This web utility may be cited in APA style in the following manner:
This web page is intended to provide a brief introduction to chi-square tests of independence and goodness-of-fit. These tests are used to detect group differences using frequency (count) data. This page also provides an interactive tool allowing researchers to conduct chi-square tests for their own research. Any introductory applied statistics text should have a good description of these chi-square tests, but following is a condensed introduction.
Often a researcher wishes to see if the frequency of cases possessing some quality varies among levels of a given factor or among combinations of levels of two or more factors. In such situations, the appropriate test is the chi-square test of goodness of fit or the chi-square test of independence for k groups.
To conduct the chi-square test, the researcher enters observed frequencies corresponding to combinations of levels of relevant factors (here, called "condition" and "group," but these are labels of convenience). Sums of elements within rows and within columns are then computed (call these marginal Ns). The chi-square test of independence is used to test the null hypothesis that the frequency within cells is what would be expected, given these marginal Ns. The chi-square test of goodness of fit is used to test the hypothesis that the total sample N is distributed evenly among all levels of the relevant factor.
The expected value within each cell, if the null condition is true (i.e., if the factors have no significant influence on observed frequencies in the population), is simply the product of the row total and column total divided by the overall sample N for the test of independence and N divided by the number of levels of the single factor for the test of goodness of fit. If Oij is the observed frequency and Eij the expected frequency for the cell corresponding to the ith condition and the jth group, then chi-square is:
If there is only one factor of interest with (k > 1) levels, the same formula will work, with i or j being set to 1. The test presented here can be used to test only 1- or 2-dimensional arrays. Arrays of higher dimension are possible, and are based on the same principle and even use the same formula, although they involve multiple nested summations.
Input observed frequencies into the white cells. I realize that not very many designs involve exactly 10 conditions and 10 groups - if your design is smaller, then choose some subset of rows and columns in which to enter your data. For example, if your design is (2 x 3), then you may choose to enter your data in the 6 cells in the upper left portion of the data table, defined by the first two Conditions and the first three Groups. You can choose any subset of rows and columns for your data. You can also opt to leave cells corresponding to observed frequencies of zero blank. Non-integer observed frequencies are allowed, although it is difficult to imagine how one would obtain these in actual research.
If you are performing a test of goodness of fit, you may choose to enter your data in any single column or row. However, observed zero frequencies need to be explicitly included (i.e., you'll need to actually type "0" in those cells, otherwise it is assumed that those cells are not part of your design). Once you have entered your data, click on the Calculate button and expect to see output in the beige cells (they should be white if you are using older versions of Netscape). Do not panic if you see scientific notation for your p-value - that simply means that p is really small.
This tool also yields a chi-square incorporating Yates' correction for continuity. This correction is often employed to improve the accuracy of the null-condition sampling distribution of chi-square. It probably should be used only for 1-df tests (i.e., goodness of fit tests or tests of independence with 2x2 contingency tables), so use at your own risk for tests with df>1.
Use of the chi-square tests is inappropriate if any expected frequency is below 1 or if the expected frequency is less than 5 in more than 20% of your cells. The status cell at the bottom of the table will let you know if there is a problem. In the 2 x 2 case of the chi-square test of independence, expected frequencies less than 5 are usually considered acceptable if Yates' correction is employed.
Finally, if any expected frequencies are 0, the chi-square statistic is undefined (division by zero). This will happen if you enter a row or column consisting entirely of 0 values. Avoid this!
When using the chi-square goodness of fit test, sometimes it is useful to be able to specify your own expected frequencies. If there is a theoretical reason for doing so, the following table will allow you to enter your own Eij's. Non-integer expected frequencies are allowed. Use as many cells in this table as necessary, making sure that (1) the marginal total is the same for both observed and expected frequencies, (2) there are no expected frequencies less than 1, and (3) no more than 20% of your expected frequencies are less than 5. If a frequency is entered in an Observed cell, then a frequency must also be entered in the corresponding Expected cell (and vice versa).