When I create graphs or calculate descriptive statistics, should I put identifiers for each data point on the graph or next to the statistics?

Not usually. The purpose of a graph is to illustrate the major features of the data. The purpose of a statistic is to provide an index to represent a feature of the data. Adding names or other identifying marks will usually just make your report more difficult to read. An exception is when you are wanting to focus on one particular case in the dataset. Then you may want to show where that case falls in the dataset. Even then you probably want to just identify that one case, and not add labels for all of the cases. The other exception is when there is a clear outlier. It is often useful to find out what case the outlier represents. This can often help us better understand the outlier.

When I create some graphs they look "flat" and are difficult to read. Why does this happen and what should I do about it?

This is caused by an extreme outlier in the data. The software is adjusting the scale in order to accommodate all data points. When there is an extreme outlier, it has a large influence on the chosen scale so that most of the graph becomes unreadable. Try removing the outlier. You can then mention this extreme outlier in your narrative or with a side comment on the graph.

What type of graphical display can be used to illustrate power estimates?

The best kind of power display is a graph that relates power to sample size or effect size. Every power calculation requires a specific sample size or effect size. Usually, though, the researcher is interested in calculating power for various sample sizes or effect sizes. For example, consider a null hypothesis that the population proportion is 0.5. To calculate the power for a test of this hypothesis, we need to know the sample size and then speculate an actual population proportion (say, something larger than 0.5). The researcher will probably want to try various sample sizes or different population proportions. Let's suppose that we fix the population proportion at 0.6 and calculate power for the test that the population proportion is 0.5 using sample sizes of 100, 200, 300, 400, and 500. For each sample size we will obtain a different power estimate. Suppose these power estimates are 0.6 (when N = 100), 0.65 (when N = 200), 0.70 (when N = 300), 0.75 (when N = 400), and 0.80 (when N = 500). A nice graphical display would be a plot (you could use a scatterplot) with N on the horizontal axis and power on the vertical axis. This will be useful to researchers who try to balance the cost of larger samples with their attempt to use a reasonably powerful test. The same procedure could be used if you keep the sample size constant, but vary the effect size. In this example, for instance, you could use a sample size of 100 and then try proportions of 0.55, 0.60, 0.65, 0.70, and so on. Again, this could be illustrated in a graph with the proportions on the horizontal axis and power on the vertical axis.

URL http://edpsych.ed.sc.edu/seaman/edrm711/questions/graphs.htm

This web was developed by Michael A. Seaman.
This page last updated on 04 February 2001 .

The views and opinions expressed in this page are strictly those of the page author. The contents of this page have not been reviewed or approved by the University of South Carolina.