The efficacy of primary prophylactic treatment for opportunistic infections can be estimated in an observational cohort study by adjusting for clinical and laboratory markers of the immunodeficiency (e.g., oral candidiasis, CD4%, lymphocyte cell counts) as time-dependent co-variates (providing that the treatment does not directly alter the markers). However, the CD4 cell count provides an incomplete measure of the protective immune response, and the efficacy of treatment may be underestimated if there is inadequate adjustment for the severity of immunodeficiency. Unlike prophylactic therapies, the efficacy of which remains relatively constant over time, antiretroviral therapy may produce only transient or time-limited benefits. This problem can be minimized by allowing the effect of antiretroviral therapy to vary over time in Cox proportional hazards models (i.e., to allow the antiretroviral therapy coefficient to change over time). Another difficulty is that CD4 cell counts may underestimate the degree of immunodeficiency after prolonged zidovudine (AZT) monotherapy. If post-antiretroviral therapy CD4 cell counts are used to adjust for the stage of immunodeficiency, it may therefore be helpful to adjust for the duration of antiretroviral therapy with the CD4 cell count at the time of starting antiretroviral therapy. It is interesting to consider statistical models of progressive HIV-induced immunodeficiency in the context of the evolution of host immunity. HIV infection results in the loss of the relatively recently evolved adaptive CD4 T cell-mediated immunity to intracellular parasites. The infected host may compensate for this by making greater use of phylogenetically ancient, more innate protective responses. Because these compensatory responses are polymorphic, this results in the appearance of differences between individuals in the immune response to HIV as the disease progresses. Data from the Western Australia HIV Cohort Study support a two-stage model of immunopathology. The first stage of this model involves a loss of mucosal immunity and occurs at a variable CD4 cell count (of between 400 cells/mm3 and zero), and is marked by a loss of cutaneous delayed-type hypersensitivity responses and oral candidiasis, seborrheic dermatitis, and Pneumocystis carinii pneumonia. The second stage of the model involves a loss of systemic immunity and requires profound CD4 T-cell lymphopenia (CD4 cell count <50 cells/mm3), and is marked by infections such as cytomegalovirus and disseminated Mycobacterium avium infection. The influence of HLA type on the risk for such opportunistic infections becomes apparent during this late phase.