As we’ve talked about before, accounting for biases is an important part of research design. One of the most powerful tools that a scientist has in accounting for biases that’s not of interest is a block design. That’s the topic of this module of Experimental Design. [credits] Look, it’s almost impossible to keep all of the conditions of an experiment constant for its entire length. It’s likely going to take more than a day. It’s likely going to involve more than one scientist, and more than one instrument. When you use a block design, you divide up the big experiment into many smaller experiments. Each block will have everything you’re doing, but fewer within-experiment replications. Let’s consider some examples. Maybe you need to block by cage. This would mean conducting your experiment on all the animals living in one cage, and considering it all by itself before moving onto the next cage. Maybe you block by room. Or building. Or even geographic location. You might block by day of the week. Or time of the day. Or age of the animal. Again, you’d consider all of the measurements and tests done in that block to be a mini-experiment on its own. These could then be checked against the other blocks. You may need to block by batch of reagents. Or by research assistant. If you are testing a new surgical procedure, you might need to block by surgeon. Blocking not only helps to account for unwanted variation, it confers statistical advantages as well. Within each block, you’re trying to reduce the amount of variation as much as you can. Then, you can include a factor in the statistical analysis that matches the blocks themselves. If we had blocked by cage, for example, then we can have a variable that represents all the animals in that cage. When an individual animal’s data is recorded, then, by controlling for the cage, we can make sure that we are helping to really measure within variability between subjects in the same cage rather than variability of subjects in different cages. Blocking can also help us to reduce bias in our experiments. Sometimes, research is not reproducible because the differences we think we’re seeing due to an intervention, are actually due to variability in the environment or from nuisance effects. These nuisance effects can bias the results. If we can make sure that the differences between treatment groups isn’t influenced by any nuisance effects that are also affecting outcomes, the study is unbiased. Let’s start with an animal model, and again, consider cages. Let’s say we have two cages, one in the front — where it happens to be warm and sunny — and one in the back of the room — where it happens to be cooler and darker. If we gave all the animals in one cage one treatment and all of the animals in the other cage the other treatment, then any differences we saw between animals that got different treatments could be overwhelmed by differences caused by the location of the cages in the room. Maybe the outcomes we observed were largely due to temperature and sunlight differences. If we treat each cage as a separate experiment, then some animals in each of the cages will get each of the treatments. We can then control for the location of the cage in the experiment, to remove that confounder. This has the effect of removing any treatment-control differences across cages, but preserves those within cages. Let’s take a human experiment. Maybe we want to see how a new educational intervention affects how students learn. But we know that different schools, in different places, with different students, of different socioeconomic status, are going to be important and influential. We should consider blocking by state, by school, even, perhaps by class. And I bet lots of you are nodding and thinking “this is so obvious.” Well, not to everyone. And knowing about blocking is just half the battle. The tricky part is the doing it. How do you block? First, you need to try and think of all the sources of variability that might affect your experiment. Do not ignore this step. Talk to other scientists. Read other studies. Really work at this. Then, you need to separate those sources of variability into components that you think are important and that could bias your results. Again, it might be cages. It might be days of the week. It might be ages. It might be developmental stages. Next, you need to assign subjects to the blocks so that they are as evenly distributed in the blocks as possible. You also need to assign interventions to the blocks as evenly as possible, too. We can’t control for everything. But blocking helps us to control for the things we can. By using it to reduce unwanted biases maximally, we can help ensure that our research is reproducible.