any other variable (does not depend on observed and unobserved data) (Scheffer, 2002).
With MCAR the missing data are a simple random sample of all data values, therefore MCAR reflects the highest degree of randomness of the missing data mechanism and shows no underlying reasons for missing observations that can potentially bias research findings. In practice this means that, under MCAR, the analysis of only those units with complete data gives valid inferences (Musil et al., 2002). For example, MCAR data can occur when respondents accidentally skip a question on a questionnaire or if the participant accidentally discarded the questionnaire. In these situations there is no underlying pattern to the missing observations that would contribute to biased data.
With MAR the missingness depends only on the components of a variable that are observed and not on those that are missing (Little and Rubin, 2002). MAR data show some randomness to the pattern of data omission: “For example, in a study of dietary intake, if participants with depression are less likely than those without depression to record their daily intake, then depression is a variable that predicts missing observations” (Musil et al., 2002:816). MAR has a very special and important role in longitudinal studies where, essentially, it implies that future drop-out is conditionally
33 independent of future values, given all observed past values. Another way of expressing MAR in the longitudinal/drop-out setting is to say that the future statistical behaviour of those who share the same history of measurements is the same whether they drop-out or not. MNAR or non-ignorable missing data occurs when missingness is related to the values that would have been observed. This is the most difficult condition to model.
Non-ignorable missing data have systematic non-random factors underlying the occurrence of the missing values that are not apparent or otherwise measured. Non-ignorable missing data affect generalisability of research findings and may bias parameter estimates, also the direction of bias is unpredictable (Musil et al., 2002).
The ELSA study is subject to missing data due to item non-response and attrition. Using a simulation study I will compare three techniques for dealing with missing data in order to find the best method to be applied to the ELSA data.
The reader is referred to Chapter 3 for a literature review on missing data and for a study of three techniques for handling missing data.
1.3.2 Self-reported measure of disease
The findings reviewed in section 1.2 of the chapter mainly come from community hospitals samples, which use a clinical diagnosis of CHD. Only two studies have used population-based samples from which data on CHD diagnosis were self-reported (Ford et al., 2008; Bjerkeset et al., 2005) and was not validated with medical records or a clinical diagnosis. Most epidemiological studies and health surveys assess the presence of chronic diseases from self-report, as opposed to clinical assessments mainly because the collection of self-reported conditions involves lower costs (Kriegsman et al., 1996).
However, to use self-reported data to assess CHD with confidence, it is important to know the validity of these measures. Clearly, inaccurate reporting of CHD by surveyed populations may result in people not being identified early for chronic disease-related illnesses or not being offered interventions, such as changes in health behaviours. In terms of findings, results from studies using self-report measures might be subject to misclassification bias.
Misclassification bias is defined as the systematic error due to erroneous classification.
When assessing misclassification bias of a test or measure sensitivity and specificity must be considered. The terms sensitivity and specificity are used to measure the
34 effectiveness of a test procedure in relation to a certain disease. Sensitivity is the proportion of those with the disease that are identified as positive by the test, therefore sensitivity measures how well the test detects a disease. Specificity is the proportion of those without the disease that are identified by the test as not having the disease. It follows that specificity refers to how well the test detects absence of disease (Armitage, Berry and Matthews, 2002). The ideal value of both sensitivity and specificity is 100%
indicating no misclassification. However, the relationship between these two measures tend to be inverse that is, the more sensitive a test procedure, the less specific it is likely to be, and vice versa.
Previous studies have reported high values of specificity and sensitivity of self-reported CHD (Haapanen et al., 1997; Baumeister et al., 2010). Although a number of validation studies have suggested that self-reported CHD is reasonably accurate when compared with medical records (Bush et al., 1989; Okura et al., 2004; Merkin et al., 2007;
Yamagishi et al., 2009; Barr et al., 2009; Lampe et al., 1999; Baumeister et al., 2010), the extent to which self-reported measures introduced bias in the findings of epidemiological studies is an issue rarely addressed quantitatively (Jurek et al., 2006).
Since it is not possible to validate self-reported CHD cases in ELSA using medical records, through a sensitivity analysis I will investigate the extent to which the self-reported measure of CHD used in this thesis may lead to biased estimates and/or different conclusions in the results.
The reader is referred to Chapter 5 for a literature review on validation studies on self-reported CHD and for a sensitivity analysis investigating bias due to misclassification of self-reported CHD.