Performing regression and trend analysis with confidence intervals

[George Freer]

Hi all, my project involves assessing the trends in prevalence of anaemia in people living with HIV (PLWH) over a 20-year study period. To do this I have performed an annual cross-sectional analysis and calculated the point prevalence anaemia in PLWH over each year of the study time frame. Having plotted these on a graph I can see a clear linear association of increasing prevalence over time.

My uncertainty lies around what further statistical analysis I can do with these findings. I can only think to plot a linear regression line but was wondering if something else is more appropriate. I am keen to generate confidence intervals which take into account the number of people in each annual analysis.

Thanks
George

[Clyde Schechter]

I understand that you have performed analyses of the annual cross sections. But what is the design of the data collection? Is this a single cohort that has been followed for 20 years? Or is the data itself a series of 20 cross-sectional data collections involving (mostly) different people at each time period? The implications, and the type of analysis you would do are rather different. Similarly, does your original data source provide only a dichotomous indicator of anemic/non-anemic, or do you have actual values of hemoglobin or hematocrit?

[George Freer]

It is a binary outcome variable of anemia. Each cross-section is taken from a single cohort of individuals. In each year there will be an overlap between the people included but it won't be identical.

[Clyde Schechter]

OK. Graphing the prevalence of anemia over time in a -line- or -connect- plot is appropriate here. It might be helpful to also, since you say the trend looks linear, to do a linear regression of prevalence against time and include the regression equation as a note on the plot or something like that.

But there are some more interesting analyses that I think you should do, assuming it is possible in your data. Anemia, really, is not a disease. It is a physiological condition that has heterogeneous underlying causal diseases with known epidemiology. For example, in children in the economically advanced world, most anemia is due to iron-deficient diets. In pre-menopausal women, most of it is due to menstrual losses with inadequate dietary iron replacement. In men of any age and post-menopausal women, chronic diseases are the commonest causes. It is possible that the overall trend you are seeing masks different trends in these different sub-populations. Also, I don't know where your data comes from: if it is from economically developing countries, the epidemiology is rather different from what I just outlined. So graphing, analyzing, and comparing the trends in the appropriate subpopulations would make for a more informative presentation, and might include findings that could be useful in guiding public health responses. Similarly, if your data distinguish between different hematologic types of anemia (microcytic, macrocytic, hemolyltic, dysplastic) seeing the time trends in those categories separately would be desirable.