Social Icons

Pages

Showing posts with label Statistics. Show all posts
Showing posts with label Statistics. Show all posts

Monday, 8 February 2010

Statistics in dental materials research: 2 things to plan ahead

A couple of statistical issues should be considered when designing a study in dental materials science.

(1) What statistical test(s) will be used to test the null hypothesis/hypotheses?

It is recommended to design the experiment in such a way that it is possible to test the data using one “global” test such as analysis of variance (ANOVA). The effect of one independent factor on one response variable is tested using one-way ANOVA in many papers on dental materials. Such study design includes e.g. the comparison of the degree of conversion (DC) of several materials cured under the same curing conditions (light, intensity, time, distance). The null hypothesis would be that there is no difference between the means for different materials. So, the independent factor is ‘material’ and the response variable is ‘DC’.

Two-way ANOVA is used to test the effect of two independent factors, e.g. material and light-curing unit (e.g. 3 materials are cured with either a halogen or an LED light-curing unit). Testing for interaction between the two factors shows whether or not the differences caused by one factor are consistent on different levels of the other factor. If so, the interaction is not significant (e.g. the DC may be higher in each material when cured with a halogen unit than an LED unit). Alternatively, if these differences are not consistent, then the interaction is significant (the DC may be higher in some materials when cured with a halogen and in others when cured with an LED unit). In this case, a series of one-way ANOVA must be used to examine this interaction more closely. This will, however, result in multiple testing which by default increases the chance of making the Type I Error (rejecting the null hypothesis when it is true) and some sort of correction is necessary to keep the overall significance level at the usual alpha=0.05. This correction most often means a decrease in the individual alpha value which also reduces the power of the statistical test.

Three-way ANOVA is sometimes used in dental materials science to study the effect of three independent factors on a particular response variable (e.g. the effect of material, light-curing unit and curing time on the DC of resin-based composites). Researchers are often tempted to test more and more factors in order to make their experiments robust. However, one has to keep in mind that the interpretation of three-way ANOVA is more difficult that that of two-way ANOVA and the post-test corrections may significantly reduce the power of the test. These are by no means the only tests used and are only an indication of the type of studies carried out in dental materials science.

(2) Power and sample size – power is the probability of not making Type II Error (failing to reject the null hypothesis when is false) i.e. power is the probability of correctly rejecting the null hypothesis when the difference between the groups truely exists. In sample size calculation prior to an experiment, the power of 80% is generally used as the cut-off point. So, when calculating the number of samples for each group, we need to know the following: number of levels (groups) that we will be comparing; significance level; power; estimated standard deviation (determined in a pilot study or taken from the literature) and the difference between the groups that we consider clinically relevant and don’t want to miss in our statistical testing. This last one may be tricky, because we often don’t know what difference between the groups is clinically relevant in a way that might affect the clinical performance of the tested materials. In this case, we may base our decision on the literature data or we can do a pilot study to find out the likely difference in the response variable between our groups which we would then use in the sample size calculations. Alternatively, if we already have a pre-determined number of samples in each group, we may be able to determine the power of our statistical test (i.e. how certain we are that our conclusion is correct).

***********************************************************************************


Monday, 1 February 2010

Statistics in dental research: A book review

In a addition to the previous post on statistics in dental research, I'd like to mention that Medical Statistics at a Glance is the best book on the subject I've seen. It contains all the basic things a dental materials scientist needs to know, from study designs, types of data and descriptive statistics to hypothesis testing, correlation and regression, survival analysis and Bayesian methods. The book is written in an exceptionally succinct and reader-friendly way, understandable to researchers with very little previous knowledge on statistics.

Theory is only given in the amount which is necessary to understand each concept. A very good feature of the book is that it explains most commonly used statistical tests in dental research: t-tests, analysis of variance (ANOVA), the non-parametric Mann-Whitney and Kruskal-Wallis test, chi-squared and McNemar's test. The assumptions for these tests are given but situations with departures from these assumptions are mentioned in terms of their effect and possible solutions.

Also, statistics for some more complex study designs is also presented, such as generalized linear models, multiple linear regression or methods for clustered data.

Medical Statistics at a Glance also serves as a fantastic reminder with an informative glossary and a detailed index of terms. It is an excellent value for money. I bought a new copy on eBay for about £20 but I'm sure it can be found elsewhere on the internet.

***********************************************************************************

Wednesday, 27 January 2010

Statistics in dental research: A challenge for a dental materials scientist

Dental research relies heavily on statistics and in the majority of studies some sort of statistics is necessary. This goes beyond the descriptive statistics (the measures of central tendency and spread) and includes hypothesis testing using parametric or non-parametric tests. Sometimes other tests are used depending on the research question and the hypothesis. As far as I can remember, the only type of research where I haven't seen any statistics done in dental materials science is finite element analysis which involves computer simulation of stresses and strains on bone and/or tooth models. This approach does not require sampling and therefore no statistics is performed.

The validity of results and conclusions depends, among other things, on the appropriate statistical test(s). I'm pretty sure dentists and material scientists who conduct research but are not familiar with statistics feel this may be their main weekness. In all research methodology courses, it is strongly advised to consult a statistician prior to conducting a study because even in the planning stage of the study, statistics is unavoidable as it is necessary to perform sample size and power calculation. However, consulting a statistician is easier said than done simply because there are not very many statistians out there available for quick (and free of charge) consultations. It seems to be a matter of personal initiative to establish some contacts since many academic institutions don't have statisticians among their employees.

Having said that, I can't help asking myself the following when I read scientific papers: how did these authors perform statistical analysis? Did they consult a statistician? Did they do statistics themselves? What's their knowledge on this subject and did they test the hypothesis based on the correct assumptions? Did they just copy the same test from a similar paper published previously? These questions arise because in many papers only the applied test and the p value are stated. Very little or nothing is known about the assumptions for parametric testing, how the departure of the required assumptions were dealt with, possible outliers and their effect on the results, correction in multiple testing etc.

I would appreciate some input from fellow scientists so please feel free to comment on this and write your opinion. Your own or other people's experience is welcome.

************************************************************************************