I'm trying to analyse a paeds HIV cohort, and thinking of ways to describe the trend in HIV VL over time, and comparing average viral loads at the start and end of the study. Is there a generally-accepted way of dealing with undetectable HIV RNA levels? ie do I call them all 0 or 40? Ditto for log viral loads - technically there is no log 0, but should I call those values 0 or log40? Thank you all!
Hugh,
First, you could categorize your data and look at the proportion of undetectable over time. That's a standard way to analyse reponse to treatment.
If you really want to compute average Vl and make comparisons, you should log-transform your data before computing mean. For undetectable VL, you have several choices, depending on the number of undetectable measurements at each timepoint : you can leave the value as it (40--> log40), or recode undetectable VL as 0 and add 1 (0+1 --> log1=0). Another technique is to divide the threshold value by 2 (40/2--> log20). Have a look between the different methods to assess any large difference.
If you go to regression analysis, you should used tobit model for left-censored data.
Hope useful.
Hugh, in our reports we use the lower detection limit of the assay as the viral load when this is undetectable. Seems to be common that some virus persist below the detection limit of common assays.
Hugh,
First, you could categorize your data and look at the proportion of undetectable over time. That's a standard way to analyse reponse to treatment.
If you really want to compute average Vl and make comparisons, you should log-transform your data before computing mean. For undetectable VL, you have several choices, depending on the number of undetectable measurements at each timepoint : you can leave the value as it (40--> log40), or recode undetectable VL as 0 and add 1 (0+1 --> log1=0). Another technique is to divide the threshold value by 2 (40/2--> log20). Have a look between the different methods to assess any large difference.
If you go to regression analysis, you should used tobit model for left-censored data.
Hope useful.
Your answer depends of the test being used. The old technology detects 400 cp and above. If that is your case undetectable is
When discussing "undetectable" viral load test results with other HIVers, I always suggest thinking of the true number at the limit of detection minus 1 -- it makes blips easier to bear. For the sake of analytics, I would suggest considering both that approach and using 0.5 times the limit of detection as substitute values. Take a data set, apply substitute values differently in separate analyses, and see what you get. Assuming that logVL is evaluated as occurs in pharma studies, you'd see the outcomes shift by roughly 0.3 (log[0.5]).
undetectable viral load is HIV-RNA below the limit of assay detection that is actually
In prior analyses our group used an assay with a cut off value of 100 copies/ml. Any values below 100 were set to 50. Since majority (80%) of the participants had undetectable viral load levels, we eventually dichotomized the variable at the cutoff value.
Hi Hugh - I recently dealt with several studies involving HIV viral loads.
Dichotomizing your data (“0”=below detection; “1”=above) may be a simple and complementary way to analyze your data. The dichotomized data is interesting if you want to answer a question related to (for example) the proportion of patients enjoying an undetectable VL after intervention. But you may kill a bit too much information if you are interested in average or distribution of viral loads.
From what you wrote, I assume your detection limit is 40 cps/ml. In papers doing a statistical analysis on HIV VL, I mainly saw two ways to deal with undetectable VL. One is to recode your “1,000), then recoding the undetectable VL at half or at the detection limit will not make a big difference on your averages. But if most of your VLs are just above, this choice may affect the value of you averages before/after intervention… so I’d be careful. If you are interested in assessing a decrease of VL, setting the undetectable at the detection limit may feel more ‘conservative’ because you’re assuming the worst case. As mention in previous posts, I would also deal with log VL (not simply VL).
Also, if the variance of VL is important in your analysis (explicitly or implicitly), I’d be careful because recoding undetectable VL with a single value will artificially decrease the variance.
A more technical method (which may be an overkill for your purpose — here again it really depends on how your VL data is distributed) is to use latent variables: undetectable log VLs are assumed to be randomly distributed (you choose the distribution - uniformly on the interval [tiny value;log(40)] for example).
I hope this helps.
David
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis