After matching, even though the stratum-specific ORs still give valid estimation of

the RR, the crude OR however, was biased towards the null value indicating that
matching has introduced negative confounding into the sample.

Again, matched analysis would give valid estimation of the underlying association
between the exposure and the disease of interest, the confounding introduced by
matching would be eliminated:

C. Confounding introduced by matching is predictable. Whether the matching

variable is a positive confounder or negative confounder, or only associated with the
exposure of interest (therefore, not a confounder) in the population which produced
the cases, the confounding introduced by matching is generally in the direction of a
bias towards the null value for a crude estimate (RR-1. AR 0, r 0, b o), since
matching a variable which is associated with exposure of interest would force the
distribution of exposure (as well as the matched variable) similar or identical
between the cases and controls.

However, in unusual situation, u e crude association between exposure and disesse

cold be fanher*vey from the null value than the association ufter stratification by
the matching verisble. This could occur only if uere is negative correlation in
exposure histories between cases and their matched cantrols either from sampling
variabiliry or from extreme effect modification.

D. Confounding introduced by matching is controllable. As shown in the previous

examples, if the matching variable is taken into consideration in the data analysis
G.e., by calculating stratum-specific estimator or summary estimate adjusted for
the matched variable), the confounding introduced by matching and the original
confounding existing in the population would both be eliminated. This is the very
reason why we utilize matching to help control confounding in a case-control study
even though matching itself can not eliminate confounding or may even introduce a
negative confounding: matching creates balanced 2x2 tables, therefore, promoting
study efficiency and guaranteeing a balanced comparison within each level of the
matched variable. On the other hand, if we stratify the data by the matched
variable, the original confounding together with the confounding introduced by
matching would both be eliminated. Therefore, matching eams study efficiency
without loss of study validity (if proper analysis is used).

3. Types of Matching

1. Full Matching (Perfect Matching

If matching eliminates the confounding effect from the matching variable in a cohort
study or in a case-control study with matched analysis, then, we say that matching
is a perfect matching.

Usually, matching by a discrete variable with a nominal scale will result in perfect
matching. A discrete variable with a nominal scale is a variable whose value has no
intermediate level, each value can only fall into one of the two or more mutually
exclusive categories which have no inherent order. These variables include
dichotomous variables (such as sex, survival status) and multichotomous variables
(such as religion, race, etc.)

Matching by a dichotomous variable is full matching since individuals can only be

allocated into either of two mutually exclusive categories: for example, gender
(male or female); patient's survival status (alive or dead).

Matching by multichotomous variable is full matching, since individuals can only be

classified into each mutually exclusive and unordered qualitative categories, the
measurements for each category is purely in nominal scales (in name), such as
matching by religion (Catholic, Protestant, Jewish, other); race (white, black,
Hispanic, other), blood type (A, B, AB, O); hair color (dark, brown, blond, red); or
country of birth, etc. Once the variable is matched and proper analysis is taken, the
confounding effect by the matching variable is eliminated.

2. Partial Matchin

There are two types of partial matching: one is referred to as the residual
confounding from matching a continuous variable or an ordinal variable; the other
implies incomplete matching of all the required matching variables.

(1) Matching variable is a continuous variable or a discrete variable with ordinal

scale

A continuous variable has a potentially infinite number possible values along a of

continuum within a specified range, such as height, weight, blood pressure,
measurements of blood level of selenium, B-carotene, cholesterol.

An ordinal variable has ordered qualitative categories, such as the stage of disease
at the diagnosis I, III, IV); social class (low, middle and high); smoking or alcohol
intake (none, light, moderate and heavy); age (young, middle age, and old); number
of liver births (0, 1-2, 3-4, 5+); height (short, medium, tall); blood pressure
(hypotensive, normotensive, hypertensive). The major feature of an ordinal variable
is their values have a distinct order.

Matching a continuous variable usually involves artificially dividing the variable into
several categories so it becomes an ordinal variable. For example, divide age into
young, middle age and old; divide blood pressure, Quintelet Index, and blood
cholesterol levels into low, normal and high.
The ability to control the confounding by the matched variable, and therefore how
strong the residual confounding existing in the data is depends on how finely the
categories are generated, and within each category whether the disease rate is
relatively homogeneous. If the study subjects at the upper or lower bounds of each
category have quite different disease rates, residual confounding could be a serious
problem.

For example, in a study of physical activity and cardiovascular disease, if the study
decided to match the following variables in the selection of controls: smoking in two
categories (ever smoker and never smoker); alcohol drinking in two categories (ever
drinker and never drinker); age in 2 categories (<50 and 250, there is potential for
residual confounding from the matching variables.

For example, in the case of age as a confounder, persons between age <50 have
very different disease rate (CVD) and physical activity (E), therefore, after matching,
age is still associated with the disease (CvD) and exposure (physical activity).
Matching by age at such wide range can not eliminate the confounding effect from
age. In fact, a misclassification of confounding could even be introduced.

(2) Incomplete matching of the required matching variables

For some studies, a large number of confounding variables is anticipated. If control

of confounding only relies on stratified analysis or multivariate analysis in these
studies, the total number of strata could be extremely large, the study subjects
would be spread too thinly over the strata, and therefore, no contrast could be
made from these strata. In such a case matching becomes unavoidable. However, it
is also possible that if every control must match all the matching variables, many
cases may have to be excluded from the study simply because the study can not
find suitable controls for the corresponding cases.

For example, in a study of physical activity and risk of CvD.a large number of
environmental, life styles and behavior factors are potential confounders of the
alleged association: such as smoking, alcohol drinking, dietary intake, serum
cholesterol level, personality, income, education level, occupational exposure, sex,
race and age. Let's assume the study has decided to match on:

sex (in two categories);

race (in four categories);

age fin five categories).

there are already 40 (2 x 4 x 5) possible combinations that have to be considered in
finding a suitable control. However, if hospital controls are used, the control disease
entities must also not be associated with the exposure of interes: and if cases are
incident cases, the controls' diseases must also be newly diagnosed. In order to
avoid selection bias, referral pattern must also be controlled.