| Newsflash | ||
|---|---|---|
|
||
| Polls |
|---|
| 20% + off all books |
|---|
|
 |
| Inference: Making Educated Guesses |
| Columns - Research to Practice | |
| Monday, 30 September 2002 | |
|
We have arrived at the next to the last of these
research/statistics columns. Everything previously discussed in this column was
to prepare you for inferential statistics. Addiction questions abound with
inference questions. For example, counselors are always asking, is the treatment
technique I selected working for this particular client? What am I doing as a
counselor that contributes to better or poorer outcomes? What variables account
for relapse in a particular clinical group or client? Prevention questions are
similar. For example, what are the effects of my drug and alcohol awareness
program in the drinking habits of early adolescents? For these, and many other
questions, we need to go beyond merely describing data. We now want to know
cause and effect, and attempt to make predictions from that information
(Kranzler & Moursund, 1999).
It is only recently that our field has
had to come to terms with inference. The reasons are the demands of managed care
and the general higher accountability facing our clinical programs. Without a
doubt, inference will take a more predominant role in the future of addictions
treatment. You may spout off some interesting testimony about how effective your
program is, but rest assured that someone, somewhere, someplace is going to say
to you ? prove it. Hence, it is critical you understand this statistical form.
Hypotheses So, you’re becoming interested in research and statistics and thinking it really could be useful in your clinical work. What is the next tangible step? Generally, you develop a scientific hypothesis. This is an objective criterion and it is used to determine if your hypothesis should be accepted as true or rejected as false (Polit & Hunger, 1989). For example, if you are interested in the client outcome from your program, you will first want to write out a research hypothesis. That hypothesis can take the form of, “More clients maintain abstinence following treatment in my program than those who attend no program.” This statement is an expectation you have of your program. Next, the researcher needs to set up a hypothesis that is the exact opposite of the research hypothesis. It will read something like, “There will be no difference in abstinence rates of clients who attend my program versus those who attend no program.” This opposite hypothesis is called a null hypothesis (H0). It is odd, but in science, we can’t directly support a truth. Basically, there is always the possibility that one day your theory will have an exception to the rule, be it water flowing uphill, to spontaneous regeneration of brain cells damaged from years of alcohol abuse. So, we have to investigate things sort of backwards. Statistics is good at this backwards approach because it is able to demonstrate that something is false. Given this state of affairs, we are in an interesting situation. In order to demonstrate that something is true, we have to use tools that can only show that something is not true (Kranzler & Moursund, 1999). Simplified, state the exact opposite of what we want to demonstrate and disprove that. What is left must be true. The whole idea of inference is to reject the null hypothesis. To illustrate, let’s make a statement like, “There is no link between certain genetic dispositions and problematic drinking.” Try to disprove that statement through experiments. The exceptions you find to this statement must be true. The null hypothesis is one of those awkward phrases that simply mean there is no relationship between the variables you are studying. It essentially takes in every possibility that exists except the one thing you would like to prove. The figure below should help illustrate this. For example, the entire block represents all the possible ways the world might be. The gray area represents everything that isn’t what you would like to prove. The clear circle represents the thing you would like to prove. If you can rule out all the gray areas, the only thing that must be true is the clear area. This is what you wanted to prove in the first place (Kranzler & Moursund, 1999). Feel confident about your results? Beware the errors If you did a study on the treatment outcome of your program, and you are feeling confident about your results, just make sure you haven’t run into any unforeseen errors. What are we taking about? Well, how probable is it that your study on outcome is due to chance alone? Remember you are trying to infer from a sample to a whole population. You can’t flatly say that, because you will rarely be able to get all the information about the entire population. So, you have to accept the idea that my hypothesis is mostly true or mostly false. That “mostly thing” presents a problem. Because your outcome study is mostly true or mostly false, you are going to have to take a risk that there is a chance that you could be wrong. This risk can lead to certain types of errors. For example, if there is no difference in outcome between your experimental group and control group, but you say, “oh no, there really is,” then you have made a type I error (rejected a true hypothesis). If, however, there is a real difference between your groups and you claim, “no there isn’t,” you have made a type II error (rejected a false hypothesis). But not all is lost, you do have some control over not making these errors. Level of significance We have taken pains to point out that anytime you do an experiment, you risk not being 100 percent confident of your results because you cannot control all the factors (including chance) that can affect your results. Yet, we want to have some confidence in our results. So we rely on a thing called significance, which is a probably that an observed value could not have occurred by chance. Using our example, you can never be positive that your program significantly helped your clients. So, you have to define a region of risk you are willing to take. Typically, two acceptable levels of significance are used, .05 and .01. If you decide to use the .05, you are saying that out of 100 samples you are willing to erroneously risk that a true null hypothesis will be rejected five times. With .01, you are willing to risk rejecting a true null hypothesis only one time out of 100. Using our outcome example, if your statistical results fall within these regions (.05.,01) that says there is a high probably that your result could not have occurred by chance. Something else is going on, so go ahead and reject the null hypothesis (which says nothing is going on). In this case, H0 is not an attractive hypothesis. Instead, you can favor your research hypothesis that something is going on, and that something is my program. Well, you say, “why not lower the type I error as much as possible?” The problem with that idea is that as you lower the type I error you inadvertently increase the type II error. To avoid that, we have to balance both and live with the risks. As usual, there is a lot more to hypotheses, errors, and significance thing, but if you understand what was just presented, you understand inference. In review Try out all this new information. Take a few minutes and think of a research project you would like to conduct. Get that project firmly in your mind, and then apply the following steps (Salkind, 2000). 1. State what you would like your outcome to look like then state the opposite or your null hypothesis. 2. Set the level of risk (.05 or .01). 3. Select the proper test statistic (See next column). 4. Compute the test statistic result. (It will be some numerical value) 5. Determine the critical value needed to reject the null hypothesis. (You usually find this critical value in one of those complicated looking tables found in the back of statistics books, or a statistical software package will do it for you.) 6. Compare the obtained value to the critical value. 7. If the obtained value is more extreme than the critical value, the null hypothesis is rejected. 8. If the obtained value does not exceed the critical value, the null hypothesis is the best explanation. This column is about as simplified and as you can get. The whole process can get a little more complicated. (See the references for more information.) But, if you understand this portion, the other parts will fall into place. For our last column, we cover some of the major statistical tests, what they mean, and how to select the best one for any research. Michael J. Taleff, PhD, CSAC, MAC, is the Alcohol & Drug Education Program Coordinator for the University of Hawai’i at Manoa. References Kranzler, G. & Moursund, J. (1999). Statistics for the terrified (2nd ed.). Upper Saddle River, NJ: Prentice Hall. Polit, D.F. & Hunger, B.P. (1989). Essentials of nursing research: Methods, appraisal, and utiliztion (2nd Ed.) Philadelphia, J.B. Lippincott Co. Salkind, N.J. (2000). Statistics for people who (think they) hate statistics. Thousand Oaks, CA: Sage. Vogt, W.P. (1999). Dictionary of statistics and methodology: A nontechnical guide for the social sciences. Thousand Oaks, CA: Sage.              |
|
| < Prev | Next > |
|---|