1
00:00:01,170 --> 00:00:03,425
Hi everyone, and welcome to this lecture

2
00:00:03,425 --> 00:00:07,833
on ways to reduce or avoid confounding.

3
00:00:09,780 --> 00:00:11,971
So just a brief recap.

4
00:00:11,971 --> 00:00:14,550
We learned this in epi one

5
00:00:14,550 --> 00:00:17,970
and then we briefly looked
at it in module one.

6
00:00:17,970 --> 00:00:18,803
So to be a confound

7
00:00:18,803 --> 00:00:22,344
where a variable must be
associated with the exposure,

8
00:00:22,344 --> 00:00:26,553
associated with the outcome,
and not on the causal pathway.

9
00:00:27,870 --> 00:00:31,195
Today we'll talk about ways
to reduce or avoid confounding

10
00:00:31,195 --> 00:00:33,164
and the different things
that we'll talk about

11
00:00:33,164 --> 00:00:35,220
are discussed below.

12
00:00:35,220 --> 00:00:38,070
So restriction, matching, randomization,

13
00:00:38,070 --> 00:00:41,023
stratification, and multivariate analysis.

14
00:00:41,023 --> 00:00:43,681
And when I go through examples,

15
00:00:43,681 --> 00:00:47,199
I like to try to talk
about apples and oranges

16
00:00:47,199 --> 00:00:49,730
or other fruits, something
that's completely

17
00:00:49,730 --> 00:00:52,438
kind of out of the medical field

18
00:00:52,438 --> 00:00:55,773
to simplify things as much as I can.

19
00:00:56,627 --> 00:00:58,830
So the first thing that we can do

20
00:00:58,830 --> 00:01:03,830
to decrease the risk of
confounding in our study

21
00:01:04,228 --> 00:01:05,694
is called restriction.

22
00:01:05,694 --> 00:01:08,710
So restriction is when we
limit the study to individuals

23
00:01:08,710 --> 00:01:13,003
who fall within a specified
category of each confounder.

24
00:01:13,003 --> 00:01:15,155
So in our example below,

25
00:01:15,155 --> 00:01:18,196
we're now comparing apples to apples.

26
00:01:18,196 --> 00:01:20,144
The limitation of restriction

27
00:01:20,144 --> 00:01:25,080
is that we reduce the
generalizability of our study.

28
00:01:25,080 --> 00:01:29,400
So we are now looking
at apples versus apples,

29
00:01:29,400 --> 00:01:30,517
we're not looking at fruits,

30
00:01:30,517 --> 00:01:32,560
and the results from our apples study

31
00:01:32,560 --> 00:01:36,696
can't necessarily be
generalized to all fruits.

32
00:01:36,696 --> 00:01:40,394
We can also have difficulties
recruiting participants.

33
00:01:40,394 --> 00:01:42,578
And this becomes more of a problem

34
00:01:42,578 --> 00:01:45,157
if we are restricting
on several variables.

35
00:01:45,157 --> 00:01:49,381
So for example, if I wanted
to look at green apples

36
00:01:49,381 --> 00:01:52,593
that were grown in Vermont
in a certain months,

37
00:01:52,593 --> 00:01:57,593
I may have trouble getting
enough apples to do a good study.

38
00:01:59,648 --> 00:02:01,998
The next thing we can do
to decrease confounding

39
00:02:01,998 --> 00:02:04,916
in our studies is called matching.

40
00:02:04,916 --> 00:02:06,587
And this is where we select subjects

41
00:02:06,587 --> 00:02:08,450
such that the confounders are distributed

42
00:02:08,450 --> 00:02:11,760
in identical manner among groups.

43
00:02:11,760 --> 00:02:15,205
This is really useful
in case control studies.

44
00:02:15,205 --> 00:02:18,081
We have to be careful of over matching,

45
00:02:18,081 --> 00:02:20,108
which occurs when cases and controls

46
00:02:20,108 --> 00:02:22,321
are inadvertently matched

47
00:02:22,321 --> 00:02:26,084
on characteristics that
is potentially causal.

48
00:02:26,084 --> 00:02:28,095
For example, if you matched

49
00:02:28,095 --> 00:02:32,610
on smoking status when the
outcome is lung cancer.

50
00:02:32,610 --> 00:02:37,610
We can't then look at the
impact of smoking on lung cancer

51
00:02:37,864 --> 00:02:40,852
because we've matched for that variable.

52
00:02:40,852 --> 00:02:43,231
There are two types of matching.

53
00:02:43,231 --> 00:02:45,244
We have individual matching,

54
00:02:45,244 --> 00:02:47,725
which is for every case you match one

55
00:02:47,725 --> 00:02:50,152
or more controls with the same value.

56
00:02:50,152 --> 00:02:53,321
So here we would match
every apple to every apple

57
00:02:53,321 --> 00:02:56,308
or every green apple to every green apple,

58
00:02:56,308 --> 00:02:58,661
or every green apple grown organically

59
00:02:58,661 --> 00:02:59,965
in the state of Vermont

60
00:02:59,965 --> 00:03:01,911
to every green apple grown organically

61
00:03:01,911 --> 00:03:03,492
in the state of Vermont.

62
00:03:03,492 --> 00:03:05,930
And then we have frequency matching,

63
00:03:05,930 --> 00:03:07,891
which is where you create a control group

64
00:03:07,891 --> 00:03:10,938
with similar means and proportions
of confounding variables

65
00:03:10,938 --> 00:03:12,792
as the case group.

66
00:03:12,792 --> 00:03:15,924
So here we would get a
group of apples and oranges

67
00:03:15,924 --> 00:03:20,410
and then match that group
characteristics to a group

68
00:03:22,088 --> 00:03:24,693
which will be our control group.

69
00:03:26,310 --> 00:03:27,959
The strengths and limitations of matching

70
00:03:27,959 --> 00:03:29,630
are shown on the screen.

71
00:03:29,630 --> 00:03:33,202
So matching is intuitive
and easy to explain.

72
00:03:33,202 --> 00:03:35,735
It guarantees a certain
degree of comparability

73
00:03:35,735 --> 00:03:39,522
even in small studies, and it's efficient,

74
00:03:39,522 --> 00:03:43,276
especially if you're matching
on a strong confounder.

75
00:03:43,276 --> 00:03:45,888
The limitations is that it's costly.

76
00:03:45,888 --> 00:03:49,922
If it's logistically complicated,
it can be an efficient,

77
00:03:49,922 --> 00:03:52,281
if you're matching on a weak confounder,

78
00:03:52,281 --> 00:03:55,033
you can't study the matching variable.

79
00:03:55,033 --> 00:03:57,120
And the more things we match on,

80
00:03:57,120 --> 00:03:59,026
the harder it is to find a match.

81
00:03:59,026 --> 00:04:03,003
And you also have the possibility
of residual confounding.

82
00:04:04,710 --> 00:04:07,328
So here's an example of a
study that I've actually done

83
00:04:07,328 --> 00:04:10,110
where we used restricting and matching

84
00:04:10,110 --> 00:04:14,040
to decrease the impact of
confounding on our results.

85
00:04:14,040 --> 00:04:15,761
So here we looked at
patient reported outcomes

86
00:04:15,761 --> 00:04:18,763
at six to 12 months among survivors

87
00:04:18,763 --> 00:04:21,316
of firearm injuries in the US.

88
00:04:21,316 --> 00:04:26,316
So we wanted to look at traumatic
injuries as our exposure

89
00:04:26,459 --> 00:04:29,758
and patient reported
outcomes as our outcome.

90
00:04:29,758 --> 00:04:34,758
And then we thought
that the green variables

91
00:04:34,800 --> 00:04:35,995
listed on the screen

92
00:04:35,995 --> 00:04:39,533
could be potential confounding variables.

93
00:04:39,533 --> 00:04:43,953
And so we restricted the
study to some categories

94
00:04:43,953 --> 00:04:48,081
and then we matched the
study on other categories.

95
00:04:48,081 --> 00:04:50,861
So we restricted our study to adults

96
00:04:50,861 --> 00:04:54,533
who were between the
age of 18 to 64 years.

97
00:04:56,818 --> 00:05:01,506
We looked only at patients
admitted in 2015 to 2018,

98
00:05:01,506 --> 00:05:04,971
and we restricted our study
to moderate to severe injuries

99
00:05:04,971 --> 00:05:08,591
from firearm injuries or
motor vehicle crashes.

100
00:05:08,591 --> 00:05:12,199
And then we matched the
firearm injury patients

101
00:05:12,199 --> 00:05:15,850
to motor vehicle crash injury patients

102
00:05:15,850 --> 00:05:19,997
on age, gender, race, and education level,

103
00:05:19,997 --> 00:05:24,997
and on a history of previous
psychiatric illness diagnosis.

104
00:05:26,223 --> 00:05:28,444
And then we found that
firearm injury survivors

105
00:05:28,444 --> 00:05:29,497
were more likely to have

106
00:05:29,497 --> 00:05:32,852
daily pain screened positive for PTSD

107
00:05:32,852 --> 00:05:34,734
and had worse physical and mental health

108
00:05:34,734 --> 00:05:35,914
related quality of life

109
00:05:35,914 --> 00:05:40,503
compared to similarly injured
motor vehicle crash survivors.

110
00:05:41,432 --> 00:05:42,979
So the next thing that we can do

111
00:05:42,979 --> 00:05:45,842
to decrease the impact of
confounding on our studies

112
00:05:45,842 --> 00:05:47,811
is called randomization,

113
00:05:47,811 --> 00:05:50,083
which is where you randomly
allocate individuals

114
00:05:50,083 --> 00:05:51,046
to study groups.

115
00:05:51,046 --> 00:05:54,409
The strengths is that the
groups are likely to be balanced

116
00:05:54,409 --> 00:05:56,410
in terms of identified confounders

117
00:05:56,410 --> 00:05:59,907
and potential confounders
that weren't identified.

118
00:05:59,907 --> 00:06:01,206
So in our example below

119
00:06:01,206 --> 00:06:05,875
we're comparing a random
allocation of apple and oranges

120
00:06:05,875 --> 00:06:09,123
to a random allocation
of apples and oranges.

121
00:06:10,380 --> 00:06:13,152
And potential unknown confounder

122
00:06:13,152 --> 00:06:16,621
that may impact our study
results could be worms,

123
00:06:16,621 --> 00:06:18,807
and we don't know which fruit have worms

124
00:06:18,807 --> 00:06:21,873
because we can't see
which ones have worms.

125
00:06:21,873 --> 00:06:24,793
But because we've randomly
allocated the fruit,

126
00:06:24,793 --> 00:06:28,331
we've also randomly allocated the worms.

127
00:06:28,331 --> 00:06:32,068
The limitations are that you
can have unlucky randomization,

128
00:06:32,068 --> 00:06:34,380
which can cause an imbalance,

129
00:06:34,380 --> 00:06:37,101
and this is more likely
with small sample sizes.

130
00:06:37,101 --> 00:06:42,033
And randomization isn't
always feasible or ethical.

131
00:06:43,083 --> 00:06:45,765
The next way that we can reduce

132
00:06:45,765 --> 00:06:49,068
the potential impacts of
confounding on our study

133
00:06:49,068 --> 00:06:51,180
is called stratification.

134
00:06:51,180 --> 00:06:52,169
And this is when

135
00:06:52,169 --> 00:06:54,613
the study population is
broken into subgroups

136
00:06:54,613 --> 00:06:58,035
according to the level of
potential confounding factors,

137
00:06:58,035 --> 00:07:03,035
and you actually produce
results for each level.

138
00:07:03,234 --> 00:07:06,420
So the strength is that you
have more granular data,

139
00:07:06,420 --> 00:07:09,728
and the limitations are that
you have smaller sample sizes

140
00:07:09,728 --> 00:07:11,548
within each subdivision.

141
00:07:11,548 --> 00:07:14,490
So here we would produce results

142
00:07:14,490 --> 00:07:17,005
for our oranges versus oranges analysis,

143
00:07:17,005 --> 00:07:20,130
and then we would also produce results

144
00:07:20,130 --> 00:07:22,519
for our apple versus apple analysis.

145
00:07:22,519 --> 00:07:24,070
Now, this becomes more difficult

146
00:07:24,070 --> 00:07:26,894
if we want to do an analysis on

147
00:07:26,894 --> 00:07:31,016
apples that are organic versus
apples that are not organic.

148
00:07:31,016 --> 00:07:32,463
So we're doing,

149
00:07:34,921 --> 00:07:37,826
it becomes more complicated

150
00:07:37,826 --> 00:07:42,826
if you stratify it down
to too many levels.

151
00:07:44,700 --> 00:07:46,179
So an example of a study that I've done,

152
00:07:46,179 --> 00:07:49,439
where we looked at,

153
00:07:49,439 --> 00:07:51,590
where we used stratification

154
00:07:51,590 --> 00:07:54,783
is this study where we
looked at the impact of

155
00:07:54,783 --> 00:07:57,598
income on emergency
general surgery outcomes

156
00:07:57,598 --> 00:08:00,693
in urban and rural areas.

157
00:08:01,560 --> 00:08:05,498
So we wanted to have
income as our exposure

158
00:08:05,498 --> 00:08:08,164
and emergency general
surgery adverse outcomes

159
00:08:08,164 --> 00:08:10,127
as our outcome.

160
00:08:10,127 --> 00:08:12,395
And then rurality was the variable

161
00:08:12,395 --> 00:08:14,791
that we thought would be
the confounding factor

162
00:08:14,791 --> 00:08:17,215
that we wanted to stratify by.

163
00:08:17,215 --> 00:08:20,470
So low income emergency
general surgery patients

164
00:08:20,470 --> 00:08:23,044
have higher rates of
postoperative adverse events

165
00:08:23,044 --> 00:08:25,350
compared to high income patients.

166
00:08:25,350 --> 00:08:28,421
And this may be related
to healthcare segregation.

167
00:08:28,421 --> 00:08:32,136
The emergent nature of emergency
general surgery conditions

168
00:08:32,136 --> 00:08:36,291
and limited number of emergency
general surgery providers

169
00:08:36,291 --> 00:08:40,140
in rural areas may result in
less healthcare segregation

170
00:08:40,140 --> 00:08:42,133
and therefore less variability

171
00:08:42,133 --> 00:08:46,383
in emergency general surgery
outcomes in rural areas.

172
00:08:48,771 --> 00:08:51,616
And so to do this analysis,

173
00:08:51,616 --> 00:08:55,889
we actually stratified our analysis

174
00:08:55,889 --> 00:08:58,732
into rural groups and urban groups,

175
00:08:58,732 --> 00:09:00,874
and we found that income was associated

176
00:09:00,874 --> 00:09:03,588
with higher postoperative
adverse events in urban,

177
00:09:03,588 --> 00:09:05,751
but not in rural settings.

178
00:09:05,751 --> 00:09:07,661
So the socioeconomic disparity

179
00:09:07,661 --> 00:09:10,596
in emergency general surgery
outcomes in urban settings

180
00:09:10,596 --> 00:09:13,144
may reflect healthcare segregation,

181
00:09:13,144 --> 00:09:14,435
which is a differential access

182
00:09:14,435 --> 00:09:17,643
to high quality healthcare
for low income patients.

183
00:09:19,727 --> 00:09:22,330
The last thing that we'll touch on today

184
00:09:22,330 --> 00:09:26,683
is multi-variable or
multivariate regression analysis.

185
00:09:26,683 --> 00:09:29,650
These are advanced statistical methods

186
00:09:29,650 --> 00:09:32,100
that allow us to remove the effects

187
00:09:32,100 --> 00:09:35,949
of specified confounding variables.

188
00:09:35,949 --> 00:09:40,949
And we'll really touch on
these analysis techniques

189
00:09:42,196 --> 00:09:45,399
later on in epi two.

190
00:09:45,399 --> 00:09:47,278
But basically they allow us to see

191
00:09:47,278 --> 00:09:48,870
if there's a difference after removing

192
00:09:48,870 --> 00:09:52,572
the effects of specified
confounding variables.

193
00:09:52,572 --> 00:09:54,450
The strengths are that it's inexpensive

194
00:09:54,450 --> 00:09:58,897
and that it can handle multiple
confounders simultaneously.

195
00:09:58,897 --> 00:10:02,379
So this is a study where we looked,

196
00:10:02,379 --> 00:10:04,765
we did a multivariate analysis.

197
00:10:04,765 --> 00:10:07,756
We looked at the lethality
of active shooter incidents

198
00:10:07,756 --> 00:10:11,460
with and without semi-automatic
rifles in the US.

199
00:10:11,460 --> 00:10:14,587
And so our exposure was the presence of

200
00:10:14,587 --> 00:10:18,510
semi-automatic rifles at
an active shooter event.

201
00:10:18,510 --> 00:10:23,142
And our outcome was the number
of persons wounded or killed.

202
00:10:23,142 --> 00:10:25,184
The variables in green

203
00:10:25,184 --> 00:10:27,532
were the potential confounding variables

204
00:10:27,532 --> 00:10:29,403
that we wanted to adjust for.

205
00:10:29,403 --> 00:10:32,460
So the place of the shooting,
the year of the shooting,

206
00:10:32,460 --> 00:10:34,860
and the presence of other firearms.

207
00:10:34,860 --> 00:10:38,006
And then we found after doing our analysis

208
00:10:38,006 --> 00:10:40,250
that more people were wounded or killed

209
00:10:40,250 --> 00:10:43,024
in incidents in which
semi-automatic rifles were used

210
00:10:43,024 --> 00:10:46,143
compared to incidents
involving other firearms.

211
00:10:47,910 --> 00:10:50,398
So today we've briefly discussed

212
00:10:50,398 --> 00:10:54,719
on the ways to reduce or avoid confounding

213
00:10:54,719 --> 00:10:57,199
when we are designing a study,

214
00:10:57,199 --> 00:11:02,155
restriction, matching,
randomization, stratification,

215
00:11:02,155 --> 00:11:05,341
and multivariate analysis.

216
00:11:05,341 --> 00:11:07,740
So that concludes this lecture.

217
00:11:07,740 --> 00:11:08,643
Take care.