Bias is because of mistake of investigator (design or conduct of study) but confounding is not because of mistake of investigator but due to some natural phenomenon. (confounding: natural)
Bias
A systematic error (not a random error) in the design or conduct of a study which results in mistaken estimate of risk (of investigator in design or conducting the study) i.e. a wrong calculation.
Selection bias:
Selection of wrong people in the group. People selected are not the representative which happens because the study subjects are not representative of the reference population. We can remove this by randomization in selection or by random group allocation i.e. selection bias is eliminated by randomization. Example of selection bias:
- Prevalence incidence bias or Neymann bias: is a selection bias which occurs in case controlled study
- Berkisonian bias: it is a type of selection bias occurring due to different rates of hospital admission in different diseases.
Randomization does not ensure that the groups will be completely comparable on baseline characteristics or in other words randomization is not a guarantee against confounding but is a guarantee against selection bias. Randomization in RCT is not for selection of study subjects but for group allocation.
Randomization simply decreases confounding but not eliminates. It tries to make groups comparable but is not a guarantee.
Information bias: If information gathered is incorrect. For example,
- Recall bias: it can be removed (by doing a prospective study) by changing the study from retrospective to prospective.
- Measurement/assortment bias: It is due to fault in technique, instrument or reagent. It can be removed by using correct technique i.e. standardized operating procedures given by WHO or use of standardized equipment or well-calibrated instruments.
- Observer interviewer bias: it happens by giving different time for history. Blinding, so that we wont, know which is case and which is control, can eliminate it.
Combination bias: it happens by making both selection bias and information bias. For example,
- Surrogate bias
- Surveillance bias
The third factor or third variable of the study, which is associated with both the exposure and the outcome and can independently cause the outcome. Exposure is the first factor. Outcome is the second factor. Third factor is the confounding factor.
It has two criteria: Associated with exposure factor and Independently cause same outcome like exposure factor. For example, smoking (first factor) is a risk factor for coronary heart disease (second factor). Alcoholism is third factor. For alcoholism to be called confounding. It is associated with smoking as smoking and alcohol go hand in hand (and it occurs naturally). Alcohol is independent risk factor for CHD. So, if I do study, and subjects also happen to be alcoholic, some of the cases in CHD can be due to alcohol and not because of smoking. When we calculate relative risk of smoking and CHD we may get overestimate of RR, which may be wrong calculation.
Methods to reduce (not eliminate) confounding:
Confounding can be decreased at two stages
- At the time of collection of data
- Matching (known)
- Randomization (best) (both known and unknown)
- Restriction (known)
- Stratification (known)
- At the time of analysis of data
- Mantal Haenzel stratified analysis
- Regression analysis
Note:
Confounding can be completely eliminated by doing cross over RCT.
Out of this, which is best to reduce confounding?
- Matching
- Randomization (correct)
- Restriction
- Stratification
Although randomization is not a guarantee, it is the best because while doing randomization both known and unknown confounding factors will be equally distributed more or less but not completely. Rest of the three can take care of or restrict entry of only known confounding. For example, not taking alcoholics. But can be done only when we know that certain criteria is confounding.
Taking all the 6 techniques, regression analysis is the best in decreasing confounding. (Not bias)
- P stands for probability
- It is the probability of committing type I error or false positive error
- How much chance I have of making false positive mistake
- It is the type I error expressed in decimal (from percentage)
- P- value does not give information about
- It cannot give information about type II error or beta error or false negative error
- It has nothing to do with power of the study.
- If p value = 0.04 what is it?
- In this study, the probability of type I error is 4 % (false positive)
- Drug A and Drug B, p value=0.04, what do you conclude?
- Probability of false negative is 4%
- Type II error is 4 %
- Power of the study is 96%
- False positive =4% (correct)
P = 0.04 means
- The probability of false positive conclusion that drug A is better than drug B when in reality it is not is 4%.
- The difference in study is chance finding and not true finding i.e. probability of difference being because of chance not because of true difference is 4%.
- If we do the study 100 times, 96 times the difference will be statistically significant (there is 96% probability of difference because of true difference) and only 4 times the difference will be statistically non-significant.
- Among second and third statements, the third statement is best.
| Decision taken based on study results | |||
|---|---|---|---|
| Accept null hypothesis | Reject null hypothesis | ||
| As it exists in reality | Null hypothesis is true | Correct decision | Type I/False positive Error |
| Null hypothesis is false | Type II/beta/False Negative Error | Correct decision | |
Type I Error/False Positive Error:
- Rejecting a null hypothesis, which in reality is true
- In reality, there is no difference in intervention but our study shows a difference, which is false positive error i.e. in reality there is no difference between the interventions but the study has shown a difference so called false positive error.
- Obviously if there is no difference in reality but our study shows difference then it is false positive (i.e. not true positive difference)
Type II Error/Beta Error/False Negative Error:
- Accepting a null hypothesis which in reality is false
- In reality there is a difference between the interventions but the study has failed to detect that difference (due to bias or confounding) and it shows no difference.
Note:
Type I and type II errors have totally no link and no relation between both of them.
- Power is the ability of the study to reject a null hypothesis, which in reality is false. In other words, ability of the study to detect a difference if such a difference exists in reality.
- Power gives no information about type I error or false positive error.
- It is the complement of beta error so we write
- Power = 1- beta error
- Power + beta error = 1
- If there is a difference in reality and we are able to detect it…complement
Concept of alpha α:
⦁ Alpha is NOT AN ERROR.
⦁ This α is compliment level of confidence. Generally we take confidence=95% so α is 5%=0.05.
⦁ Alpha is a test criteria or a cut off value at which we accept or reject the null hypothesis
⦁ Level of confidence is decided IN ADVANCE before study.
⦁ It is surety 95% CI means 95% sure and 5% mistake.
⦁ Once we decide level of CI, α is fixed i.e. cut off is fixed so it is decided before sample size.
⦁ P value of study we get from test of significance i.e. occurs after study is conducted. P is actual mistake in the study as it is occurred after conduction of study.
⦁ If p value < α set for the study, we reject null hypothesis and declare result statistically significant.
| Level of confidence | Complement of confidence (α) |
|---|---|
| 99% | 1% = 0.01 |
| 95% | 5% = 0.05 |
Note:
- If level of confidence is 95% i.e. α=0.05, but we get p value of 0.04, do we call it significant?
- The answer is yes as it is less than 0.05.
- If the level of confidence is raised to 99% i.e. α=0.01 and p value of the study is 0.04, then this becomes non-significant as it is more than 0.04.
- So on increasing level of confidence previously significant value becomes non-significant.
