Smith, CJ; (2007) Methodological issues when investigating the prevalence and incidence of antiretroviral-related toxicities amongst HIV positive individuals. Doctoral thesis , UCL (University College London).

Preview

Text Smith.Colette.Joanne_thesis.Redacted.pdf Download (38MB) | Preview

Abstract

The introduction of highly active antiretroviral therapy (HAART) for the treatment of HIV infection has led to dramatic reductions in morbidity. However, these highly potent drugs have been associated with a number of side effects. It is important to accurately quantify the prevalence and incidence of these toxicities, and to be able to compare the impact of specific antiretrovirals on toxicity rates in an unbiased manner. As we have become aware of potential HAART-related toxicities, there has been an increase in the frequency of monitoring of associated laboratory markers. Thus, diagnoses are made more quickly, and randomly abnormal values are more likely to be observed. There have also been changes over time in the specific antiretrovirals prescribed in HAART regimens. Consequently, newer antiretrovirals may seem to be associated with greater toxicity, even if this is not the case. Data simulations performed suggest that, when comparing the impact of specific HAART regimens on the occurrence of toxicity, considering the first measurement in a specified window leads to the most unbiased results. The choice of most appropriate cut-off of a surrogate laboratory marker of toxicity can also lead to differences in prevlance estimates. My investigations suggest that the most appropriate cut-off for each toxicity and each laboratory marker must be considered individually, but that definitions in which a confirmatory measurement is required are often appropriate. Results of a simulation study investigating a method proposed to account for unmeasured confounding, sample selection models, found this method gives unbiased treatment estimates in the situations we investigated. However, this method can lead to a lack of precision, and requires identification of a variable that is associated with treatment allocation. Thus, the use of these models in real-life settings may be limited. Mis-specification of non-linear associations between variables in regression models can lead to biased estimates. I have found that the use of multi-fractional polynomials to systematically consider the most appropriate relationship between variables is a method that is easy to apply and leads to plausible results.

Download activity - last month

Export as CSVXMLJSON

Download activity - last 12 months

Export as JSONXMLCSV

Downloads by country - last 12 months

Export as JSONXMLCSV

Archive Staff Only

View Item