We liked the paper and appreciated the diagrams. I thought it interesting that Hulbert refers to statistics as possessing an impoverished vocabulary, with the word “error” (for example) being used in reference to many different issues. Only the “pseudo-replication” term has really been adopted. The paper is clearly well-cited. But as part of a new discussion that broke out 20 years later on (Hulbert SH. 2004. On Misinterpretations of presudoreplication and related matters: a reply to Oksanen. Oikos 104:591-597) Hulbert suspects that not everyone who has cited the paper has read it. (“Pseudo-issue” – funny – and specifically notes that he did not originally cite Likens et al. 1970 – our Week 4 reading – as a culprit.) So I’m glad that we read it. Did we understand it? Below are five paragraphs below – un-edited post-tutorial contributions from our group.
There were two aspects to the side challenge that led to these paragraphs. One was to correctly summarise the issue of pseudo-replication in 100 words. The other was to do so in a way that can be readily understood. This is a reasonable request. Marketing campaigns are aimed at people with seven full time years of education (say aged 11). In general we should aim no higher than those with nine. Scientific journal articles might aim a bit higher – but not much – another couple of years. So there are various readability scores available and Word calculates Flesch–Kincaid for you, if you tick the right box under “grammar check”. Websites such as www.readability-score.com are useful if the piece is critical. They provide a result for other indices as well – though generally fairly similar results. Does each index work equally for scientific and non-scientific writing? Do they work well in general, for scientific writing? The paragraphs are arranged in order of ascending Flesch-Kincaid scores. So we were aiming for a score of 9.0. The first one below hits 9.8!
- Coming to the wrong conclusion in your experiment because you did not do enough replicates, or did not replicate correctly so are unable to trust that your result is correct and due to the reason you were testing, rather than outside factors. For example, you are testing 2 types of potatoes to see which grows bigger. You decide to plant 20 of each type (20 replicates). You plant type 1 at the top of a field and type 2 at the bottom, however these replicate are not spaced out and are not true replicates. You may conclude that type 1 grows bigger; however, this is actually due to drainage within the field. The bottom of the field is waterlogged and type 2 is unable to grow. To correct this, you could randomly assign each of the 40 potatoes a position in the field.
- Let’s say you want to find out which of two crops are more suitable for your field because they grow better. So you plant crop A in the western half of the field and crop B in the eastern half. After your experimental season is over crop A has developed much less than crop B so you figure, that crop B must be a more suitable crop, since it grows faster. This is pseudoreplication, because the fact that one crop grows better than the other is due to factors you haven’t controlled for: i.e. the western half gets much more intense sun hours, or has less nutrients, (worse soil quality, etc.) so you cannot say that the performance of the crop isn’t due to properties of the field. If you do true replication you would randomly allocate your crops within your field or do a random block design. Now all treatments have an equal probability of being affected by unknown sources of variance.
- Let’s take an example. You want to investigate whether forests or grasslands have higher soil moisture content. Pseudoreplication happens when you take only one patch of forest and grassland each and take multiple measurements of soil moisture only from these 2 locations. It does not represent the behaviour of forests and grassland in general as was first intended, but only compares properties of those 2 locations. To avoid pseudoreplication one must take measurements from multiple patches of forest and grassland to get an idea how these two habitats behave generally. To avoid pseudoreplication in most of the experimental designs, samples should be independent, i.e not sharing characteristics that are different from other samples.
- For the purpose of this explanation we investigate the effectiveness of a de-worming treatment on Sheep (worms being parasitic, harmful to sheep survival and development). The worm density should be measured with (a) and without the treatment (b). Pseudoreplication, which is bad experimental practice, would happen here if the repeated treatments (a&b) were to happen on individual sheep only, the same species of sheep only or on sheep found in one location only. The measurements of worm density in the sheep gut are all linked by shared organism, species-level traits or environmental condition no matter how many measurements are recorded. To gain independent measurements of treatments we need to capture the range conditions we can see.
- If the repetition of your experiment is to produce multiple copies of the same treatment which share the entirely same variables/ manipulation/ situation/ site, then the factors are not independent and using the results of this in your experiment will mean it is pseudoreplicated. E.g if you are testing the number of rabbits in coniferous forests, if you do multiple replicates in one or two sites of one confer forest this will be pseudoreplication because the effect you find will not be independent. Instead you should do fewer replicates in multiple forests.