I'm coming across questions in Biostats which ask to test the hypothesis through χ2 test that $-$

  1. A drug it better than a placebo.

    Not that the effect of the drug is not the same as the placebo.

  2. Heterozygotes with sickle cell trait are more resistant to malaria than individuals homozygous for the normal gene. (Q-27 of this book)

    Not that the proportion of the individual with sickled cell trait with heavy infection is not the same as the proportion of the normal individual with heavy infection.

What I understand of Chi-square:

The chi square is a measure (square) of deviation of an observed sample from a standard data/proportion/frequency, it does not indicate the direction of the deviation (i.e. it is not positive,negative and zero as z-score but is only positive and zero). If the data is a good fit with theoretical, the χ2 would be less or in fact zero (something out of shear chance) and the value will increase with deviation.

So, should (is) χ2 test be used to determine the direction of deviation?

  • $\begingroup$ I don't understand why you are asking this, since you have already remarked that the $\chi^2$ statistic "does not indicate the direction of the deviation." Because the test is based only on that statistic and the degrees of freedom parameter--which merely counts data--isn't it obvious you cannot obtain directional information from nothing? I suspect you really want to ask a different question about how one might make a directional test. $\endgroup$ – whuber Apr 20 '17 at 15:03
  • $\begingroup$ It could be obvious for those who have a complete idea of hypothesis testing and not a starter, to me it is just not semantic to make conclusion from χ2 statistic about the direction of deviation.I was told 'Researchers will often get a statistically-significant 'not equal', and then look at the contrast and interpret it as greater than or less than based on the observed direction,but that is not what the test formally tests.' which is also what the book I've linked has done in the mentioned sum.So my question is basically if there is a non-standard way of determining the direction,what is it? $\endgroup$ – Tyto alba Apr 20 '17 at 15:30
  • $\begingroup$ These are both tests of 2X2 contingency tables, so you could do a one-sided Fisher exact test. (Also, as long as the n's are big enough the P-value is very significant, I wouldn't be quite so dismissive of interpreting the result of the chi-square test in this way. Yes, it's not formally correct, but in practice a fixed-up one-sided test is going to give you the same answer.) $\endgroup$ – David Wright Apr 20 '17 at 20:48

Directionality in $2\times 2$ tables

Let us restrisct ourselves to $2\times2$ contingency tables first. I'm taking example from Wikipedia (distribution of white and blue collar workers in neighbourhoods $A$ and $B$). The observed values are as follows:

enter image description here .

Now we calculate deviations from expected values called residuals in the following way:

$$n_{i,j}-\frac{n_{i,\bullet}\times n_{\bullet,j}}{n_{\bullet,\bullet}}$$

This is simply substracting expected value of a cell from a observed value of a cell. Expected value for a cell $n_{W,A}=120\times 150/230$. We get the following table of residuals:

enter image description here

First thing you notice is that residuals sum to zero and that absolute observed values of residuals are all the same (we only work with one $df$). You notice there is 11.74 more White collar workers in neighbourhood $A$ than expected. So it follows there is 11.74 more Blue collar workers in neighbourhood $B$. Neighbourhood $A$ is more White collar and $B$ is more Blue collar. Directionality in $2\times 2$ tables is obvious. If the value of $\chi^2$ test is statistically significant we can conclude that: $(1)$ neighbourhood of residence is not independent from occupational classification AND $(2)$ that neighbourhood $A$ is more White collar and $B$ is more Blue collar.

Directionality in large tables

$\chi^2$ test in cases where $df>1$ acts as an omnibus test. It tests whether data as a whole is independently distributed. To asses directionality in larger tables you have to use additional post-hoc tests to pinpoint contributing sources of dependence. For this you can use three methods:

  • Comparing cells
  • Partitioning
  • Ransacking

You can find more information about these methods in Sharpe (2015) Your Chi-Square Test is Statistically Significant: Now What? Practical Assessment, Research & Evaluation.

| cite | improve this answer | |
  • $\begingroup$ @Tyto alba are you satisfied with the answer? I think it is quite thorough so give it a look and let me know if you have additional questions. $\endgroup$ – Vivaldi Apr 29 '17 at 11:48
  • $\begingroup$ Sorry for the undesirable delay in responding, I'm in the mid of an exam. I'd love to read your entry and after some verification will respond, it is due to the inevitable latter that I'm delaying reading this thread. Hope you can bear with me. $\endgroup$ – Tyto alba Apr 29 '17 at 13:19

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.